program(1.3)
[mldb_token = string("mldb-5qbg63zgxe")]
{
    func foundInApp<ios16>(tensor<int32, [1, 256, 1, 3]> mlm_input) {
            int32 var_6 = const()[name = string("op_6"), val = int32(0)];
            tensor<int32, [3]> var_13 = const()[name = string("op_13"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_14_axis_0 = const()[name = string("op_14_axis_0"), val = int32(-1)];
            tensor<int32, [1, 256, 1, 1]> var_14_0, tensor<int32, [1, 256, 1, 1]> var_14_1, tensor<int32, [1, 256, 1, 1]> var_14_2 = split(axis = var_14_axis_0, split_sizes = var_13, x = mlm_input)[name = string("op_14")];
            tensor<int32, [1]> var_18_axes_0 = const()[name = string("op_18_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_18 = squeeze(axes = var_18_axes_0, x = var_14_0)[name = string("op_18")];
            tensor<int32, [1]> tok_ids_1_axes_0 = const()[name = string("tok_ids_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256]> tok_ids_1 = squeeze(axes = tok_ids_1_axes_0, x = var_18)[name = string("tok_ids_1")];
            tensor<int32, [1]> var_20_axes_0 = const()[name = string("op_20_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_20 = squeeze(axes = var_20_axes_0, x = var_14_1)[name = string("op_20")];
            tensor<int32, [1]> var_22_axes_0 = const()[name = string("op_22_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_22 = squeeze(axes = var_22_axes_0, x = var_14_2)[name = string("op_22")];
            tensor<bool, [1, 256]> var_24 = not_equal(x = tok_ids_1, y = var_6)[name = string("op_24")];
            fp16 var_8_to_fp16 = const()[name = string("op_8_to_fp16"), val = fp16(1)];
            string var_24_to_fp32_to_fp16_dtype_0 = const()[name = string("op_24_to_fp32_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256]> cast_1 = cast(dtype = var_24_to_fp32_to_fp16_dtype_0, x = var_24)[name = string("cast_1")];
            tensor<fp16, [1, 256]> var_29_cast_fp16 = sub(x = var_8_to_fp16, y = cast_1)[name = string("op_29_cast_fp16")];
            fp16 var_30_to_fp16 = const()[name = string("op_30_to_fp16"), val = fp16(-10000)];
            tensor<fp16, [1, 256]> padding_mask0_1_cast_fp16 = mul(x = var_29_cast_fp16, y = var_30_to_fp16)[name = string("padding_mask0_1_cast_fp16")];
            tensor<int32, [4]> var_32 = const()[name = string("op_32"), val = tensor<int32, [4]>([-1, 256, 1, 1])];
            tensor<fp16, [1, 256, 1, 1]> var_33_cast_fp16 = reshape(shape = var_32, x = padding_mask0_1_cast_fp16)[name = string("op_33_cast_fp16")];
            int32 var_46 = const()[name = string("op_46"), val = int32(1)];
            int32 var_47 = const()[name = string("op_47"), val = int32(2)];
            tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 256, 1]> input_3 = expand_dims(axes = input_3_axes_0, x = tok_ids_1)[name = string("input_3")];
            int32 var_54_axis_0 = const()[name = string("op_54_axis_0"), val = int32(0)];
            tensor<fp16, [50000, 512]> nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [50000, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(150272))), scale = tensor<fp16, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50176))), zero_point = tensor<int8, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(64)))];
            int32 op_54_cast_fp16_batch_dims_0 = const()[name = string("op_54_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_54_cast_fp16 = gather(axis = var_54_axis_0, batch_dims = op_54_cast_fp16_batch_dims_0, indices = input_3, x = nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized)[name = string("op_54_cast_fp16")];
            int32 var_57_axis_0 = const()[name = string("op_57_axis_0"), val = int32(0)];
            tensor<fp16, [256, 512]> nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [256, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25751232))), scale = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750656))), zero_point = tensor<int8, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750336)))];
            int32 op_57_cast_fp16_batch_dims_0 = const()[name = string("op_57_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_57_cast_fp16 = gather(axis = var_57_axis_0, batch_dims = op_57_cast_fp16_batch_dims_0, indices = var_20, x = nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized)[name = string("op_57_cast_fp16")];
            int32 var_60_axis_0 = const()[name = string("op_60_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512]> nlp_net_default_encoder_seg_embed_weight_to_fp16 = const()[name = string("nlp_net_default_encoder_seg_embed_weight_to_fp16"), val = tensor<fp16, [1, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25882368)))];
            int32 op_60_cast_fp16_batch_dims_0 = const()[name = string("op_60_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_60_cast_fp16 = gather(axis = var_60_axis_0, batch_dims = op_60_cast_fp16_batch_dims_0, indices = var_22, x = nlp_net_default_encoder_seg_embed_weight_to_fp16)[name = string("op_60_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_62_cast_fp16 = add(x = var_54_cast_fp16, y = var_57_cast_fp16)[name = string("op_62_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_63_cast_fp16 = add(x = var_62_cast_fp16, y = var_60_cast_fp16)[name = string("op_63_cast_fp16")];
            tensor<int32, [4]> t_1_perm_0 = const()[name = string("t_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_3_gamma_0_to_fp16 = const()[name = string("k_3_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53441792)))];
            tensor<fp16, [512]> k_3_beta_0_to_fp16 = const()[name = string("k_3_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53442880)))];
            fp16 var_96_to_fp16 = const()[name = string("op_96_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> transpose_9 = transpose(perm = t_1_perm_0, x = var_63_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 512, 1, 256]> k_3_cast_fp16 = layer_norm(axes = k_3_axes_0, beta = k_3_beta_0_to_fp16, epsilon = var_96_to_fp16, gamma = k_3_gamma_0_to_fp16, x = transpose_9)[name = string("k_3_cast_fp16")];
            tensor<int32, [2]> var_115 = const()[name = string("op_115"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_117 = const()[name = string("op_117"), val = tensor<int32, [2]>([1, 1])];
            string var_119_pad_type_0 = const()[name = string("op_119_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_119_pad_0 = const()[name = string("op_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25887296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25886208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26149504)))];
            tensor<fp16, [1, 512, 1, 256]> var_119_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16, dilations = var_117, groups = var_46, pad = var_119_pad_0, pad_type = var_119_pad_type_0, strides = var_115, weight = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_119_cast_fp16")];
            tensor<int32, [2]> var_122 = const()[name = string("op_122"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_124 = const()[name = string("op_124"), val = tensor<int32, [2]>([1, 1])];
            string k_5_pad_type_0 = const()[name = string("k_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_5_pad_0 = const()[name = string("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26151680))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26150592))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26413888)))];
            tensor<fp16, [1, 512, 1, 256]> k_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16, dilations = var_124, groups = var_46, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = var_122, weight = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [2]> var_129 = const()[name = string("op_129"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_131 = const()[name = string("op_131"), val = tensor<int32, [2]>([1, 1])];
            string var_133_pad_type_0 = const()[name = string("op_133_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_133_pad_0 = const()[name = string("op_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26416064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26414976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26678272)))];
            tensor<fp16, [1, 512, 1, 256]> var_133_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16, dilations = var_131, groups = var_46, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_129, weight = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [8]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_134_axis_0 = const()[name = string("op_134_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_7 = split(axis = var_134_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = string("op_134_cast_fp16")];
            tensor<int32, [4]> var_143_perm_0 = const()[name = string("op_143_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_144_axis_0 = const()[name = string("op_144_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_8 = transpose(perm = var_143_perm_0, x = k_5_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_7 = split(axis = var_144_axis_0, split_sizes = tile_1, x = transpose_8)[name = string("op_144_cast_fp16")];
            tensor<int32, [8]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_153_axis_0 = const()[name = string("op_153_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_7 = split(axis = var_153_axis_0, split_sizes = tile_2, x = var_133_cast_fp16)[name = string("op_153_cast_fp16")];
            string var_163_equation_0 = const()[name = string("op_163_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_144_cast_fp16_0, var_134_cast_fp16_0))[name = string("op_163_cast_fp16")];
            fp16 var_164_to_fp16 = const()[name = string("op_164_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_165_cast_fp16 = mul(x = var_163_cast_fp16, y = var_164_to_fp16)[name = string("op_165_cast_fp16")];
            string var_167_equation_0 = const()[name = string("op_167_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_144_cast_fp16_1, var_134_cast_fp16_1))[name = string("op_167_cast_fp16")];
            fp16 var_168_to_fp16 = const()[name = string("op_168_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_169_cast_fp16 = mul(x = var_167_cast_fp16, y = var_168_to_fp16)[name = string("op_169_cast_fp16")];
            string var_171_equation_0 = const()[name = string("op_171_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_144_cast_fp16_2, var_134_cast_fp16_2))[name = string("op_171_cast_fp16")];
            fp16 var_172_to_fp16 = const()[name = string("op_172_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_173_cast_fp16 = mul(x = var_171_cast_fp16, y = var_172_to_fp16)[name = string("op_173_cast_fp16")];
            string var_175_equation_0 = const()[name = string("op_175_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_175_cast_fp16 = einsum(equation = var_175_equation_0, values = (var_144_cast_fp16_3, var_134_cast_fp16_3))[name = string("op_175_cast_fp16")];
            fp16 var_176_to_fp16 = const()[name = string("op_176_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_177_cast_fp16 = mul(x = var_175_cast_fp16, y = var_176_to_fp16)[name = string("op_177_cast_fp16")];
            string var_179_equation_0 = const()[name = string("op_179_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_179_cast_fp16 = einsum(equation = var_179_equation_0, values = (var_144_cast_fp16_4, var_134_cast_fp16_4))[name = string("op_179_cast_fp16")];
            fp16 var_180_to_fp16 = const()[name = string("op_180_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_181_cast_fp16 = mul(x = var_179_cast_fp16, y = var_180_to_fp16)[name = string("op_181_cast_fp16")];
            string var_183_equation_0 = const()[name = string("op_183_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_183_cast_fp16 = einsum(equation = var_183_equation_0, values = (var_144_cast_fp16_5, var_134_cast_fp16_5))[name = string("op_183_cast_fp16")];
            fp16 var_184_to_fp16 = const()[name = string("op_184_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_185_cast_fp16 = mul(x = var_183_cast_fp16, y = var_184_to_fp16)[name = string("op_185_cast_fp16")];
            string var_187_equation_0 = const()[name = string("op_187_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_187_cast_fp16 = einsum(equation = var_187_equation_0, values = (var_144_cast_fp16_6, var_134_cast_fp16_6))[name = string("op_187_cast_fp16")];
            fp16 var_188_to_fp16 = const()[name = string("op_188_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_189_cast_fp16 = mul(x = var_187_cast_fp16, y = var_188_to_fp16)[name = string("op_189_cast_fp16")];
            string var_191_equation_0 = const()[name = string("op_191_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_191_cast_fp16 = einsum(equation = var_191_equation_0, values = (var_144_cast_fp16_7, var_134_cast_fp16_7))[name = string("op_191_cast_fp16")];
            fp16 var_192_to_fp16 = const()[name = string("op_192_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_193_cast_fp16 = mul(x = var_191_cast_fp16, y = var_192_to_fp16)[name = string("op_193_cast_fp16")];
            bool attn_weights_2_interleave_0 = const()[name = string("attn_weights_2_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_2_cast_fp16 = concat(axis = var_47, interleave = attn_weights_2_interleave_0, values = (var_165_cast_fp16, var_169_cast_fp16, var_173_cast_fp16, var_177_cast_fp16, var_181_cast_fp16, var_185_cast_fp16, var_189_cast_fp16, var_193_cast_fp16))[name = string("attn_weights_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_2_cast_fp16 = add(x = attn_weights_2_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_7_cast_fp16 = softmax(axis = var_46, x = attn_weights0_2_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<int32, [8]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_199_axis_0 = const()[name = string("op_199_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_7 = split(axis = var_199_axis_0, split_sizes = tile_3, x = input_7_cast_fp16)[name = string("op_199_cast_fp16")];
            string var_209_equation_0 = const()[name = string("op_209_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_209_cast_fp16 = einsum(equation = var_209_equation_0, values = (var_153_cast_fp16_0, var_199_cast_fp16_0))[name = string("op_209_cast_fp16")];
            string var_211_equation_0 = const()[name = string("op_211_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_211_cast_fp16 = einsum(equation = var_211_equation_0, values = (var_153_cast_fp16_1, var_199_cast_fp16_1))[name = string("op_211_cast_fp16")];
            string var_213_equation_0 = const()[name = string("op_213_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_213_cast_fp16 = einsum(equation = var_213_equation_0, values = (var_153_cast_fp16_2, var_199_cast_fp16_2))[name = string("op_213_cast_fp16")];
            string var_215_equation_0 = const()[name = string("op_215_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_215_cast_fp16 = einsum(equation = var_215_equation_0, values = (var_153_cast_fp16_3, var_199_cast_fp16_3))[name = string("op_215_cast_fp16")];
            string var_217_equation_0 = const()[name = string("op_217_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_217_cast_fp16 = einsum(equation = var_217_equation_0, values = (var_153_cast_fp16_4, var_199_cast_fp16_4))[name = string("op_217_cast_fp16")];
            string var_219_equation_0 = const()[name = string("op_219_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_219_cast_fp16 = einsum(equation = var_219_equation_0, values = (var_153_cast_fp16_5, var_199_cast_fp16_5))[name = string("op_219_cast_fp16")];
            string var_221_equation_0 = const()[name = string("op_221_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_221_cast_fp16 = einsum(equation = var_221_equation_0, values = (var_153_cast_fp16_6, var_199_cast_fp16_6))[name = string("op_221_cast_fp16")];
            string var_223_equation_0 = const()[name = string("op_223_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_223_cast_fp16 = einsum(equation = var_223_equation_0, values = (var_153_cast_fp16_7, var_199_cast_fp16_7))[name = string("op_223_cast_fp16")];
            bool attn_5_interleave_0 = const()[name = string("attn_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_5_cast_fp16 = concat(axis = var_46, interleave = attn_5_interleave_0, values = (var_209_cast_fp16, var_211_cast_fp16, var_213_cast_fp16, var_215_cast_fp16, var_217_cast_fp16, var_219_cast_fp16, var_221_cast_fp16, var_223_cast_fp16))[name = string("attn_5_cast_fp16")];
            tensor<int32, [2]> var_231 = const()[name = string("op_231"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_233 = const()[name = string("op_233"), val = tensor<int32, [2]>([1, 1])];
            string input0_5_pad_type_0 = const()[name = string("input0_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_5_pad_0 = const()[name = string("input0_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26680448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26679360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26942656)))];
            tensor<fp16, [1, 512, 1, 256]> input0_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16, dilations = var_233, groups = var_46, pad = input0_5_pad_0, pad_type = input0_5_pad_type_0, strides = var_231, weight = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_5_cast_fp16)[name = string("input0_5_cast_fp16")];
            tensor<int32, [2]> var_240 = const()[name = string("op_240"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_242 = const()[name = string("op_242"), val = tensor<int32, [2]>([1, 1])];
            string x_6_pad_type_0 = const()[name = string("x_6_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_6_pad_0 = const()[name = string("x_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53444288))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53443968))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53509888)))];
            tensor<fp16, [1, 128, 1, 256]> x_6_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16, dilations = var_242, groups = var_46, pad = x_6_pad_0, pad_type = x_6_pad_type_0, strides = var_240, weight = nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_5_cast_fp16)[name = string("x_6_cast_fp16")];
            fp16 var_245_to_fp16 = const()[name = string("op_245_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_246_cast_fp16 = mul(x = x_6_cast_fp16, y = var_245_to_fp16)[name = string("op_246_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_247_cast_fp16 = sigmoid(x = var_246_cast_fp16)[name = string("op_247_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_9_cast_fp16 = mul(x = x_6_cast_fp16, y = var_247_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<int32, [2]> var_251 = const()[name = string("op_251"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_253 = const()[name = string("op_253"), val = tensor<int32, [2]>([1, 1])];
            string x_8_pad_type_0 = const()[name = string("x_8_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_8_pad_0 = const()[name = string("x_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53511296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53510208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53576896)))];
            tensor<fp16, [1, 512, 1, 256]> x_8_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16, dilations = var_253, groups = var_46, pad = x_8_pad_0, pad_type = x_8_pad_type_0, strides = var_251, weight = nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_9_cast_fp16)[name = string("x_8_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_7_cast_fp16 = add(x = x_8_cast_fp16, y = input0_5_cast_fp16)[name = string("attn_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_1_cast_fp16 = add(x = transpose_9, y = attn_7_cast_fp16)[name = string("inputs_1_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_11_gamma_0_to_fp16 = const()[name = string("input_11_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53577984)))];
            tensor<fp16, [512]> input_11_beta_0_to_fp16 = const()[name = string("input_11_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53579072)))];
            fp16 var_266_to_fp16 = const()[name = string("op_266_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = input_11_beta_0_to_fp16, epsilon = var_266_to_fp16, gamma = input_11_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<int32, [2]> var_280 = const()[name = string("op_280"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_282 = const()[name = string("op_282"), val = tensor<int32, [2]>([1, 1])];
            string x_10_pad_type_0 = const()[name = string("x_10_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_10_pad_0 = const()[name = string("x_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26952192))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26948032))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28004992))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28000832))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_10_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_282, groups = var_46, pad = x_10_pad_0, pad_type = x_10_pad_type_0, strides = var_280, weight = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_11_cast_fp16)[name = string("x_10_cast_fp16")];
            fp16 var_285_to_fp16 = const()[name = string("op_285_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_286_cast_fp16 = mul(x = x_10_cast_fp16, y = var_285_to_fp16)[name = string("op_286_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_287_cast_fp16 = sigmoid(x = var_286_cast_fp16)[name = string("op_287_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_13_cast_fp16 = mul(x = x_10_cast_fp16, y = var_287_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<int32, [2]> var_291 = const()[name = string("op_291"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_293 = const()[name = string("op_293"), val = tensor<int32, [2]>([1, 1])];
            string input0_7_pad_type_0 = const()[name = string("input0_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_7_pad_0 = const()[name = string("input0_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28008192))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28007104))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29056832)))];
            tensor<fp16, [1, 512, 1, 256]> input0_7_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16, dilations = var_293, groups = var_46, pad = input0_7_pad_0, pad_type = input0_7_pad_type_0, strides = var_291, weight = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_13_cast_fp16)[name = string("input0_7_cast_fp16")];
            tensor<int32, [2]> var_301 = const()[name = string("op_301"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_303 = const()[name = string("op_303"), val = tensor<int32, [2]>([1, 1])];
            string x_12_pad_type_0 = const()[name = string("x_12_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_12_pad_0 = const()[name = string("x_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53580480))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53580160))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53646080)))];
            tensor<fp16, [1, 128, 1, 256]> x_12_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_303, groups = var_46, pad = x_12_pad_0, pad_type = x_12_pad_type_0, strides = var_301, weight = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_7_cast_fp16)[name = string("x_12_cast_fp16")];
            fp16 var_306_to_fp16 = const()[name = string("op_306_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_307_cast_fp16 = mul(x = x_12_cast_fp16, y = var_306_to_fp16)[name = string("op_307_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_308_cast_fp16 = sigmoid(x = var_307_cast_fp16)[name = string("op_308_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_17_cast_fp16 = mul(x = x_12_cast_fp16, y = var_308_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<int32, [2]> var_312 = const()[name = string("op_312"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_314 = const()[name = string("op_314"), val = tensor<int32, [2]>([1, 1])];
            string x_14_pad_type_0 = const()[name = string("x_14_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_14_pad_0 = const()[name = string("x_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53647488))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53646400))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53713088)))];
            tensor<fp16, [1, 512, 1, 256]> x_14_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_314, groups = var_46, pad = x_14_pad_0, pad_type = x_14_pad_type_0, strides = var_312, weight = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_17_cast_fp16)[name = string("x_14_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_2_cast_fp16 = add(x = x_14_cast_fp16, y = input0_7_cast_fp16)[name = string("f_2_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_2_cast_fp16 = add(x = f_2_cast_fp16, y = inputs_1_cast_fp16)[name = string("x1_2_cast_fp16")];
            fp16 var_319_to_fp16 = const()[name = string("op_319_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_320_cast_fp16 = mul(x = transpose_9, y = var_319_to_fp16)[name = string("op_320_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_2_cast_fp16 = add(x = var_320_cast_fp16, y = x1_2_cast_fp16)[name = string("inputs_2_cast_fp16")];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_7_gamma_0_to_fp16 = const()[name = string("k_7_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53714176)))];
            tensor<fp16, [512]> k_7_beta_0_to_fp16 = const()[name = string("k_7_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53715264)))];
            fp16 var_338_to_fp16 = const()[name = string("op_338_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_7_cast_fp16 = layer_norm(axes = k_7_axes_0, beta = k_7_beta_0_to_fp16, epsilon = var_338_to_fp16, gamma = k_7_gamma_0_to_fp16, x = inputs_2_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<int32, [2]> var_357 = const()[name = string("op_357"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_359 = const()[name = string("op_359"), val = tensor<int32, [2]>([1, 1])];
            string var_361_pad_type_0 = const()[name = string("op_361_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_361_pad_0 = const()[name = string("op_361_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29061184))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29060096))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29323392)))];
            tensor<fp16, [1, 512, 1, 256]> var_361_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16, dilations = var_359, groups = var_46, pad = var_361_pad_0, pad_type = var_361_pad_type_0, strides = var_357, weight = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_361_cast_fp16")];
            tensor<int32, [2]> var_364 = const()[name = string("op_364"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_366 = const()[name = string("op_366"), val = tensor<int32, [2]>([1, 1])];
            string k_9_pad_type_0 = const()[name = string("k_9_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_9_pad_0 = const()[name = string("k_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29325568))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29324480))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29587776)))];
            tensor<fp16, [1, 512, 1, 256]> k_9_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16, dilations = var_366, groups = var_46, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = var_364, weight = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [2]> var_371 = const()[name = string("op_371"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_373 = const()[name = string("op_373"), val = tensor<int32, [2]>([1, 1])];
            string var_375_pad_type_0 = const()[name = string("op_375_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_375_pad_0 = const()[name = string("op_375_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29589952))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29588864))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29852160)))];
            tensor<fp16, [1, 512, 1, 256]> var_375_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16, dilations = var_373, groups = var_46, pad = var_375_pad_0, pad_type = var_375_pad_type_0, strides = var_371, weight = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_375_cast_fp16")];
            tensor<int32, [8]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_376_axis_0 = const()[name = string("op_376_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_376_cast_fp16_7 = split(axis = var_376_axis_0, split_sizes = tile_4, x = var_361_cast_fp16)[name = string("op_376_cast_fp16")];
            tensor<int32, [4]> var_385_perm_0 = const()[name = string("op_385_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_386_axis_0 = const()[name = string("op_386_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_7 = transpose(perm = var_385_perm_0, x = k_9_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_386_cast_fp16_7 = split(axis = var_386_axis_0, split_sizes = tile_5, x = transpose_7)[name = string("op_386_cast_fp16")];
            tensor<int32, [8]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_395_axis_0 = const()[name = string("op_395_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_395_cast_fp16_7 = split(axis = var_395_axis_0, split_sizes = tile_6, x = var_375_cast_fp16)[name = string("op_395_cast_fp16")];
            string var_405_equation_0 = const()[name = string("op_405_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_405_cast_fp16 = einsum(equation = var_405_equation_0, values = (var_386_cast_fp16_0, var_376_cast_fp16_0))[name = string("op_405_cast_fp16")];
            fp16 var_406_to_fp16 = const()[name = string("op_406_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_407_cast_fp16 = mul(x = var_405_cast_fp16, y = var_406_to_fp16)[name = string("op_407_cast_fp16")];
            string var_409_equation_0 = const()[name = string("op_409_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_409_cast_fp16 = einsum(equation = var_409_equation_0, values = (var_386_cast_fp16_1, var_376_cast_fp16_1))[name = string("op_409_cast_fp16")];
            fp16 var_410_to_fp16 = const()[name = string("op_410_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_411_cast_fp16 = mul(x = var_409_cast_fp16, y = var_410_to_fp16)[name = string("op_411_cast_fp16")];
            string var_413_equation_0 = const()[name = string("op_413_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_413_cast_fp16 = einsum(equation = var_413_equation_0, values = (var_386_cast_fp16_2, var_376_cast_fp16_2))[name = string("op_413_cast_fp16")];
            fp16 var_414_to_fp16 = const()[name = string("op_414_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_415_cast_fp16 = mul(x = var_413_cast_fp16, y = var_414_to_fp16)[name = string("op_415_cast_fp16")];
            string var_417_equation_0 = const()[name = string("op_417_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_417_cast_fp16 = einsum(equation = var_417_equation_0, values = (var_386_cast_fp16_3, var_376_cast_fp16_3))[name = string("op_417_cast_fp16")];
            fp16 var_418_to_fp16 = const()[name = string("op_418_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_419_cast_fp16 = mul(x = var_417_cast_fp16, y = var_418_to_fp16)[name = string("op_419_cast_fp16")];
            string var_421_equation_0 = const()[name = string("op_421_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_421_cast_fp16 = einsum(equation = var_421_equation_0, values = (var_386_cast_fp16_4, var_376_cast_fp16_4))[name = string("op_421_cast_fp16")];
            fp16 var_422_to_fp16 = const()[name = string("op_422_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_423_cast_fp16 = mul(x = var_421_cast_fp16, y = var_422_to_fp16)[name = string("op_423_cast_fp16")];
            string var_425_equation_0 = const()[name = string("op_425_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_425_cast_fp16 = einsum(equation = var_425_equation_0, values = (var_386_cast_fp16_5, var_376_cast_fp16_5))[name = string("op_425_cast_fp16")];
            fp16 var_426_to_fp16 = const()[name = string("op_426_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = string("op_427_cast_fp16")];
            string var_429_equation_0 = const()[name = string("op_429_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_429_cast_fp16 = einsum(equation = var_429_equation_0, values = (var_386_cast_fp16_6, var_376_cast_fp16_6))[name = string("op_429_cast_fp16")];
            fp16 var_430_to_fp16 = const()[name = string("op_430_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_431_cast_fp16 = mul(x = var_429_cast_fp16, y = var_430_to_fp16)[name = string("op_431_cast_fp16")];
            string var_433_equation_0 = const()[name = string("op_433_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_433_cast_fp16 = einsum(equation = var_433_equation_0, values = (var_386_cast_fp16_7, var_376_cast_fp16_7))[name = string("op_433_cast_fp16")];
            fp16 var_434_to_fp16 = const()[name = string("op_434_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_435_cast_fp16 = mul(x = var_433_cast_fp16, y = var_434_to_fp16)[name = string("op_435_cast_fp16")];
            bool attn_weights_4_interleave_0 = const()[name = string("attn_weights_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_4_cast_fp16 = concat(axis = var_47, interleave = attn_weights_4_interleave_0, values = (var_407_cast_fp16, var_411_cast_fp16, var_415_cast_fp16, var_419_cast_fp16, var_423_cast_fp16, var_427_cast_fp16, var_431_cast_fp16, var_435_cast_fp16))[name = string("attn_weights_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_4_cast_fp16 = add(x = attn_weights_4_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_19_cast_fp16 = softmax(axis = var_46, x = attn_weights0_4_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<int32, [8]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_441_axis_0 = const()[name = string("op_441_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_441_cast_fp16_7 = split(axis = var_441_axis_0, split_sizes = tile_7, x = input_19_cast_fp16)[name = string("op_441_cast_fp16")];
            string var_451_equation_0 = const()[name = string("op_451_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_451_cast_fp16 = einsum(equation = var_451_equation_0, values = (var_395_cast_fp16_0, var_441_cast_fp16_0))[name = string("op_451_cast_fp16")];
            string var_453_equation_0 = const()[name = string("op_453_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_453_cast_fp16 = einsum(equation = var_453_equation_0, values = (var_395_cast_fp16_1, var_441_cast_fp16_1))[name = string("op_453_cast_fp16")];
            string var_455_equation_0 = const()[name = string("op_455_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_455_cast_fp16 = einsum(equation = var_455_equation_0, values = (var_395_cast_fp16_2, var_441_cast_fp16_2))[name = string("op_455_cast_fp16")];
            string var_457_equation_0 = const()[name = string("op_457_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_457_cast_fp16 = einsum(equation = var_457_equation_0, values = (var_395_cast_fp16_3, var_441_cast_fp16_3))[name = string("op_457_cast_fp16")];
            string var_459_equation_0 = const()[name = string("op_459_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_459_cast_fp16 = einsum(equation = var_459_equation_0, values = (var_395_cast_fp16_4, var_441_cast_fp16_4))[name = string("op_459_cast_fp16")];
            string var_461_equation_0 = const()[name = string("op_461_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_461_cast_fp16 = einsum(equation = var_461_equation_0, values = (var_395_cast_fp16_5, var_441_cast_fp16_5))[name = string("op_461_cast_fp16")];
            string var_463_equation_0 = const()[name = string("op_463_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_463_cast_fp16 = einsum(equation = var_463_equation_0, values = (var_395_cast_fp16_6, var_441_cast_fp16_6))[name = string("op_463_cast_fp16")];
            string var_465_equation_0 = const()[name = string("op_465_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_465_cast_fp16 = einsum(equation = var_465_equation_0, values = (var_395_cast_fp16_7, var_441_cast_fp16_7))[name = string("op_465_cast_fp16")];
            bool attn_11_interleave_0 = const()[name = string("attn_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_11_cast_fp16 = concat(axis = var_46, interleave = attn_11_interleave_0, values = (var_451_cast_fp16, var_453_cast_fp16, var_455_cast_fp16, var_457_cast_fp16, var_459_cast_fp16, var_461_cast_fp16, var_463_cast_fp16, var_465_cast_fp16))[name = string("attn_11_cast_fp16")];
            tensor<int32, [2]> var_473 = const()[name = string("op_473"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_475 = const()[name = string("op_475"), val = tensor<int32, [2]>([1, 1])];
            string input0_11_pad_type_0 = const()[name = string("input0_11_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_11_pad_0 = const()[name = string("input0_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29854336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29853248))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30116544)))];
            tensor<fp16, [1, 512, 1, 256]> input0_11_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16, dilations = var_475, groups = var_46, pad = input0_11_pad_0, pad_type = input0_11_pad_type_0, strides = var_473, weight = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_11_cast_fp16)[name = string("input0_11_cast_fp16")];
            tensor<int32, [2]> var_482 = const()[name = string("op_482"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_484 = const()[name = string("op_484"), val = tensor<int32, [2]>([1, 1])];
            string x_16_pad_type_0 = const()[name = string("x_16_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_16_pad_0 = const()[name = string("x_16_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53716672))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53716352))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53782272)))];
            tensor<fp16, [1, 128, 1, 256]> x_16_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16, dilations = var_484, groups = var_46, pad = x_16_pad_0, pad_type = x_16_pad_type_0, strides = var_482, weight = nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_11_cast_fp16)[name = string("x_16_cast_fp16")];
            fp16 var_487_to_fp16 = const()[name = string("op_487_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_488_cast_fp16 = mul(x = x_16_cast_fp16, y = var_487_to_fp16)[name = string("op_488_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_489_cast_fp16 = sigmoid(x = var_488_cast_fp16)[name = string("op_489_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_21_cast_fp16 = mul(x = x_16_cast_fp16, y = var_489_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [2]> var_493 = const()[name = string("op_493"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_495 = const()[name = string("op_495"), val = tensor<int32, [2]>([1, 1])];
            string x_18_pad_type_0 = const()[name = string("x_18_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_18_pad_0 = const()[name = string("x_18_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53783680))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53782592))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53849280)))];
            tensor<fp16, [1, 512, 1, 256]> x_18_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16, dilations = var_495, groups = var_46, pad = x_18_pad_0, pad_type = x_18_pad_type_0, strides = var_493, weight = nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_21_cast_fp16)[name = string("x_18_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_13_cast_fp16 = add(x = x_18_cast_fp16, y = input0_11_cast_fp16)[name = string("attn_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_4_cast_fp16 = add(x = inputs_2_cast_fp16, y = attn_13_cast_fp16)[name = string("inputs0_4_cast_fp16")];
            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_23_gamma_0_to_fp16 = const()[name = string("input_23_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53850368)))];
            tensor<fp16, [512]> input_23_beta_0_to_fp16 = const()[name = string("input_23_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53851456)))];
            fp16 var_508_to_fp16 = const()[name = string("op_508_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_508_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs0_4_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<int32, [2]> var_522 = const()[name = string("op_522"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_524 = const()[name = string("op_524"), val = tensor<int32, [2]>([1, 1])];
            string x_20_pad_type_0 = const()[name = string("x_20_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_20_pad_0 = const()[name = string("x_20_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30123968))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30119808))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31176768))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31172608))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_20_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_524, groups = var_46, pad = x_20_pad_0, pad_type = x_20_pad_type_0, strides = var_522, weight = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_23_cast_fp16)[name = string("x_20_cast_fp16")];
            fp16 var_527_to_fp16 = const()[name = string("op_527_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_528_cast_fp16 = mul(x = x_20_cast_fp16, y = var_527_to_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_529_cast_fp16 = sigmoid(x = var_528_cast_fp16)[name = string("op_529_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_25_cast_fp16 = mul(x = x_20_cast_fp16, y = var_529_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<int32, [2]> var_533 = const()[name = string("op_533"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_535 = const()[name = string("op_535"), val = tensor<int32, [2]>([1, 1])];
            string input0_13_pad_type_0 = const()[name = string("input0_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_13_pad_0 = const()[name = string("input0_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31179968))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31178880))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32228608)))];
            tensor<fp16, [1, 512, 1, 256]> input0_13_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16, dilations = var_535, groups = var_46, pad = input0_13_pad_0, pad_type = input0_13_pad_type_0, strides = var_533, weight = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_25_cast_fp16)[name = string("input0_13_cast_fp16")];
            tensor<int32, [2]> var_543 = const()[name = string("op_543"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_545 = const()[name = string("op_545"), val = tensor<int32, [2]>([1, 1])];
            string x_22_pad_type_0 = const()[name = string("x_22_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_22_pad_0 = const()[name = string("x_22_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53852864))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53852544))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53918464)))];
            tensor<fp16, [1, 128, 1, 256]> x_22_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_545, groups = var_46, pad = x_22_pad_0, pad_type = x_22_pad_type_0, strides = var_543, weight = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_13_cast_fp16)[name = string("x_22_cast_fp16")];
            fp16 var_548_to_fp16 = const()[name = string("op_548_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_549_cast_fp16 = mul(x = x_22_cast_fp16, y = var_548_to_fp16)[name = string("op_549_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_550_cast_fp16 = sigmoid(x = var_549_cast_fp16)[name = string("op_550_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_29_cast_fp16 = mul(x = x_22_cast_fp16, y = var_550_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [2]> var_554 = const()[name = string("op_554"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_556 = const()[name = string("op_556"), val = tensor<int32, [2]>([1, 1])];
            string x_24_pad_type_0 = const()[name = string("x_24_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_24_pad_0 = const()[name = string("x_24_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53919872))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53918784))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53985472)))];
            tensor<fp16, [1, 512, 1, 256]> x_24_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_556, groups = var_46, pad = x_24_pad_0, pad_type = x_24_pad_type_0, strides = var_554, weight = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_29_cast_fp16)[name = string("x_24_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_4_cast_fp16 = add(x = x_24_cast_fp16, y = input0_13_cast_fp16)[name = string("f_4_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_4_cast_fp16 = add(x = f_4_cast_fp16, y = inputs0_4_cast_fp16)[name = string("x1_4_cast_fp16")];
            fp16 var_561_to_fp16 = const()[name = string("op_561_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_562_cast_fp16 = mul(x = inputs_2_cast_fp16, y = var_561_to_fp16)[name = string("op_562_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_2_cast_fp16 = add(x = var_562_cast_fp16, y = x1_4_cast_fp16)[name = string("inputs0_2_cast_fp16")];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_11_gamma_0_to_fp16 = const()[name = string("k_11_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53986560)))];
            tensor<fp16, [512]> k_11_beta_0_to_fp16 = const()[name = string("k_11_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53987648)))];
            fp16 var_580_to_fp16 = const()[name = string("op_580_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_11_cast_fp16 = layer_norm(axes = k_11_axes_0, beta = k_11_beta_0_to_fp16, epsilon = var_580_to_fp16, gamma = k_11_gamma_0_to_fp16, x = inputs0_2_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [2]> var_599 = const()[name = string("op_599"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_601 = const()[name = string("op_601"), val = tensor<int32, [2]>([1, 1])];
            string var_603_pad_type_0 = const()[name = string("op_603_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_603_pad_0 = const()[name = string("op_603_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32232960))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32231872))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32495168)))];
            tensor<fp16, [1, 512, 1, 256]> var_603_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16, dilations = var_601, groups = var_46, pad = var_603_pad_0, pad_type = var_603_pad_type_0, strides = var_599, weight = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_603_cast_fp16")];
            tensor<int32, [2]> var_606 = const()[name = string("op_606"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_608 = const()[name = string("op_608"), val = tensor<int32, [2]>([1, 1])];
            string k_13_pad_type_0 = const()[name = string("k_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_13_pad_0 = const()[name = string("k_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32497344))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32496256))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32759552)))];
            tensor<fp16, [1, 512, 1, 256]> k_13_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16, dilations = var_608, groups = var_46, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = var_606, weight = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [2]> var_613 = const()[name = string("op_613"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_615 = const()[name = string("op_615"), val = tensor<int32, [2]>([1, 1])];
            string var_617_pad_type_0 = const()[name = string("op_617_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_617_pad_0 = const()[name = string("op_617_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32761728))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32760640))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33023936)))];
            tensor<fp16, [1, 512, 1, 256]> var_617_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16, dilations = var_615, groups = var_46, pad = var_617_pad_0, pad_type = var_617_pad_type_0, strides = var_613, weight = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_617_cast_fp16")];
            tensor<int32, [8]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_618_axis_0 = const()[name = string("op_618_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_618_cast_fp16_7 = split(axis = var_618_axis_0, split_sizes = tile_8, x = var_603_cast_fp16)[name = string("op_618_cast_fp16")];
            tensor<int32, [4]> var_627_perm_0 = const()[name = string("op_627_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_628_axis_0 = const()[name = string("op_628_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_6 = transpose(perm = var_627_perm_0, x = k_13_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_628_cast_fp16_7 = split(axis = var_628_axis_0, split_sizes = tile_9, x = transpose_6)[name = string("op_628_cast_fp16")];
            tensor<int32, [8]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_637_axis_0 = const()[name = string("op_637_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_637_cast_fp16_7 = split(axis = var_637_axis_0, split_sizes = tile_10, x = var_617_cast_fp16)[name = string("op_637_cast_fp16")];
            string var_647_equation_0 = const()[name = string("op_647_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_647_cast_fp16 = einsum(equation = var_647_equation_0, values = (var_628_cast_fp16_0, var_618_cast_fp16_0))[name = string("op_647_cast_fp16")];
            fp16 var_648_to_fp16 = const()[name = string("op_648_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_649_cast_fp16 = mul(x = var_647_cast_fp16, y = var_648_to_fp16)[name = string("op_649_cast_fp16")];
            string var_651_equation_0 = const()[name = string("op_651_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_651_cast_fp16 = einsum(equation = var_651_equation_0, values = (var_628_cast_fp16_1, var_618_cast_fp16_1))[name = string("op_651_cast_fp16")];
            fp16 var_652_to_fp16 = const()[name = string("op_652_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_653_cast_fp16 = mul(x = var_651_cast_fp16, y = var_652_to_fp16)[name = string("op_653_cast_fp16")];
            string var_655_equation_0 = const()[name = string("op_655_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_655_cast_fp16 = einsum(equation = var_655_equation_0, values = (var_628_cast_fp16_2, var_618_cast_fp16_2))[name = string("op_655_cast_fp16")];
            fp16 var_656_to_fp16 = const()[name = string("op_656_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_657_cast_fp16 = mul(x = var_655_cast_fp16, y = var_656_to_fp16)[name = string("op_657_cast_fp16")];
            string var_659_equation_0 = const()[name = string("op_659_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_659_cast_fp16 = einsum(equation = var_659_equation_0, values = (var_628_cast_fp16_3, var_618_cast_fp16_3))[name = string("op_659_cast_fp16")];
            fp16 var_660_to_fp16 = const()[name = string("op_660_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_661_cast_fp16 = mul(x = var_659_cast_fp16, y = var_660_to_fp16)[name = string("op_661_cast_fp16")];
            string var_663_equation_0 = const()[name = string("op_663_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_663_cast_fp16 = einsum(equation = var_663_equation_0, values = (var_628_cast_fp16_4, var_618_cast_fp16_4))[name = string("op_663_cast_fp16")];
            fp16 var_664_to_fp16 = const()[name = string("op_664_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_665_cast_fp16 = mul(x = var_663_cast_fp16, y = var_664_to_fp16)[name = string("op_665_cast_fp16")];
            string var_667_equation_0 = const()[name = string("op_667_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_667_cast_fp16 = einsum(equation = var_667_equation_0, values = (var_628_cast_fp16_5, var_618_cast_fp16_5))[name = string("op_667_cast_fp16")];
            fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_669_cast_fp16 = mul(x = var_667_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")];
            string var_671_equation_0 = const()[name = string("op_671_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_671_cast_fp16 = einsum(equation = var_671_equation_0, values = (var_628_cast_fp16_6, var_618_cast_fp16_6))[name = string("op_671_cast_fp16")];
            fp16 var_672_to_fp16 = const()[name = string("op_672_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_673_cast_fp16 = mul(x = var_671_cast_fp16, y = var_672_to_fp16)[name = string("op_673_cast_fp16")];
            string var_675_equation_0 = const()[name = string("op_675_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_675_cast_fp16 = einsum(equation = var_675_equation_0, values = (var_628_cast_fp16_7, var_618_cast_fp16_7))[name = string("op_675_cast_fp16")];
            fp16 var_676_to_fp16 = const()[name = string("op_676_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_677_cast_fp16 = mul(x = var_675_cast_fp16, y = var_676_to_fp16)[name = string("op_677_cast_fp16")];
            bool attn_weights_6_interleave_0 = const()[name = string("attn_weights_6_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_6_cast_fp16 = concat(axis = var_47, interleave = attn_weights_6_interleave_0, values = (var_649_cast_fp16, var_653_cast_fp16, var_657_cast_fp16, var_661_cast_fp16, var_665_cast_fp16, var_669_cast_fp16, var_673_cast_fp16, var_677_cast_fp16))[name = string("attn_weights_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_6_cast_fp16 = add(x = attn_weights_6_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_31_cast_fp16 = softmax(axis = var_46, x = attn_weights0_6_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<int32, [8]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_683_axis_0 = const()[name = string("op_683_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_683_cast_fp16_7 = split(axis = var_683_axis_0, split_sizes = tile_11, x = input_31_cast_fp16)[name = string("op_683_cast_fp16")];
            string var_693_equation_0 = const()[name = string("op_693_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_693_cast_fp16 = einsum(equation = var_693_equation_0, values = (var_637_cast_fp16_0, var_683_cast_fp16_0))[name = string("op_693_cast_fp16")];
            string var_695_equation_0 = const()[name = string("op_695_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_695_cast_fp16 = einsum(equation = var_695_equation_0, values = (var_637_cast_fp16_1, var_683_cast_fp16_1))[name = string("op_695_cast_fp16")];
            string var_697_equation_0 = const()[name = string("op_697_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_697_cast_fp16 = einsum(equation = var_697_equation_0, values = (var_637_cast_fp16_2, var_683_cast_fp16_2))[name = string("op_697_cast_fp16")];
            string var_699_equation_0 = const()[name = string("op_699_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_699_cast_fp16 = einsum(equation = var_699_equation_0, values = (var_637_cast_fp16_3, var_683_cast_fp16_3))[name = string("op_699_cast_fp16")];
            string var_701_equation_0 = const()[name = string("op_701_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_701_cast_fp16 = einsum(equation = var_701_equation_0, values = (var_637_cast_fp16_4, var_683_cast_fp16_4))[name = string("op_701_cast_fp16")];
            string var_703_equation_0 = const()[name = string("op_703_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_703_cast_fp16 = einsum(equation = var_703_equation_0, values = (var_637_cast_fp16_5, var_683_cast_fp16_5))[name = string("op_703_cast_fp16")];
            string var_705_equation_0 = const()[name = string("op_705_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_705_cast_fp16 = einsum(equation = var_705_equation_0, values = (var_637_cast_fp16_6, var_683_cast_fp16_6))[name = string("op_705_cast_fp16")];
            string var_707_equation_0 = const()[name = string("op_707_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_707_cast_fp16 = einsum(equation = var_707_equation_0, values = (var_637_cast_fp16_7, var_683_cast_fp16_7))[name = string("op_707_cast_fp16")];
            bool attn_17_interleave_0 = const()[name = string("attn_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_17_cast_fp16 = concat(axis = var_46, interleave = attn_17_interleave_0, values = (var_693_cast_fp16, var_695_cast_fp16, var_697_cast_fp16, var_699_cast_fp16, var_701_cast_fp16, var_703_cast_fp16, var_705_cast_fp16, var_707_cast_fp16))[name = string("attn_17_cast_fp16")];
            tensor<int32, [2]> var_715 = const()[name = string("op_715"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_717 = const()[name = string("op_717"), val = tensor<int32, [2]>([1, 1])];
            string input0_17_pad_type_0 = const()[name = string("input0_17_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_17_pad_0 = const()[name = string("input0_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33026112))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33025024))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33288320)))];
            tensor<fp16, [1, 512, 1, 256]> input0_17_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16, dilations = var_717, groups = var_46, pad = input0_17_pad_0, pad_type = input0_17_pad_type_0, strides = var_715, weight = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_17_cast_fp16)[name = string("input0_17_cast_fp16")];
            tensor<int32, [2]> var_724 = const()[name = string("op_724"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_726 = const()[name = string("op_726"), val = tensor<int32, [2]>([1, 1])];
            string x_26_pad_type_0 = const()[name = string("x_26_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_26_pad_0 = const()[name = string("x_26_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53989056))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53988736))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54054656)))];
            tensor<fp16, [1, 128, 1, 256]> x_26_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16, dilations = var_726, groups = var_46, pad = x_26_pad_0, pad_type = x_26_pad_type_0, strides = var_724, weight = nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_17_cast_fp16)[name = string("x_26_cast_fp16")];
            fp16 var_729_to_fp16 = const()[name = string("op_729_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_730_cast_fp16 = mul(x = x_26_cast_fp16, y = var_729_to_fp16)[name = string("op_730_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_731_cast_fp16 = sigmoid(x = var_730_cast_fp16)[name = string("op_731_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_33_cast_fp16 = mul(x = x_26_cast_fp16, y = var_731_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [2]> var_735 = const()[name = string("op_735"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_737 = const()[name = string("op_737"), val = tensor<int32, [2]>([1, 1])];
            string x_28_pad_type_0 = const()[name = string("x_28_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_28_pad_0 = const()[name = string("x_28_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54056064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54054976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54121664)))];
            tensor<fp16, [1, 512, 1, 256]> x_28_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16, dilations = var_737, groups = var_46, pad = x_28_pad_0, pad_type = x_28_pad_type_0, strides = var_735, weight = nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_33_cast_fp16)[name = string("x_28_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_19_cast_fp16 = add(x = x_28_cast_fp16, y = input0_17_cast_fp16)[name = string("attn_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_6_cast_fp16 = add(x = inputs0_2_cast_fp16, y = attn_19_cast_fp16)[name = string("inputs0_6_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_35_gamma_0_to_fp16 = const()[name = string("input_35_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54122752)))];
            tensor<fp16, [512]> input_35_beta_0_to_fp16 = const()[name = string("input_35_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54123840)))];
            fp16 var_750_to_fp16 = const()[name = string("op_750_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = input_35_beta_0_to_fp16, epsilon = var_750_to_fp16, gamma = input_35_gamma_0_to_fp16, x = inputs0_6_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<int32, [2]> var_764 = const()[name = string("op_764"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_766 = const()[name = string("op_766"), val = tensor<int32, [2]>([1, 1])];
            string x_30_pad_type_0 = const()[name = string("x_30_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_30_pad_0 = const()[name = string("x_30_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33295744))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33291584))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34348544))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34344384))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_30_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_766, groups = var_46, pad = x_30_pad_0, pad_type = x_30_pad_type_0, strides = var_764, weight = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_35_cast_fp16)[name = string("x_30_cast_fp16")];
            fp16 var_769_to_fp16 = const()[name = string("op_769_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_770_cast_fp16 = mul(x = x_30_cast_fp16, y = var_769_to_fp16)[name = string("op_770_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_771_cast_fp16 = sigmoid(x = var_770_cast_fp16)[name = string("op_771_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_37_cast_fp16 = mul(x = x_30_cast_fp16, y = var_771_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<int32, [2]> var_775 = const()[name = string("op_775"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_777 = const()[name = string("op_777"), val = tensor<int32, [2]>([1, 1])];
            string input0_19_pad_type_0 = const()[name = string("input0_19_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_19_pad_0 = const()[name = string("input0_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34351744))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34350656))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35400384)))];
            tensor<fp16, [1, 512, 1, 256]> input0_19_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16, dilations = var_777, groups = var_46, pad = input0_19_pad_0, pad_type = input0_19_pad_type_0, strides = var_775, weight = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_37_cast_fp16)[name = string("input0_19_cast_fp16")];
            tensor<int32, [2]> var_785 = const()[name = string("op_785"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_787 = const()[name = string("op_787"), val = tensor<int32, [2]>([1, 1])];
            string x_32_pad_type_0 = const()[name = string("x_32_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_32_pad_0 = const()[name = string("x_32_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54125248))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54124928))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54190848)))];
            tensor<fp16, [1, 128, 1, 256]> x_32_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_787, groups = var_46, pad = x_32_pad_0, pad_type = x_32_pad_type_0, strides = var_785, weight = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_19_cast_fp16)[name = string("x_32_cast_fp16")];
            fp16 var_790_to_fp16 = const()[name = string("op_790_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_791_cast_fp16 = mul(x = x_32_cast_fp16, y = var_790_to_fp16)[name = string("op_791_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_792_cast_fp16 = sigmoid(x = var_791_cast_fp16)[name = string("op_792_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_41_cast_fp16 = mul(x = x_32_cast_fp16, y = var_792_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<int32, [2]> var_796 = const()[name = string("op_796"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_798 = const()[name = string("op_798"), val = tensor<int32, [2]>([1, 1])];
            string x_34_pad_type_0 = const()[name = string("x_34_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_34_pad_0 = const()[name = string("x_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54192256))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54191168))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54257856)))];
            tensor<fp16, [1, 512, 1, 256]> x_34_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_798, groups = var_46, pad = x_34_pad_0, pad_type = x_34_pad_type_0, strides = var_796, weight = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_41_cast_fp16)[name = string("x_34_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_6_cast_fp16 = add(x = x_34_cast_fp16, y = input0_19_cast_fp16)[name = string("f_6_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_6_cast_fp16 = add(x = f_6_cast_fp16, y = inputs0_6_cast_fp16)[name = string("x1_6_cast_fp16")];
            fp16 var_803_to_fp16 = const()[name = string("op_803_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_804_cast_fp16 = mul(x = inputs0_2_cast_fp16, y = var_803_to_fp16)[name = string("op_804_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs1_1_cast_fp16 = add(x = var_804_cast_fp16, y = x1_6_cast_fp16)[name = string("inputs1_1_cast_fp16")];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_15_gamma_0_to_fp16 = const()[name = string("k_15_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54258944)))];
            tensor<fp16, [512]> k_15_beta_0_to_fp16 = const()[name = string("k_15_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54260032)))];
            fp16 var_822_to_fp16 = const()[name = string("op_822_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_15_cast_fp16 = layer_norm(axes = k_15_axes_0, beta = k_15_beta_0_to_fp16, epsilon = var_822_to_fp16, gamma = k_15_gamma_0_to_fp16, x = inputs1_1_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [2]> var_841 = const()[name = string("op_841"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_843 = const()[name = string("op_843"), val = tensor<int32, [2]>([1, 1])];
            string var_845_pad_type_0 = const()[name = string("op_845_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_845_pad_0 = const()[name = string("op_845_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35404736))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35403648))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35666944)))];
            tensor<fp16, [1, 512, 1, 256]> var_845_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16, dilations = var_843, groups = var_46, pad = var_845_pad_0, pad_type = var_845_pad_type_0, strides = var_841, weight = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_845_cast_fp16")];
            tensor<int32, [2]> var_848 = const()[name = string("op_848"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_850 = const()[name = string("op_850"), val = tensor<int32, [2]>([1, 1])];
            string k_17_pad_type_0 = const()[name = string("k_17_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_17_pad_0 = const()[name = string("k_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35669120))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35668032))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35931328)))];
            tensor<fp16, [1, 512, 1, 256]> k_17_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16, dilations = var_850, groups = var_46, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = var_848, weight = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [2]> var_855 = const()[name = string("op_855"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_857 = const()[name = string("op_857"), val = tensor<int32, [2]>([1, 1])];
            string var_859_pad_type_0 = const()[name = string("op_859_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_859_pad_0 = const()[name = string("op_859_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35933504))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35932416))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36195712)))];
            tensor<fp16, [1, 512, 1, 256]> var_859_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16, dilations = var_857, groups = var_46, pad = var_859_pad_0, pad_type = var_859_pad_type_0, strides = var_855, weight = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_859_cast_fp16")];
            tensor<int32, [8]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_860_axis_0 = const()[name = string("op_860_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_860_cast_fp16_7 = split(axis = var_860_axis_0, split_sizes = tile_12, x = var_845_cast_fp16)[name = string("op_860_cast_fp16")];
            tensor<int32, [4]> var_869_perm_0 = const()[name = string("op_869_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_870_axis_0 = const()[name = string("op_870_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_5 = transpose(perm = var_869_perm_0, x = k_17_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_870_cast_fp16_7 = split(axis = var_870_axis_0, split_sizes = tile_13, x = transpose_5)[name = string("op_870_cast_fp16")];
            tensor<int32, [8]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_879_axis_0 = const()[name = string("op_879_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_879_cast_fp16_7 = split(axis = var_879_axis_0, split_sizes = tile_14, x = var_859_cast_fp16)[name = string("op_879_cast_fp16")];
            string var_889_equation_0 = const()[name = string("op_889_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_889_cast_fp16 = einsum(equation = var_889_equation_0, values = (var_870_cast_fp16_0, var_860_cast_fp16_0))[name = string("op_889_cast_fp16")];
            fp16 var_890_to_fp16 = const()[name = string("op_890_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_891_cast_fp16 = mul(x = var_889_cast_fp16, y = var_890_to_fp16)[name = string("op_891_cast_fp16")];
            string var_893_equation_0 = const()[name = string("op_893_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_893_cast_fp16 = einsum(equation = var_893_equation_0, values = (var_870_cast_fp16_1, var_860_cast_fp16_1))[name = string("op_893_cast_fp16")];
            fp16 var_894_to_fp16 = const()[name = string("op_894_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_895_cast_fp16 = mul(x = var_893_cast_fp16, y = var_894_to_fp16)[name = string("op_895_cast_fp16")];
            string var_897_equation_0 = const()[name = string("op_897_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_897_cast_fp16 = einsum(equation = var_897_equation_0, values = (var_870_cast_fp16_2, var_860_cast_fp16_2))[name = string("op_897_cast_fp16")];
            fp16 var_898_to_fp16 = const()[name = string("op_898_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_899_cast_fp16 = mul(x = var_897_cast_fp16, y = var_898_to_fp16)[name = string("op_899_cast_fp16")];
            string var_901_equation_0 = const()[name = string("op_901_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_901_cast_fp16 = einsum(equation = var_901_equation_0, values = (var_870_cast_fp16_3, var_860_cast_fp16_3))[name = string("op_901_cast_fp16")];
            fp16 var_902_to_fp16 = const()[name = string("op_902_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_903_cast_fp16 = mul(x = var_901_cast_fp16, y = var_902_to_fp16)[name = string("op_903_cast_fp16")];
            string var_905_equation_0 = const()[name = string("op_905_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_905_cast_fp16 = einsum(equation = var_905_equation_0, values = (var_870_cast_fp16_4, var_860_cast_fp16_4))[name = string("op_905_cast_fp16")];
            fp16 var_906_to_fp16 = const()[name = string("op_906_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_907_cast_fp16 = mul(x = var_905_cast_fp16, y = var_906_to_fp16)[name = string("op_907_cast_fp16")];
            string var_909_equation_0 = const()[name = string("op_909_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_909_cast_fp16 = einsum(equation = var_909_equation_0, values = (var_870_cast_fp16_5, var_860_cast_fp16_5))[name = string("op_909_cast_fp16")];
            fp16 var_910_to_fp16 = const()[name = string("op_910_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_911_cast_fp16 = mul(x = var_909_cast_fp16, y = var_910_to_fp16)[name = string("op_911_cast_fp16")];
            string var_913_equation_0 = const()[name = string("op_913_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_913_cast_fp16 = einsum(equation = var_913_equation_0, values = (var_870_cast_fp16_6, var_860_cast_fp16_6))[name = string("op_913_cast_fp16")];
            fp16 var_914_to_fp16 = const()[name = string("op_914_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = string("op_915_cast_fp16")];
            string var_917_equation_0 = const()[name = string("op_917_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_917_cast_fp16 = einsum(equation = var_917_equation_0, values = (var_870_cast_fp16_7, var_860_cast_fp16_7))[name = string("op_917_cast_fp16")];
            fp16 var_918_to_fp16 = const()[name = string("op_918_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_919_cast_fp16 = mul(x = var_917_cast_fp16, y = var_918_to_fp16)[name = string("op_919_cast_fp16")];
            bool attn_weights_8_interleave_0 = const()[name = string("attn_weights_8_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_8_cast_fp16 = concat(axis = var_47, interleave = attn_weights_8_interleave_0, values = (var_891_cast_fp16, var_895_cast_fp16, var_899_cast_fp16, var_903_cast_fp16, var_907_cast_fp16, var_911_cast_fp16, var_915_cast_fp16, var_919_cast_fp16))[name = string("attn_weights_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_8_cast_fp16 = add(x = attn_weights_8_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_43_cast_fp16 = softmax(axis = var_46, x = attn_weights0_8_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<int32, [8]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_925_axis_0 = const()[name = string("op_925_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_925_cast_fp16_7 = split(axis = var_925_axis_0, split_sizes = tile_15, x = input_43_cast_fp16)[name = string("op_925_cast_fp16")];
            string var_935_equation_0 = const()[name = string("op_935_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_935_cast_fp16 = einsum(equation = var_935_equation_0, values = (var_879_cast_fp16_0, var_925_cast_fp16_0))[name = string("op_935_cast_fp16")];
            string var_937_equation_0 = const()[name = string("op_937_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_937_cast_fp16 = einsum(equation = var_937_equation_0, values = (var_879_cast_fp16_1, var_925_cast_fp16_1))[name = string("op_937_cast_fp16")];
            string var_939_equation_0 = const()[name = string("op_939_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_939_cast_fp16 = einsum(equation = var_939_equation_0, values = (var_879_cast_fp16_2, var_925_cast_fp16_2))[name = string("op_939_cast_fp16")];
            string var_941_equation_0 = const()[name = string("op_941_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_941_cast_fp16 = einsum(equation = var_941_equation_0, values = (var_879_cast_fp16_3, var_925_cast_fp16_3))[name = string("op_941_cast_fp16")];
            string var_943_equation_0 = const()[name = string("op_943_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16 = einsum(equation = var_943_equation_0, values = (var_879_cast_fp16_4, var_925_cast_fp16_4))[name = string("op_943_cast_fp16")];
            string var_945_equation_0 = const()[name = string("op_945_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_945_cast_fp16 = einsum(equation = var_945_equation_0, values = (var_879_cast_fp16_5, var_925_cast_fp16_5))[name = string("op_945_cast_fp16")];
            string var_947_equation_0 = const()[name = string("op_947_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_947_cast_fp16 = einsum(equation = var_947_equation_0, values = (var_879_cast_fp16_6, var_925_cast_fp16_6))[name = string("op_947_cast_fp16")];
            string var_949_equation_0 = const()[name = string("op_949_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_949_cast_fp16 = einsum(equation = var_949_equation_0, values = (var_879_cast_fp16_7, var_925_cast_fp16_7))[name = string("op_949_cast_fp16")];
            bool attn_23_interleave_0 = const()[name = string("attn_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_23_cast_fp16 = concat(axis = var_46, interleave = attn_23_interleave_0, values = (var_935_cast_fp16, var_937_cast_fp16, var_939_cast_fp16, var_941_cast_fp16, var_943_cast_fp16, var_945_cast_fp16, var_947_cast_fp16, var_949_cast_fp16))[name = string("attn_23_cast_fp16")];
            tensor<int32, [2]> var_957 = const()[name = string("op_957"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_959 = const()[name = string("op_959"), val = tensor<int32, [2]>([1, 1])];
            string input0_23_pad_type_0 = const()[name = string("input0_23_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_23_pad_0 = const()[name = string("input0_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36197888))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36196800))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36460096)))];
            tensor<fp16, [1, 512, 1, 256]> input0_23_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16, dilations = var_959, groups = var_46, pad = input0_23_pad_0, pad_type = input0_23_pad_type_0, strides = var_957, weight = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_23_cast_fp16)[name = string("input0_23_cast_fp16")];
            tensor<int32, [2]> var_966 = const()[name = string("op_966"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_968 = const()[name = string("op_968"), val = tensor<int32, [2]>([1, 1])];
            string x_36_pad_type_0 = const()[name = string("x_36_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_36_pad_0 = const()[name = string("x_36_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54261440))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54261120))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54327040)))];
            tensor<fp16, [1, 128, 1, 256]> x_36_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16, dilations = var_968, groups = var_46, pad = x_36_pad_0, pad_type = x_36_pad_type_0, strides = var_966, weight = nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_23_cast_fp16)[name = string("x_36_cast_fp16")];
            fp16 var_971_to_fp16 = const()[name = string("op_971_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_972_cast_fp16 = mul(x = x_36_cast_fp16, y = var_971_to_fp16)[name = string("op_972_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_973_cast_fp16 = sigmoid(x = var_972_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_45_cast_fp16 = mul(x = x_36_cast_fp16, y = var_973_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [2]> var_977 = const()[name = string("op_977"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_979 = const()[name = string("op_979"), val = tensor<int32, [2]>([1, 1])];
            string x_38_pad_type_0 = const()[name = string("x_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_38_pad_0 = const()[name = string("x_38_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54328448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54327360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54394048)))];
            tensor<fp16, [1, 512, 1, 256]> x_38_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16, dilations = var_979, groups = var_46, pad = x_38_pad_0, pad_type = x_38_pad_type_0, strides = var_977, weight = nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_45_cast_fp16)[name = string("x_38_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_25_cast_fp16 = add(x = x_38_cast_fp16, y = input0_23_cast_fp16)[name = string("attn_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_8_cast_fp16 = add(x = inputs1_1_cast_fp16, y = attn_25_cast_fp16)[name = string("inputs0_8_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_47_gamma_0_to_fp16 = const()[name = string("input_47_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54395136)))];
            tensor<fp16, [512]> input_47_beta_0_to_fp16 = const()[name = string("input_47_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54396224)))];
            fp16 var_992_to_fp16 = const()[name = string("op_992_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_992_to_fp16, gamma = input_47_gamma_0_to_fp16, x = inputs0_8_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<int32, [2]> var_1006 = const()[name = string("op_1006"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1008 = const()[name = string("op_1008"), val = tensor<int32, [2]>([1, 1])];
            string x_40_pad_type_0 = const()[name = string("x_40_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_40_pad_0 = const()[name = string("x_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36467520))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36463360))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37520320))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37516160))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_40_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1008, groups = var_46, pad = x_40_pad_0, pad_type = x_40_pad_type_0, strides = var_1006, weight = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_47_cast_fp16)[name = string("x_40_cast_fp16")];
            fp16 var_1011_to_fp16 = const()[name = string("op_1011_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1012_cast_fp16 = mul(x = x_40_cast_fp16, y = var_1011_to_fp16)[name = string("op_1012_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1013_cast_fp16 = sigmoid(x = var_1012_cast_fp16)[name = string("op_1013_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_49_cast_fp16 = mul(x = x_40_cast_fp16, y = var_1013_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<int32, [2]> var_1017 = const()[name = string("op_1017"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1019 = const()[name = string("op_1019"), val = tensor<int32, [2]>([1, 1])];
            string input0_25_pad_type_0 = const()[name = string("input0_25_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_25_pad_0 = const()[name = string("input0_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37523520))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37522432))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38572160)))];
            tensor<fp16, [1, 512, 1, 256]> input0_25_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16, dilations = var_1019, groups = var_46, pad = input0_25_pad_0, pad_type = input0_25_pad_type_0, strides = var_1017, weight = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_49_cast_fp16)[name = string("input0_25_cast_fp16")];
            tensor<int32, [2]> var_1027 = const()[name = string("op_1027"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1029 = const()[name = string("op_1029"), val = tensor<int32, [2]>([1, 1])];
            string x_42_pad_type_0 = const()[name = string("x_42_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_42_pad_0 = const()[name = string("x_42_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54397632))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54397312))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54463232)))];
            tensor<fp16, [1, 128, 1, 256]> x_42_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1029, groups = var_46, pad = x_42_pad_0, pad_type = x_42_pad_type_0, strides = var_1027, weight = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_25_cast_fp16)[name = string("x_42_cast_fp16")];
            fp16 var_1032_to_fp16 = const()[name = string("op_1032_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1033_cast_fp16 = mul(x = x_42_cast_fp16, y = var_1032_to_fp16)[name = string("op_1033_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1034_cast_fp16 = sigmoid(x = var_1033_cast_fp16)[name = string("op_1034_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_53_cast_fp16 = mul(x = x_42_cast_fp16, y = var_1034_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<int32, [2]> var_1038 = const()[name = string("op_1038"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1040 = const()[name = string("op_1040"), val = tensor<int32, [2]>([1, 1])];
            string x_44_pad_type_0 = const()[name = string("x_44_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_44_pad_0 = const()[name = string("x_44_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54464640))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54463552))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54530240)))];
            tensor<fp16, [1, 512, 1, 256]> x_44_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1040, groups = var_46, pad = x_44_pad_0, pad_type = x_44_pad_type_0, strides = var_1038, weight = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_53_cast_fp16)[name = string("x_44_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_8_cast_fp16 = add(x = x_44_cast_fp16, y = input0_25_cast_fp16)[name = string("f_8_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_8_cast_fp16 = add(x = f_8_cast_fp16, y = inputs0_8_cast_fp16)[name = string("x1_8_cast_fp16")];
            fp16 var_1045_to_fp16 = const()[name = string("op_1045_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1046_cast_fp16 = mul(x = inputs1_1_cast_fp16, y = var_1045_to_fp16)[name = string("op_1046_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs2_1_cast_fp16 = add(x = var_1046_cast_fp16, y = x1_8_cast_fp16)[name = string("inputs2_1_cast_fp16")];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_19_gamma_0_to_fp16 = const()[name = string("k_19_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54531328)))];
            tensor<fp16, [512]> k_19_beta_0_to_fp16 = const()[name = string("k_19_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54532416)))];
            fp16 var_1064_to_fp16 = const()[name = string("op_1064_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_19_cast_fp16 = layer_norm(axes = k_19_axes_0, beta = k_19_beta_0_to_fp16, epsilon = var_1064_to_fp16, gamma = k_19_gamma_0_to_fp16, x = inputs2_1_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [2]> var_1083 = const()[name = string("op_1083"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1085 = const()[name = string("op_1085"), val = tensor<int32, [2]>([1, 1])];
            string var_1087_pad_type_0 = const()[name = string("op_1087_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1087_pad_0 = const()[name = string("op_1087_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38576512))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38575424))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38838720)))];
            tensor<fp16, [1, 512, 1, 256]> var_1087_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16, dilations = var_1085, groups = var_46, pad = var_1087_pad_0, pad_type = var_1087_pad_type_0, strides = var_1083, weight = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_1087_cast_fp16")];
            tensor<int32, [2]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1092 = const()[name = string("op_1092"), val = tensor<int32, [2]>([1, 1])];
            string k_21_pad_type_0 = const()[name = string("k_21_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_21_pad_0 = const()[name = string("k_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38840896))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38839808))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39103104)))];
            tensor<fp16, [1, 512, 1, 256]> k_21_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16, dilations = var_1092, groups = var_46, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = var_1090, weight = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [2]> var_1097 = const()[name = string("op_1097"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1099 = const()[name = string("op_1099"), val = tensor<int32, [2]>([1, 1])];
            string var_1101_pad_type_0 = const()[name = string("op_1101_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1101_pad_0 = const()[name = string("op_1101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39105280))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39104192))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39367488)))];
            tensor<fp16, [1, 512, 1, 256]> var_1101_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16, dilations = var_1099, groups = var_46, pad = var_1101_pad_0, pad_type = var_1101_pad_type_0, strides = var_1097, weight = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_1101_cast_fp16")];
            tensor<int32, [8]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1102_axis_0 = const()[name = string("op_1102_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1102_cast_fp16_7 = split(axis = var_1102_axis_0, split_sizes = tile_16, x = var_1087_cast_fp16)[name = string("op_1102_cast_fp16")];
            tensor<int32, [4]> var_1111_perm_0 = const()[name = string("op_1111_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1112_axis_0 = const()[name = string("op_1112_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_4 = transpose(perm = var_1111_perm_0, x = k_21_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1112_cast_fp16_7 = split(axis = var_1112_axis_0, split_sizes = tile_17, x = transpose_4)[name = string("op_1112_cast_fp16")];
            tensor<int32, [8]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1121_axis_0 = const()[name = string("op_1121_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1121_cast_fp16_7 = split(axis = var_1121_axis_0, split_sizes = tile_18, x = var_1101_cast_fp16)[name = string("op_1121_cast_fp16")];
            string var_1131_equation_0 = const()[name = string("op_1131_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1131_cast_fp16 = einsum(equation = var_1131_equation_0, values = (var_1112_cast_fp16_0, var_1102_cast_fp16_0))[name = string("op_1131_cast_fp16")];
            fp16 var_1132_to_fp16 = const()[name = string("op_1132_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1133_cast_fp16 = mul(x = var_1131_cast_fp16, y = var_1132_to_fp16)[name = string("op_1133_cast_fp16")];
            string var_1135_equation_0 = const()[name = string("op_1135_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1135_cast_fp16 = einsum(equation = var_1135_equation_0, values = (var_1112_cast_fp16_1, var_1102_cast_fp16_1))[name = string("op_1135_cast_fp16")];
            fp16 var_1136_to_fp16 = const()[name = string("op_1136_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1137_cast_fp16 = mul(x = var_1135_cast_fp16, y = var_1136_to_fp16)[name = string("op_1137_cast_fp16")];
            string var_1139_equation_0 = const()[name = string("op_1139_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1139_cast_fp16 = einsum(equation = var_1139_equation_0, values = (var_1112_cast_fp16_2, var_1102_cast_fp16_2))[name = string("op_1139_cast_fp16")];
            fp16 var_1140_to_fp16 = const()[name = string("op_1140_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1141_cast_fp16 = mul(x = var_1139_cast_fp16, y = var_1140_to_fp16)[name = string("op_1141_cast_fp16")];
            string var_1143_equation_0 = const()[name = string("op_1143_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1143_cast_fp16 = einsum(equation = var_1143_equation_0, values = (var_1112_cast_fp16_3, var_1102_cast_fp16_3))[name = string("op_1143_cast_fp16")];
            fp16 var_1144_to_fp16 = const()[name = string("op_1144_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1145_cast_fp16 = mul(x = var_1143_cast_fp16, y = var_1144_to_fp16)[name = string("op_1145_cast_fp16")];
            string var_1147_equation_0 = const()[name = string("op_1147_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1147_cast_fp16 = einsum(equation = var_1147_equation_0, values = (var_1112_cast_fp16_4, var_1102_cast_fp16_4))[name = string("op_1147_cast_fp16")];
            fp16 var_1148_to_fp16 = const()[name = string("op_1148_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1149_cast_fp16 = mul(x = var_1147_cast_fp16, y = var_1148_to_fp16)[name = string("op_1149_cast_fp16")];
            string var_1151_equation_0 = const()[name = string("op_1151_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1151_cast_fp16 = einsum(equation = var_1151_equation_0, values = (var_1112_cast_fp16_5, var_1102_cast_fp16_5))[name = string("op_1151_cast_fp16")];
            fp16 var_1152_to_fp16 = const()[name = string("op_1152_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = string("op_1153_cast_fp16")];
            string var_1155_equation_0 = const()[name = string("op_1155_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1155_cast_fp16 = einsum(equation = var_1155_equation_0, values = (var_1112_cast_fp16_6, var_1102_cast_fp16_6))[name = string("op_1155_cast_fp16")];
            fp16 var_1156_to_fp16 = const()[name = string("op_1156_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1157_cast_fp16 = mul(x = var_1155_cast_fp16, y = var_1156_to_fp16)[name = string("op_1157_cast_fp16")];
            string var_1159_equation_0 = const()[name = string("op_1159_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1159_cast_fp16 = einsum(equation = var_1159_equation_0, values = (var_1112_cast_fp16_7, var_1102_cast_fp16_7))[name = string("op_1159_cast_fp16")];
            fp16 var_1160_to_fp16 = const()[name = string("op_1160_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1161_cast_fp16 = mul(x = var_1159_cast_fp16, y = var_1160_to_fp16)[name = string("op_1161_cast_fp16")];
            bool attn_weights_10_interleave_0 = const()[name = string("attn_weights_10_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_10_cast_fp16 = concat(axis = var_47, interleave = attn_weights_10_interleave_0, values = (var_1133_cast_fp16, var_1137_cast_fp16, var_1141_cast_fp16, var_1145_cast_fp16, var_1149_cast_fp16, var_1153_cast_fp16, var_1157_cast_fp16, var_1161_cast_fp16))[name = string("attn_weights_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_10_cast_fp16 = add(x = attn_weights_10_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_55_cast_fp16 = softmax(axis = var_46, x = attn_weights0_10_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<int32, [8]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1167_axis_0 = const()[name = string("op_1167_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16_7 = split(axis = var_1167_axis_0, split_sizes = tile_19, x = input_55_cast_fp16)[name = string("op_1167_cast_fp16")];
            string var_1177_equation_0 = const()[name = string("op_1177_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1177_cast_fp16 = einsum(equation = var_1177_equation_0, values = (var_1121_cast_fp16_0, var_1167_cast_fp16_0))[name = string("op_1177_cast_fp16")];
            string var_1179_equation_0 = const()[name = string("op_1179_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1179_cast_fp16 = einsum(equation = var_1179_equation_0, values = (var_1121_cast_fp16_1, var_1167_cast_fp16_1))[name = string("op_1179_cast_fp16")];
            string var_1181_equation_0 = const()[name = string("op_1181_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1181_cast_fp16 = einsum(equation = var_1181_equation_0, values = (var_1121_cast_fp16_2, var_1167_cast_fp16_2))[name = string("op_1181_cast_fp16")];
            string var_1183_equation_0 = const()[name = string("op_1183_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1183_cast_fp16 = einsum(equation = var_1183_equation_0, values = (var_1121_cast_fp16_3, var_1167_cast_fp16_3))[name = string("op_1183_cast_fp16")];
            string var_1185_equation_0 = const()[name = string("op_1185_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1185_cast_fp16 = einsum(equation = var_1185_equation_0, values = (var_1121_cast_fp16_4, var_1167_cast_fp16_4))[name = string("op_1185_cast_fp16")];
            string var_1187_equation_0 = const()[name = string("op_1187_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1187_cast_fp16 = einsum(equation = var_1187_equation_0, values = (var_1121_cast_fp16_5, var_1167_cast_fp16_5))[name = string("op_1187_cast_fp16")];
            string var_1189_equation_0 = const()[name = string("op_1189_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1189_cast_fp16 = einsum(equation = var_1189_equation_0, values = (var_1121_cast_fp16_6, var_1167_cast_fp16_6))[name = string("op_1189_cast_fp16")];
            string var_1191_equation_0 = const()[name = string("op_1191_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1191_cast_fp16 = einsum(equation = var_1191_equation_0, values = (var_1121_cast_fp16_7, var_1167_cast_fp16_7))[name = string("op_1191_cast_fp16")];
            bool attn_29_interleave_0 = const()[name = string("attn_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_29_cast_fp16 = concat(axis = var_46, interleave = attn_29_interleave_0, values = (var_1177_cast_fp16, var_1179_cast_fp16, var_1181_cast_fp16, var_1183_cast_fp16, var_1185_cast_fp16, var_1187_cast_fp16, var_1189_cast_fp16, var_1191_cast_fp16))[name = string("attn_29_cast_fp16")];
            tensor<int32, [2]> var_1199 = const()[name = string("op_1199"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1201 = const()[name = string("op_1201"), val = tensor<int32, [2]>([1, 1])];
            string input0_29_pad_type_0 = const()[name = string("input0_29_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_29_pad_0 = const()[name = string("input0_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39369664))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39368576))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39631872)))];
            tensor<fp16, [1, 512, 1, 256]> input0_29_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16, dilations = var_1201, groups = var_46, pad = input0_29_pad_0, pad_type = input0_29_pad_type_0, strides = var_1199, weight = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_29_cast_fp16)[name = string("input0_29_cast_fp16")];
            tensor<int32, [2]> var_1208 = const()[name = string("op_1208"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1210 = const()[name = string("op_1210"), val = tensor<int32, [2]>([1, 1])];
            string x_46_pad_type_0 = const()[name = string("x_46_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_46_pad_0 = const()[name = string("x_46_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54533824))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54533504))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54599424)))];
            tensor<fp16, [1, 128, 1, 256]> x_46_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1210, groups = var_46, pad = x_46_pad_0, pad_type = x_46_pad_type_0, strides = var_1208, weight = nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_29_cast_fp16)[name = string("x_46_cast_fp16")];
            fp16 var_1213_to_fp16 = const()[name = string("op_1213_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1214_cast_fp16 = mul(x = x_46_cast_fp16, y = var_1213_to_fp16)[name = string("op_1214_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1215_cast_fp16 = sigmoid(x = var_1214_cast_fp16)[name = string("op_1215_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_57_cast_fp16 = mul(x = x_46_cast_fp16, y = var_1215_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [2]> var_1219 = const()[name = string("op_1219"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [2]>([1, 1])];
            string x_48_pad_type_0 = const()[name = string("x_48_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_48_pad_0 = const()[name = string("x_48_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54600832))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54599744))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54666432)))];
            tensor<fp16, [1, 512, 1, 256]> x_48_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1221, groups = var_46, pad = x_48_pad_0, pad_type = x_48_pad_type_0, strides = var_1219, weight = nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_57_cast_fp16)[name = string("x_48_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_31_cast_fp16 = add(x = x_48_cast_fp16, y = input0_29_cast_fp16)[name = string("attn_31_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_10_cast_fp16 = add(x = inputs2_1_cast_fp16, y = attn_31_cast_fp16)[name = string("inputs0_10_cast_fp16")];
            tensor<int32, [1]> input_59_axes_0 = const()[name = string("input_59_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_59_gamma_0_to_fp16 = const()[name = string("input_59_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54667520)))];
            tensor<fp16, [512]> input_59_beta_0_to_fp16 = const()[name = string("input_59_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54668608)))];
            fp16 var_1234_to_fp16 = const()[name = string("op_1234_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = input_59_beta_0_to_fp16, epsilon = var_1234_to_fp16, gamma = input_59_gamma_0_to_fp16, x = inputs0_10_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<int32, [2]> var_1248 = const()[name = string("op_1248"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1250 = const()[name = string("op_1250"), val = tensor<int32, [2]>([1, 1])];
            string x_50_pad_type_0 = const()[name = string("x_50_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_50_pad_0 = const()[name = string("x_50_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39639296))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39635136))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40692096))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40687936))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_50_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1250, groups = var_46, pad = x_50_pad_0, pad_type = x_50_pad_type_0, strides = var_1248, weight = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_59_cast_fp16)[name = string("x_50_cast_fp16")];
            fp16 var_1253_to_fp16 = const()[name = string("op_1253_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1254_cast_fp16 = mul(x = x_50_cast_fp16, y = var_1253_to_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1255_cast_fp16 = sigmoid(x = var_1254_cast_fp16)[name = string("op_1255_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_61_cast_fp16 = mul(x = x_50_cast_fp16, y = var_1255_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<int32, [2]> var_1259 = const()[name = string("op_1259"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1261 = const()[name = string("op_1261"), val = tensor<int32, [2]>([1, 1])];
            string input0_31_pad_type_0 = const()[name = string("input0_31_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_31_pad_0 = const()[name = string("input0_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40695296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40694208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41743936)))];
            tensor<fp16, [1, 512, 1, 256]> input0_31_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16, dilations = var_1261, groups = var_46, pad = input0_31_pad_0, pad_type = input0_31_pad_type_0, strides = var_1259, weight = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_61_cast_fp16)[name = string("input0_31_cast_fp16")];
            tensor<int32, [2]> var_1269 = const()[name = string("op_1269"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1271 = const()[name = string("op_1271"), val = tensor<int32, [2]>([1, 1])];
            string x_52_pad_type_0 = const()[name = string("x_52_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_52_pad_0 = const()[name = string("x_52_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54670016))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54669696))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54735616)))];
            tensor<fp16, [1, 128, 1, 256]> x_52_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1271, groups = var_46, pad = x_52_pad_0, pad_type = x_52_pad_type_0, strides = var_1269, weight = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_31_cast_fp16)[name = string("x_52_cast_fp16")];
            fp16 var_1274_to_fp16 = const()[name = string("op_1274_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1275_cast_fp16 = mul(x = x_52_cast_fp16, y = var_1274_to_fp16)[name = string("op_1275_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1276_cast_fp16 = sigmoid(x = var_1275_cast_fp16)[name = string("op_1276_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_65_cast_fp16 = mul(x = x_52_cast_fp16, y = var_1276_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [2]> var_1280 = const()[name = string("op_1280"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1282 = const()[name = string("op_1282"), val = tensor<int32, [2]>([1, 1])];
            string x_54_pad_type_0 = const()[name = string("x_54_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_54_pad_0 = const()[name = string("x_54_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54737024))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54735936))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54802624)))];
            tensor<fp16, [1, 512, 1, 256]> x_54_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1282, groups = var_46, pad = x_54_pad_0, pad_type = x_54_pad_type_0, strides = var_1280, weight = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_65_cast_fp16)[name = string("x_54_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_10_cast_fp16 = add(x = x_54_cast_fp16, y = input0_31_cast_fp16)[name = string("f_10_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_10_cast_fp16 = add(x = f_10_cast_fp16, y = inputs0_10_cast_fp16)[name = string("x1_10_cast_fp16")];
            fp16 var_1287_to_fp16 = const()[name = string("op_1287_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1288_cast_fp16 = mul(x = inputs2_1_cast_fp16, y = var_1287_to_fp16)[name = string("op_1288_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs3_1_cast_fp16 = add(x = var_1288_cast_fp16, y = x1_10_cast_fp16)[name = string("inputs3_1_cast_fp16")];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_23_gamma_0_to_fp16 = const()[name = string("k_23_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54803712)))];
            tensor<fp16, [512]> k_23_beta_0_to_fp16 = const()[name = string("k_23_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54804800)))];
            fp16 var_1306_to_fp16 = const()[name = string("op_1306_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_23_cast_fp16 = layer_norm(axes = k_23_axes_0, beta = k_23_beta_0_to_fp16, epsilon = var_1306_to_fp16, gamma = k_23_gamma_0_to_fp16, x = inputs3_1_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [2]> var_1325 = const()[name = string("op_1325"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1327 = const()[name = string("op_1327"), val = tensor<int32, [2]>([1, 1])];
            string var_1329_pad_type_0 = const()[name = string("op_1329_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1329_pad_0 = const()[name = string("op_1329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41748288))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41747200))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42010496)))];
            tensor<fp16, [1, 512, 1, 256]> var_1329_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16, dilations = var_1327, groups = var_46, pad = var_1329_pad_0, pad_type = var_1329_pad_type_0, strides = var_1325, weight = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1329_cast_fp16")];
            tensor<int32, [2]> var_1332 = const()[name = string("op_1332"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1334 = const()[name = string("op_1334"), val = tensor<int32, [2]>([1, 1])];
            string k_25_pad_type_0 = const()[name = string("k_25_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_25_pad_0 = const()[name = string("k_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42012672))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42011584))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42274880)))];
            tensor<fp16, [1, 512, 1, 256]> k_25_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16, dilations = var_1334, groups = var_46, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = var_1332, weight = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [2]> var_1339 = const()[name = string("op_1339"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1341 = const()[name = string("op_1341"), val = tensor<int32, [2]>([1, 1])];
            string var_1343_pad_type_0 = const()[name = string("op_1343_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1343_pad_0 = const()[name = string("op_1343_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42277056))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42275968))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42539264)))];
            tensor<fp16, [1, 512, 1, 256]> var_1343_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16, dilations = var_1341, groups = var_46, pad = var_1343_pad_0, pad_type = var_1343_pad_type_0, strides = var_1339, weight = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1343_cast_fp16")];
            tensor<int32, [8]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1344_axis_0 = const()[name = string("op_1344_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1344_cast_fp16_7 = split(axis = var_1344_axis_0, split_sizes = tile_20, x = var_1329_cast_fp16)[name = string("op_1344_cast_fp16")];
            tensor<int32, [4]> var_1353_perm_0 = const()[name = string("op_1353_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1354_axis_0 = const()[name = string("op_1354_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_3 = transpose(perm = var_1353_perm_0, x = k_25_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1354_cast_fp16_7 = split(axis = var_1354_axis_0, split_sizes = tile_21, x = transpose_3)[name = string("op_1354_cast_fp16")];
            tensor<int32, [8]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1363_axis_0 = const()[name = string("op_1363_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1363_cast_fp16_7 = split(axis = var_1363_axis_0, split_sizes = tile_22, x = var_1343_cast_fp16)[name = string("op_1363_cast_fp16")];
            string var_1373_equation_0 = const()[name = string("op_1373_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1373_cast_fp16 = einsum(equation = var_1373_equation_0, values = (var_1354_cast_fp16_0, var_1344_cast_fp16_0))[name = string("op_1373_cast_fp16")];
            fp16 var_1374_to_fp16 = const()[name = string("op_1374_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1375_cast_fp16 = mul(x = var_1373_cast_fp16, y = var_1374_to_fp16)[name = string("op_1375_cast_fp16")];
            string var_1377_equation_0 = const()[name = string("op_1377_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1377_cast_fp16 = einsum(equation = var_1377_equation_0, values = (var_1354_cast_fp16_1, var_1344_cast_fp16_1))[name = string("op_1377_cast_fp16")];
            fp16 var_1378_to_fp16 = const()[name = string("op_1378_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1379_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = string("op_1379_cast_fp16")];
            string var_1381_equation_0 = const()[name = string("op_1381_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1381_cast_fp16 = einsum(equation = var_1381_equation_0, values = (var_1354_cast_fp16_2, var_1344_cast_fp16_2))[name = string("op_1381_cast_fp16")];
            fp16 var_1382_to_fp16 = const()[name = string("op_1382_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1383_cast_fp16 = mul(x = var_1381_cast_fp16, y = var_1382_to_fp16)[name = string("op_1383_cast_fp16")];
            string var_1385_equation_0 = const()[name = string("op_1385_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16 = einsum(equation = var_1385_equation_0, values = (var_1354_cast_fp16_3, var_1344_cast_fp16_3))[name = string("op_1385_cast_fp16")];
            fp16 var_1386_to_fp16 = const()[name = string("op_1386_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1387_cast_fp16 = mul(x = var_1385_cast_fp16, y = var_1386_to_fp16)[name = string("op_1387_cast_fp16")];
            string var_1389_equation_0 = const()[name = string("op_1389_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1389_cast_fp16 = einsum(equation = var_1389_equation_0, values = (var_1354_cast_fp16_4, var_1344_cast_fp16_4))[name = string("op_1389_cast_fp16")];
            fp16 var_1390_to_fp16 = const()[name = string("op_1390_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1391_cast_fp16 = mul(x = var_1389_cast_fp16, y = var_1390_to_fp16)[name = string("op_1391_cast_fp16")];
            string var_1393_equation_0 = const()[name = string("op_1393_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1393_cast_fp16 = einsum(equation = var_1393_equation_0, values = (var_1354_cast_fp16_5, var_1344_cast_fp16_5))[name = string("op_1393_cast_fp16")];
            fp16 var_1394_to_fp16 = const()[name = string("op_1394_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1395_cast_fp16 = mul(x = var_1393_cast_fp16, y = var_1394_to_fp16)[name = string("op_1395_cast_fp16")];
            string var_1397_equation_0 = const()[name = string("op_1397_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1397_cast_fp16 = einsum(equation = var_1397_equation_0, values = (var_1354_cast_fp16_6, var_1344_cast_fp16_6))[name = string("op_1397_cast_fp16")];
            fp16 var_1398_to_fp16 = const()[name = string("op_1398_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1399_cast_fp16 = mul(x = var_1397_cast_fp16, y = var_1398_to_fp16)[name = string("op_1399_cast_fp16")];
            string var_1401_equation_0 = const()[name = string("op_1401_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1401_cast_fp16 = einsum(equation = var_1401_equation_0, values = (var_1354_cast_fp16_7, var_1344_cast_fp16_7))[name = string("op_1401_cast_fp16")];
            fp16 var_1402_to_fp16 = const()[name = string("op_1402_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = string("op_1403_cast_fp16")];
            bool attn_weights_12_interleave_0 = const()[name = string("attn_weights_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_12_cast_fp16 = concat(axis = var_47, interleave = attn_weights_12_interleave_0, values = (var_1375_cast_fp16, var_1379_cast_fp16, var_1383_cast_fp16, var_1387_cast_fp16, var_1391_cast_fp16, var_1395_cast_fp16, var_1399_cast_fp16, var_1403_cast_fp16))[name = string("attn_weights_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_12_cast_fp16 = add(x = attn_weights_12_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_67_cast_fp16 = softmax(axis = var_46, x = attn_weights0_12_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<int32, [8]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1409_axis_0 = const()[name = string("op_1409_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1409_cast_fp16_7 = split(axis = var_1409_axis_0, split_sizes = tile_23, x = input_67_cast_fp16)[name = string("op_1409_cast_fp16")];
            string var_1419_equation_0 = const()[name = string("op_1419_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1419_cast_fp16 = einsum(equation = var_1419_equation_0, values = (var_1363_cast_fp16_0, var_1409_cast_fp16_0))[name = string("op_1419_cast_fp16")];
            string var_1421_equation_0 = const()[name = string("op_1421_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1421_cast_fp16 = einsum(equation = var_1421_equation_0, values = (var_1363_cast_fp16_1, var_1409_cast_fp16_1))[name = string("op_1421_cast_fp16")];
            string var_1423_equation_0 = const()[name = string("op_1423_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1423_cast_fp16 = einsum(equation = var_1423_equation_0, values = (var_1363_cast_fp16_2, var_1409_cast_fp16_2))[name = string("op_1423_cast_fp16")];
            string var_1425_equation_0 = const()[name = string("op_1425_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1425_cast_fp16 = einsum(equation = var_1425_equation_0, values = (var_1363_cast_fp16_3, var_1409_cast_fp16_3))[name = string("op_1425_cast_fp16")];
            string var_1427_equation_0 = const()[name = string("op_1427_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1427_cast_fp16 = einsum(equation = var_1427_equation_0, values = (var_1363_cast_fp16_4, var_1409_cast_fp16_4))[name = string("op_1427_cast_fp16")];
            string var_1429_equation_0 = const()[name = string("op_1429_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1429_cast_fp16 = einsum(equation = var_1429_equation_0, values = (var_1363_cast_fp16_5, var_1409_cast_fp16_5))[name = string("op_1429_cast_fp16")];
            string var_1431_equation_0 = const()[name = string("op_1431_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1431_cast_fp16 = einsum(equation = var_1431_equation_0, values = (var_1363_cast_fp16_6, var_1409_cast_fp16_6))[name = string("op_1431_cast_fp16")];
            string var_1433_equation_0 = const()[name = string("op_1433_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1433_cast_fp16 = einsum(equation = var_1433_equation_0, values = (var_1363_cast_fp16_7, var_1409_cast_fp16_7))[name = string("op_1433_cast_fp16")];
            bool attn_35_interleave_0 = const()[name = string("attn_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_35_cast_fp16 = concat(axis = var_46, interleave = attn_35_interleave_0, values = (var_1419_cast_fp16, var_1421_cast_fp16, var_1423_cast_fp16, var_1425_cast_fp16, var_1427_cast_fp16, var_1429_cast_fp16, var_1431_cast_fp16, var_1433_cast_fp16))[name = string("attn_35_cast_fp16")];
            tensor<int32, [2]> var_1441 = const()[name = string("op_1441"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1443 = const()[name = string("op_1443"), val = tensor<int32, [2]>([1, 1])];
            string input0_35_pad_type_0 = const()[name = string("input0_35_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_35_pad_0 = const()[name = string("input0_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42541440))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42540352))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42803648)))];
            tensor<fp16, [1, 512, 1, 256]> input0_35_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16, dilations = var_1443, groups = var_46, pad = input0_35_pad_0, pad_type = input0_35_pad_type_0, strides = var_1441, weight = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_35_cast_fp16)[name = string("input0_35_cast_fp16")];
            tensor<int32, [2]> var_1450 = const()[name = string("op_1450"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [2]>([1, 1])];
            string x_56_pad_type_0 = const()[name = string("x_56_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_56_pad_0 = const()[name = string("x_56_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54806208))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54805888))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54871808)))];
            tensor<fp16, [1, 128, 1, 256]> x_56_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1452, groups = var_46, pad = x_56_pad_0, pad_type = x_56_pad_type_0, strides = var_1450, weight = nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_35_cast_fp16)[name = string("x_56_cast_fp16")];
            fp16 var_1455_to_fp16 = const()[name = string("op_1455_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1456_cast_fp16 = mul(x = x_56_cast_fp16, y = var_1455_to_fp16)[name = string("op_1456_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1457_cast_fp16 = sigmoid(x = var_1456_cast_fp16)[name = string("op_1457_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_69_cast_fp16 = mul(x = x_56_cast_fp16, y = var_1457_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<int32, [2]> var_1461 = const()[name = string("op_1461"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1463 = const()[name = string("op_1463"), val = tensor<int32, [2]>([1, 1])];
            string x_58_pad_type_0 = const()[name = string("x_58_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_58_pad_0 = const()[name = string("x_58_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54873216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54872128))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54938816)))];
            tensor<fp16, [1, 512, 1, 256]> x_58_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1463, groups = var_46, pad = x_58_pad_0, pad_type = x_58_pad_type_0, strides = var_1461, weight = nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_69_cast_fp16)[name = string("x_58_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_37_cast_fp16 = add(x = x_58_cast_fp16, y = input0_35_cast_fp16)[name = string("attn_37_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_12_cast_fp16 = add(x = inputs3_1_cast_fp16, y = attn_37_cast_fp16)[name = string("inputs0_12_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_71_gamma_0_to_fp16 = const()[name = string("input_71_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54939904)))];
            tensor<fp16, [512]> input_71_beta_0_to_fp16 = const()[name = string("input_71_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54940992)))];
            fp16 var_1476_to_fp16 = const()[name = string("op_1476_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = input_71_beta_0_to_fp16, epsilon = var_1476_to_fp16, gamma = input_71_gamma_0_to_fp16, x = inputs0_12_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<int32, [2]> var_1490 = const()[name = string("op_1490"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1492 = const()[name = string("op_1492"), val = tensor<int32, [2]>([1, 1])];
            string x_60_pad_type_0 = const()[name = string("x_60_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_60_pad_0 = const()[name = string("x_60_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42811072))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42806912))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43863872))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43859712))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_60_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1492, groups = var_46, pad = x_60_pad_0, pad_type = x_60_pad_type_0, strides = var_1490, weight = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_71_cast_fp16)[name = string("x_60_cast_fp16")];
            fp16 var_1495_to_fp16 = const()[name = string("op_1495_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1496_cast_fp16 = mul(x = x_60_cast_fp16, y = var_1495_to_fp16)[name = string("op_1496_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1497_cast_fp16 = sigmoid(x = var_1496_cast_fp16)[name = string("op_1497_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_73_cast_fp16 = mul(x = x_60_cast_fp16, y = var_1497_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<int32, [2]> var_1501 = const()[name = string("op_1501"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1503 = const()[name = string("op_1503"), val = tensor<int32, [2]>([1, 1])];
            string input0_37_pad_type_0 = const()[name = string("input0_37_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_37_pad_0 = const()[name = string("input0_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43867072))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43865984))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44915712)))];
            tensor<fp16, [1, 512, 1, 256]> input0_37_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16, dilations = var_1503, groups = var_46, pad = input0_37_pad_0, pad_type = input0_37_pad_type_0, strides = var_1501, weight = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_73_cast_fp16)[name = string("input0_37_cast_fp16")];
            tensor<int32, [2]> var_1511 = const()[name = string("op_1511"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1513 = const()[name = string("op_1513"), val = tensor<int32, [2]>([1, 1])];
            string x_62_pad_type_0 = const()[name = string("x_62_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_62_pad_0 = const()[name = string("x_62_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54942400))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(54942080))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55008000)))];
            tensor<fp16, [1, 128, 1, 256]> x_62_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1513, groups = var_46, pad = x_62_pad_0, pad_type = x_62_pad_type_0, strides = var_1511, weight = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_37_cast_fp16)[name = string("x_62_cast_fp16")];
            fp16 var_1516_to_fp16 = const()[name = string("op_1516_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1517_cast_fp16 = mul(x = x_62_cast_fp16, y = var_1516_to_fp16)[name = string("op_1517_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1518_cast_fp16 = sigmoid(x = var_1517_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_77_cast_fp16 = mul(x = x_62_cast_fp16, y = var_1518_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<int32, [2]> var_1522 = const()[name = string("op_1522"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [2]>([1, 1])];
            string x_64_pad_type_0 = const()[name = string("x_64_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_64_pad_0 = const()[name = string("x_64_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55009408))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55008320))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55075008)))];
            tensor<fp16, [1, 512, 1, 256]> x_64_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1524, groups = var_46, pad = x_64_pad_0, pad_type = x_64_pad_type_0, strides = var_1522, weight = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_77_cast_fp16)[name = string("x_64_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_12_cast_fp16 = add(x = x_64_cast_fp16, y = input0_37_cast_fp16)[name = string("f_12_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_12_cast_fp16 = add(x = f_12_cast_fp16, y = inputs0_12_cast_fp16)[name = string("x1_12_cast_fp16")];
            fp16 var_1529_to_fp16 = const()[name = string("op_1529_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1530_cast_fp16 = mul(x = inputs3_1_cast_fp16, y = var_1529_to_fp16)[name = string("op_1530_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs4_1_cast_fp16 = add(x = var_1530_cast_fp16, y = x1_12_cast_fp16)[name = string("inputs4_1_cast_fp16")];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_27_gamma_0_to_fp16 = const()[name = string("k_27_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55076096)))];
            tensor<fp16, [512]> k_27_beta_0_to_fp16 = const()[name = string("k_27_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55077184)))];
            fp16 var_1548_to_fp16 = const()[name = string("op_1548_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_27_cast_fp16 = layer_norm(axes = k_27_axes_0, beta = k_27_beta_0_to_fp16, epsilon = var_1548_to_fp16, gamma = k_27_gamma_0_to_fp16, x = inputs4_1_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [2]> var_1567 = const()[name = string("op_1567"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1569 = const()[name = string("op_1569"), val = tensor<int32, [2]>([1, 1])];
            string var_1571_pad_type_0 = const()[name = string("op_1571_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1571_pad_0 = const()[name = string("op_1571_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44920064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44918976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45182272)))];
            tensor<fp16, [1, 512, 1, 256]> var_1571_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16, dilations = var_1569, groups = var_46, pad = var_1571_pad_0, pad_type = var_1571_pad_type_0, strides = var_1567, weight = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1571_cast_fp16")];
            tensor<int32, [2]> var_1574 = const()[name = string("op_1574"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1576 = const()[name = string("op_1576"), val = tensor<int32, [2]>([1, 1])];
            string k_29_pad_type_0 = const()[name = string("k_29_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_29_pad_0 = const()[name = string("k_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45184448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45183360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45446656)))];
            tensor<fp16, [1, 512, 1, 256]> k_29_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16, dilations = var_1576, groups = var_46, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = var_1574, weight = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [2]> var_1581 = const()[name = string("op_1581"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1583 = const()[name = string("op_1583"), val = tensor<int32, [2]>([1, 1])];
            string var_1585_pad_type_0 = const()[name = string("op_1585_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1585_pad_0 = const()[name = string("op_1585_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45448832))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45447744))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45711040)))];
            tensor<fp16, [1, 512, 1, 256]> var_1585_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16, dilations = var_1583, groups = var_46, pad = var_1585_pad_0, pad_type = var_1585_pad_type_0, strides = var_1581, weight = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1585_cast_fp16")];
            tensor<int32, [8]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1586_axis_0 = const()[name = string("op_1586_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1586_cast_fp16_7 = split(axis = var_1586_axis_0, split_sizes = tile_24, x = var_1571_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<int32, [4]> var_1595_perm_0 = const()[name = string("op_1595_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1596_axis_0 = const()[name = string("op_1596_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_2 = transpose(perm = var_1595_perm_0, x = k_29_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1596_cast_fp16_7 = split(axis = var_1596_axis_0, split_sizes = tile_25, x = transpose_2)[name = string("op_1596_cast_fp16")];
            tensor<int32, [8]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1605_axis_0 = const()[name = string("op_1605_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16_7 = split(axis = var_1605_axis_0, split_sizes = tile_26, x = var_1585_cast_fp16)[name = string("op_1605_cast_fp16")];
            string var_1615_equation_0 = const()[name = string("op_1615_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1615_cast_fp16 = einsum(equation = var_1615_equation_0, values = (var_1596_cast_fp16_0, var_1586_cast_fp16_0))[name = string("op_1615_cast_fp16")];
            fp16 var_1616_to_fp16 = const()[name = string("op_1616_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1617_cast_fp16 = mul(x = var_1615_cast_fp16, y = var_1616_to_fp16)[name = string("op_1617_cast_fp16")];
            string var_1619_equation_0 = const()[name = string("op_1619_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1619_cast_fp16 = einsum(equation = var_1619_equation_0, values = (var_1596_cast_fp16_1, var_1586_cast_fp16_1))[name = string("op_1619_cast_fp16")];
            fp16 var_1620_to_fp16 = const()[name = string("op_1620_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1621_cast_fp16 = mul(x = var_1619_cast_fp16, y = var_1620_to_fp16)[name = string("op_1621_cast_fp16")];
            string var_1623_equation_0 = const()[name = string("op_1623_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1623_cast_fp16 = einsum(equation = var_1623_equation_0, values = (var_1596_cast_fp16_2, var_1586_cast_fp16_2))[name = string("op_1623_cast_fp16")];
            fp16 var_1624_to_fp16 = const()[name = string("op_1624_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1625_cast_fp16 = mul(x = var_1623_cast_fp16, y = var_1624_to_fp16)[name = string("op_1625_cast_fp16")];
            string var_1627_equation_0 = const()[name = string("op_1627_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1627_cast_fp16 = einsum(equation = var_1627_equation_0, values = (var_1596_cast_fp16_3, var_1586_cast_fp16_3))[name = string("op_1627_cast_fp16")];
            fp16 var_1628_to_fp16 = const()[name = string("op_1628_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1629_cast_fp16 = mul(x = var_1627_cast_fp16, y = var_1628_to_fp16)[name = string("op_1629_cast_fp16")];
            string var_1631_equation_0 = const()[name = string("op_1631_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1631_cast_fp16 = einsum(equation = var_1631_equation_0, values = (var_1596_cast_fp16_4, var_1586_cast_fp16_4))[name = string("op_1631_cast_fp16")];
            fp16 var_1632_to_fp16 = const()[name = string("op_1632_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = string("op_1633_cast_fp16")];
            string var_1635_equation_0 = const()[name = string("op_1635_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1635_cast_fp16 = einsum(equation = var_1635_equation_0, values = (var_1596_cast_fp16_5, var_1586_cast_fp16_5))[name = string("op_1635_cast_fp16")];
            fp16 var_1636_to_fp16 = const()[name = string("op_1636_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1637_cast_fp16 = mul(x = var_1635_cast_fp16, y = var_1636_to_fp16)[name = string("op_1637_cast_fp16")];
            string var_1639_equation_0 = const()[name = string("op_1639_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1639_cast_fp16 = einsum(equation = var_1639_equation_0, values = (var_1596_cast_fp16_6, var_1586_cast_fp16_6))[name = string("op_1639_cast_fp16")];
            fp16 var_1640_to_fp16 = const()[name = string("op_1640_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1641_cast_fp16 = mul(x = var_1639_cast_fp16, y = var_1640_to_fp16)[name = string("op_1641_cast_fp16")];
            string var_1643_equation_0 = const()[name = string("op_1643_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1643_cast_fp16 = einsum(equation = var_1643_equation_0, values = (var_1596_cast_fp16_7, var_1586_cast_fp16_7))[name = string("op_1643_cast_fp16")];
            fp16 var_1644_to_fp16 = const()[name = string("op_1644_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1645_cast_fp16 = mul(x = var_1643_cast_fp16, y = var_1644_to_fp16)[name = string("op_1645_cast_fp16")];
            bool attn_weights_14_interleave_0 = const()[name = string("attn_weights_14_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_14_cast_fp16 = concat(axis = var_47, interleave = attn_weights_14_interleave_0, values = (var_1617_cast_fp16, var_1621_cast_fp16, var_1625_cast_fp16, var_1629_cast_fp16, var_1633_cast_fp16, var_1637_cast_fp16, var_1641_cast_fp16, var_1645_cast_fp16))[name = string("attn_weights_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_14_cast_fp16 = add(x = attn_weights_14_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_79_cast_fp16 = softmax(axis = var_46, x = attn_weights0_14_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<int32, [8]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1651_axis_0 = const()[name = string("op_1651_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1651_cast_fp16_7 = split(axis = var_1651_axis_0, split_sizes = tile_27, x = input_79_cast_fp16)[name = string("op_1651_cast_fp16")];
            string var_1661_equation_0 = const()[name = string("op_1661_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1661_cast_fp16 = einsum(equation = var_1661_equation_0, values = (var_1605_cast_fp16_0, var_1651_cast_fp16_0))[name = string("op_1661_cast_fp16")];
            string var_1663_equation_0 = const()[name = string("op_1663_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1663_cast_fp16 = einsum(equation = var_1663_equation_0, values = (var_1605_cast_fp16_1, var_1651_cast_fp16_1))[name = string("op_1663_cast_fp16")];
            string var_1665_equation_0 = const()[name = string("op_1665_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1665_cast_fp16 = einsum(equation = var_1665_equation_0, values = (var_1605_cast_fp16_2, var_1651_cast_fp16_2))[name = string("op_1665_cast_fp16")];
            string var_1667_equation_0 = const()[name = string("op_1667_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1667_cast_fp16 = einsum(equation = var_1667_equation_0, values = (var_1605_cast_fp16_3, var_1651_cast_fp16_3))[name = string("op_1667_cast_fp16")];
            string var_1669_equation_0 = const()[name = string("op_1669_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1669_cast_fp16 = einsum(equation = var_1669_equation_0, values = (var_1605_cast_fp16_4, var_1651_cast_fp16_4))[name = string("op_1669_cast_fp16")];
            string var_1671_equation_0 = const()[name = string("op_1671_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1671_cast_fp16 = einsum(equation = var_1671_equation_0, values = (var_1605_cast_fp16_5, var_1651_cast_fp16_5))[name = string("op_1671_cast_fp16")];
            string var_1673_equation_0 = const()[name = string("op_1673_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1673_cast_fp16 = einsum(equation = var_1673_equation_0, values = (var_1605_cast_fp16_6, var_1651_cast_fp16_6))[name = string("op_1673_cast_fp16")];
            string var_1675_equation_0 = const()[name = string("op_1675_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1675_cast_fp16 = einsum(equation = var_1675_equation_0, values = (var_1605_cast_fp16_7, var_1651_cast_fp16_7))[name = string("op_1675_cast_fp16")];
            bool attn_41_interleave_0 = const()[name = string("attn_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_41_cast_fp16 = concat(axis = var_46, interleave = attn_41_interleave_0, values = (var_1661_cast_fp16, var_1663_cast_fp16, var_1665_cast_fp16, var_1667_cast_fp16, var_1669_cast_fp16, var_1671_cast_fp16, var_1673_cast_fp16, var_1675_cast_fp16))[name = string("attn_41_cast_fp16")];
            tensor<int32, [2]> var_1683 = const()[name = string("op_1683"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1685 = const()[name = string("op_1685"), val = tensor<int32, [2]>([1, 1])];
            string input0_41_pad_type_0 = const()[name = string("input0_41_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_41_pad_0 = const()[name = string("input0_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45713216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45712128))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45975424)))];
            tensor<fp16, [1, 512, 1, 256]> input0_41_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16, dilations = var_1685, groups = var_46, pad = input0_41_pad_0, pad_type = input0_41_pad_type_0, strides = var_1683, weight = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_41_cast_fp16)[name = string("input0_41_cast_fp16")];
            tensor<int32, [2]> var_1692 = const()[name = string("op_1692"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1694 = const()[name = string("op_1694"), val = tensor<int32, [2]>([1, 1])];
            string x_66_pad_type_0 = const()[name = string("x_66_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_66_pad_0 = const()[name = string("x_66_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55078592))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55078272))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55144192)))];
            tensor<fp16, [1, 128, 1, 256]> x_66_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1694, groups = var_46, pad = x_66_pad_0, pad_type = x_66_pad_type_0, strides = var_1692, weight = nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_41_cast_fp16)[name = string("x_66_cast_fp16")];
            fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1698_cast_fp16 = mul(x = x_66_cast_fp16, y = var_1697_to_fp16)[name = string("op_1698_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1699_cast_fp16 = sigmoid(x = var_1698_cast_fp16)[name = string("op_1699_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_81_cast_fp16 = mul(x = x_66_cast_fp16, y = var_1699_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<int32, [2]> var_1703 = const()[name = string("op_1703"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1705 = const()[name = string("op_1705"), val = tensor<int32, [2]>([1, 1])];
            string x_68_pad_type_0 = const()[name = string("x_68_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_68_pad_0 = const()[name = string("x_68_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55145600))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55144512))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55211200)))];
            tensor<fp16, [1, 512, 1, 256]> x_68_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1705, groups = var_46, pad = x_68_pad_0, pad_type = x_68_pad_type_0, strides = var_1703, weight = nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_81_cast_fp16)[name = string("x_68_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_43_cast_fp16 = add(x = x_68_cast_fp16, y = input0_41_cast_fp16)[name = string("attn_43_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_14_cast_fp16 = add(x = inputs4_1_cast_fp16, y = attn_43_cast_fp16)[name = string("inputs0_14_cast_fp16")];
            tensor<int32, [1]> input_83_axes_0 = const()[name = string("input_83_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_83_gamma_0_to_fp16 = const()[name = string("input_83_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55212288)))];
            tensor<fp16, [512]> input_83_beta_0_to_fp16 = const()[name = string("input_83_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55213376)))];
            fp16 var_1718_to_fp16 = const()[name = string("op_1718_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = input_83_beta_0_to_fp16, epsilon = var_1718_to_fp16, gamma = input_83_gamma_0_to_fp16, x = inputs0_14_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [2]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1734 = const()[name = string("op_1734"), val = tensor<int32, [2]>([1, 1])];
            string x_70_pad_type_0 = const()[name = string("x_70_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_70_pad_0 = const()[name = string("x_70_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45982848))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45978688))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47035648))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47031488))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_70_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1734, groups = var_46, pad = x_70_pad_0, pad_type = x_70_pad_type_0, strides = var_1732, weight = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_83_cast_fp16)[name = string("x_70_cast_fp16")];
            fp16 var_1737_to_fp16 = const()[name = string("op_1737_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1738_cast_fp16 = mul(x = x_70_cast_fp16, y = var_1737_to_fp16)[name = string("op_1738_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1739_cast_fp16 = sigmoid(x = var_1738_cast_fp16)[name = string("op_1739_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_85_cast_fp16 = mul(x = x_70_cast_fp16, y = var_1739_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [2]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1745 = const()[name = string("op_1745"), val = tensor<int32, [2]>([1, 1])];
            string input0_43_pad_type_0 = const()[name = string("input0_43_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_43_pad_0 = const()[name = string("input0_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47038848))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47037760))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48087488)))];
            tensor<fp16, [1, 512, 1, 256]> input0_43_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16, dilations = var_1745, groups = var_46, pad = input0_43_pad_0, pad_type = input0_43_pad_type_0, strides = var_1743, weight = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_85_cast_fp16)[name = string("input0_43_cast_fp16")];
            tensor<int32, [2]> var_1753 = const()[name = string("op_1753"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1755 = const()[name = string("op_1755"), val = tensor<int32, [2]>([1, 1])];
            string x_72_pad_type_0 = const()[name = string("x_72_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_72_pad_0 = const()[name = string("x_72_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55214784))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55214464))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55280384)))];
            tensor<fp16, [1, 128, 1, 256]> x_72_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1755, groups = var_46, pad = x_72_pad_0, pad_type = x_72_pad_type_0, strides = var_1753, weight = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_43_cast_fp16)[name = string("x_72_cast_fp16")];
            fp16 var_1758_to_fp16 = const()[name = string("op_1758_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1759_cast_fp16 = mul(x = x_72_cast_fp16, y = var_1758_to_fp16)[name = string("op_1759_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1760_cast_fp16 = sigmoid(x = var_1759_cast_fp16)[name = string("op_1760_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_89_cast_fp16 = mul(x = x_72_cast_fp16, y = var_1760_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<int32, [2]> var_1764 = const()[name = string("op_1764"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1766 = const()[name = string("op_1766"), val = tensor<int32, [2]>([1, 1])];
            string x_74_pad_type_0 = const()[name = string("x_74_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_74_pad_0 = const()[name = string("x_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55281792))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55280704))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55347392)))];
            tensor<fp16, [1, 512, 1, 256]> x_74_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1766, groups = var_46, pad = x_74_pad_0, pad_type = x_74_pad_type_0, strides = var_1764, weight = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_89_cast_fp16)[name = string("x_74_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_14_cast_fp16 = add(x = x_74_cast_fp16, y = input0_43_cast_fp16)[name = string("f_14_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_14_cast_fp16 = add(x = f_14_cast_fp16, y = inputs0_14_cast_fp16)[name = string("x1_14_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1772_cast_fp16 = mul(x = inputs4_1_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs5_1_cast_fp16 = add(x = var_1772_cast_fp16, y = x1_14_cast_fp16)[name = string("inputs5_1_cast_fp16")];
            tensor<int32, [1]> k_2_axes_0 = const()[name = string("k_2_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_2_gamma_0_to_fp16 = const()[name = string("k_2_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55348480)))];
            tensor<fp16, [512]> k_2_beta_0_to_fp16 = const()[name = string("k_2_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55349568)))];
            fp16 var_1790_to_fp16 = const()[name = string("op_1790_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_2_cast_fp16 = layer_norm(axes = k_2_axes_0, beta = k_2_beta_0_to_fp16, epsilon = var_1790_to_fp16, gamma = k_2_gamma_0_to_fp16, x = inputs5_1_cast_fp16)[name = string("k_2_cast_fp16")];
            tensor<int32, [2]> var_1809 = const()[name = string("op_1809"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1811 = const()[name = string("op_1811"), val = tensor<int32, [2]>([1, 1])];
            string var_1813_pad_type_0 = const()[name = string("op_1813_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1813_pad_0 = const()[name = string("op_1813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48091840))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48090752))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48354048)))];
            tensor<fp16, [1, 512, 1, 256]> var_1813_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16, dilations = var_1811, groups = var_46, pad = var_1813_pad_0, pad_type = var_1813_pad_type_0, strides = var_1809, weight = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_1813_cast_fp16")];
            tensor<int32, [2]> var_1816 = const()[name = string("op_1816"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1818 = const()[name = string("op_1818"), val = tensor<int32, [2]>([1, 1])];
            string k_1_pad_type_0 = const()[name = string("k_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_1_pad_0 = const()[name = string("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48356224))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48355136))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48618432)))];
            tensor<fp16, [1, 512, 1, 256]> k_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16, dilations = var_1818, groups = var_46, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = var_1816, weight = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [2]> var_1823 = const()[name = string("op_1823"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [2]>([1, 1])];
            string var_1827_pad_type_0 = const()[name = string("op_1827_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1827_pad_0 = const()[name = string("op_1827_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48620608))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48619520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48882816)))];
            tensor<fp16, [1, 512, 1, 256]> var_1827_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16, dilations = var_1825, groups = var_46, pad = var_1827_pad_0, pad_type = var_1827_pad_type_0, strides = var_1823, weight = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_1827_cast_fp16")];
            tensor<int32, [8]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1828_axis_0 = const()[name = string("op_1828_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1828_cast_fp16_7 = split(axis = var_1828_axis_0, split_sizes = tile_28, x = var_1813_cast_fp16)[name = string("op_1828_cast_fp16")];
            tensor<int32, [4]> var_1837_perm_0 = const()[name = string("op_1837_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1838_axis_0 = const()[name = string("op_1838_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_1 = transpose(perm = var_1837_perm_0, x = k_1_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1838_cast_fp16_7 = split(axis = var_1838_axis_0, split_sizes = tile_29, x = transpose_1)[name = string("op_1838_cast_fp16")];
            tensor<int32, [8]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1847_axis_0 = const()[name = string("op_1847_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1847_cast_fp16_7 = split(axis = var_1847_axis_0, split_sizes = tile_30, x = var_1827_cast_fp16)[name = string("op_1847_cast_fp16")];
            string var_1857_equation_0 = const()[name = string("op_1857_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1857_cast_fp16 = einsum(equation = var_1857_equation_0, values = (var_1838_cast_fp16_0, var_1828_cast_fp16_0))[name = string("op_1857_cast_fp16")];
            fp16 var_1858_to_fp16 = const()[name = string("op_1858_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1859_cast_fp16 = mul(x = var_1857_cast_fp16, y = var_1858_to_fp16)[name = string("op_1859_cast_fp16")];
            string var_1861_equation_0 = const()[name = string("op_1861_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1861_cast_fp16 = einsum(equation = var_1861_equation_0, values = (var_1838_cast_fp16_1, var_1828_cast_fp16_1))[name = string("op_1861_cast_fp16")];
            fp16 var_1862_to_fp16 = const()[name = string("op_1862_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1863_cast_fp16 = mul(x = var_1861_cast_fp16, y = var_1862_to_fp16)[name = string("op_1863_cast_fp16")];
            string var_1865_equation_0 = const()[name = string("op_1865_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1865_cast_fp16 = einsum(equation = var_1865_equation_0, values = (var_1838_cast_fp16_2, var_1828_cast_fp16_2))[name = string("op_1865_cast_fp16")];
            fp16 var_1866_to_fp16 = const()[name = string("op_1866_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1867_cast_fp16 = mul(x = var_1865_cast_fp16, y = var_1866_to_fp16)[name = string("op_1867_cast_fp16")];
            string var_1869_equation_0 = const()[name = string("op_1869_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1869_cast_fp16 = einsum(equation = var_1869_equation_0, values = (var_1838_cast_fp16_3, var_1828_cast_fp16_3))[name = string("op_1869_cast_fp16")];
            fp16 var_1870_to_fp16 = const()[name = string("op_1870_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1871_cast_fp16 = mul(x = var_1869_cast_fp16, y = var_1870_to_fp16)[name = string("op_1871_cast_fp16")];
            string var_1873_equation_0 = const()[name = string("op_1873_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1873_cast_fp16 = einsum(equation = var_1873_equation_0, values = (var_1838_cast_fp16_4, var_1828_cast_fp16_4))[name = string("op_1873_cast_fp16")];
            fp16 var_1874_to_fp16 = const()[name = string("op_1874_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1875_cast_fp16 = mul(x = var_1873_cast_fp16, y = var_1874_to_fp16)[name = string("op_1875_cast_fp16")];
            string var_1877_equation_0 = const()[name = string("op_1877_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1877_cast_fp16 = einsum(equation = var_1877_equation_0, values = (var_1838_cast_fp16_5, var_1828_cast_fp16_5))[name = string("op_1877_cast_fp16")];
            fp16 var_1878_to_fp16 = const()[name = string("op_1878_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16 = mul(x = var_1877_cast_fp16, y = var_1878_to_fp16)[name = string("op_1879_cast_fp16")];
            string var_1881_equation_0 = const()[name = string("op_1881_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1881_cast_fp16 = einsum(equation = var_1881_equation_0, values = (var_1838_cast_fp16_6, var_1828_cast_fp16_6))[name = string("op_1881_cast_fp16")];
            fp16 var_1882_to_fp16 = const()[name = string("op_1882_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1883_cast_fp16 = mul(x = var_1881_cast_fp16, y = var_1882_to_fp16)[name = string("op_1883_cast_fp16")];
            string var_1885_equation_0 = const()[name = string("op_1885_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1885_cast_fp16 = einsum(equation = var_1885_equation_0, values = (var_1838_cast_fp16_7, var_1828_cast_fp16_7))[name = string("op_1885_cast_fp16")];
            fp16 var_1886_to_fp16 = const()[name = string("op_1886_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1887_cast_fp16 = mul(x = var_1885_cast_fp16, y = var_1886_to_fp16)[name = string("op_1887_cast_fp16")];
            bool attn_weights_1_interleave_0 = const()[name = string("attn_weights_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_1_cast_fp16 = concat(axis = var_47, interleave = attn_weights_1_interleave_0, values = (var_1859_cast_fp16, var_1863_cast_fp16, var_1867_cast_fp16, var_1871_cast_fp16, var_1875_cast_fp16, var_1879_cast_fp16, var_1883_cast_fp16, var_1887_cast_fp16))[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_1_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_6_cast_fp16 = softmax(axis = var_46, x = attn_weights0_1_cast_fp16)[name = string("input_6_cast_fp16")];
            tensor<int32, [8]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1893_axis_0 = const()[name = string("op_1893_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1893_cast_fp16_7 = split(axis = var_1893_axis_0, split_sizes = tile_31, x = input_6_cast_fp16)[name = string("op_1893_cast_fp16")];
            string var_1903_equation_0 = const()[name = string("op_1903_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1903_cast_fp16 = einsum(equation = var_1903_equation_0, values = (var_1847_cast_fp16_0, var_1893_cast_fp16_0))[name = string("op_1903_cast_fp16")];
            string var_1905_equation_0 = const()[name = string("op_1905_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1905_cast_fp16 = einsum(equation = var_1905_equation_0, values = (var_1847_cast_fp16_1, var_1893_cast_fp16_1))[name = string("op_1905_cast_fp16")];
            string var_1907_equation_0 = const()[name = string("op_1907_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1907_cast_fp16 = einsum(equation = var_1907_equation_0, values = (var_1847_cast_fp16_2, var_1893_cast_fp16_2))[name = string("op_1907_cast_fp16")];
            string var_1909_equation_0 = const()[name = string("op_1909_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1909_cast_fp16 = einsum(equation = var_1909_equation_0, values = (var_1847_cast_fp16_3, var_1893_cast_fp16_3))[name = string("op_1909_cast_fp16")];
            string var_1911_equation_0 = const()[name = string("op_1911_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1911_cast_fp16 = einsum(equation = var_1911_equation_0, values = (var_1847_cast_fp16_4, var_1893_cast_fp16_4))[name = string("op_1911_cast_fp16")];
            string var_1913_equation_0 = const()[name = string("op_1913_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1913_cast_fp16 = einsum(equation = var_1913_equation_0, values = (var_1847_cast_fp16_5, var_1893_cast_fp16_5))[name = string("op_1913_cast_fp16")];
            string var_1915_equation_0 = const()[name = string("op_1915_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1915_cast_fp16 = einsum(equation = var_1915_equation_0, values = (var_1847_cast_fp16_6, var_1893_cast_fp16_6))[name = string("op_1915_cast_fp16")];
            string var_1917_equation_0 = const()[name = string("op_1917_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1917_cast_fp16 = einsum(equation = var_1917_equation_0, values = (var_1847_cast_fp16_7, var_1893_cast_fp16_7))[name = string("op_1917_cast_fp16")];
            bool attn_4_interleave_0 = const()[name = string("attn_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_4_cast_fp16 = concat(axis = var_46, interleave = attn_4_interleave_0, values = (var_1903_cast_fp16, var_1905_cast_fp16, var_1907_cast_fp16, var_1909_cast_fp16, var_1911_cast_fp16, var_1913_cast_fp16, var_1915_cast_fp16, var_1917_cast_fp16))[name = string("attn_4_cast_fp16")];
            tensor<int32, [2]> var_1925 = const()[name = string("op_1925"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1927 = const()[name = string("op_1927"), val = tensor<int32, [2]>([1, 1])];
            string input0_3_pad_type_0 = const()[name = string("input0_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_3_pad_0 = const()[name = string("input0_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48884992))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48883904))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49147200)))];
            tensor<fp16, [1, 512, 1, 256]> input0_3_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16, dilations = var_1927, groups = var_46, pad = input0_3_pad_0, pad_type = input0_3_pad_type_0, strides = var_1925, weight = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_4_cast_fp16)[name = string("input0_3_cast_fp16")];
            tensor<int32, [2]> var_1934 = const()[name = string("op_1934"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1936 = const()[name = string("op_1936"), val = tensor<int32, [2]>([1, 1])];
            string x_5_pad_type_0 = const()[name = string("x_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_5_pad_0 = const()[name = string("x_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55350976))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55350656))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55416576)))];
            tensor<fp16, [1, 128, 1, 256]> x_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1936, groups = var_46, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = var_1934, weight = nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_3_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1940_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1939_to_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1941_cast_fp16 = sigmoid(x = var_1940_cast_fp16)[name = string("op_1941_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_4_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1941_cast_fp16)[name = string("input_4_cast_fp16")];
            tensor<int32, [2]> var_1945 = const()[name = string("op_1945"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1947 = const()[name = string("op_1947"), val = tensor<int32, [2]>([1, 1])];
            string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55417984))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55416896))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55483584)))];
            tensor<fp16, [1, 512, 1, 256]> x_3_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1947, groups = var_46, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = var_1945, weight = nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_4_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_1_cast_fp16 = add(x = x_3_cast_fp16, y = input0_3_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_1_cast_fp16 = add(x = inputs5_1_cast_fp16, y = attn_1_cast_fp16)[name = string("inputs0_1_cast_fp16")];
            tensor<int32, [1]> input_8_axes_0 = const()[name = string("input_8_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_8_gamma_0_to_fp16 = const()[name = string("input_8_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55484672)))];
            tensor<fp16, [512]> input_8_beta_0_to_fp16 = const()[name = string("input_8_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55485760)))];
            fp16 var_1960_to_fp16 = const()[name = string("op_1960_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_8_cast_fp16 = layer_norm(axes = input_8_axes_0, beta = input_8_beta_0_to_fp16, epsilon = var_1960_to_fp16, gamma = input_8_gamma_0_to_fp16, x = inputs0_1_cast_fp16)[name = string("input_8_cast_fp16")];
            tensor<int32, [2]> var_1974 = const()[name = string("op_1974"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [2]>([1, 1])];
            string x_7_pad_type_0 = const()[name = string("x_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_7_pad_0 = const()[name = string("x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49154624))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49150464))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50207424))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50203264))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_7_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1976, groups = var_46, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = var_1974, weight = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_8_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1979_to_fp16 = const()[name = string("op_1979_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1980_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1979_to_fp16)[name = string("op_1980_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1981_cast_fp16 = sigmoid(x = var_1980_cast_fp16)[name = string("op_1981_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_2_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1981_cast_fp16)[name = string("input_2_cast_fp16")];
            tensor<int32, [2]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1987 = const()[name = string("op_1987"), val = tensor<int32, [2]>([1, 1])];
            string input0_1_pad_type_0 = const()[name = string("input0_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_1_pad_0 = const()[name = string("input0_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50210624))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50209536))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51259264)))];
            tensor<fp16, [1, 512, 1, 256]> input0_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16, dilations = var_1987, groups = var_46, pad = input0_1_pad_0, pad_type = input0_1_pad_type_0, strides = var_1985, weight = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_2_cast_fp16)[name = string("input0_1_cast_fp16")];
            tensor<int32, [2]> var_1995 = const()[name = string("op_1995"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1997 = const()[name = string("op_1997"), val = tensor<int32, [2]>([1, 1])];
            string x_2_pad_type_0 = const()[name = string("x_2_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_2_pad_0 = const()[name = string("x_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55487168))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55486848))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55552768)))];
            tensor<fp16, [1, 128, 1, 256]> x_2_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1997, groups = var_46, pad = x_2_pad_0, pad_type = x_2_pad_type_0, strides = var_1995, weight = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input0_1_cast_fp16)[name = string("x_2_cast_fp16")];
            fp16 var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_2001_cast_fp16 = mul(x = x_2_cast_fp16, y = var_2000_to_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_2002_cast_fp16 = sigmoid(x = var_2001_cast_fp16)[name = string("op_2002_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_1_cast_fp16 = mul(x = x_2_cast_fp16, y = var_2002_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [2]> var_2006 = const()[name = string("op_2006"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2008 = const()[name = string("op_2008"), val = tensor<int32, [2]>([1, 1])];
            string x_1_pad_type_0 = const()[name = string("x_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_1_pad_0 = const()[name = string("x_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55554176))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55553088))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55619776)))];
            tensor<fp16, [1, 512, 1, 256]> x_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_2008, groups = var_46, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = var_2006, weight = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_1_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_1_cast_fp16 = add(x = x_1_cast_fp16, y = input0_1_cast_fp16)[name = string("f_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_1_cast_fp16 = add(x = f_1_cast_fp16, y = inputs0_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            fp16 var_2013_to_fp16 = const()[name = string("op_2013_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_2014_cast_fp16 = mul(x = inputs5_1_cast_fp16, y = var_2013_to_fp16)[name = string("op_2014_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs6_1_cast_fp16 = add(x = var_2014_cast_fp16, y = x1_1_cast_fp16)[name = string("inputs6_1_cast_fp16")];
            tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> embeddings_1_gamma_0_to_fp16 = const()[name = string("embeddings_1_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55620864)))];
            tensor<fp16, [512]> embeddings_1_beta_0_to_fp16 = const()[name = string("embeddings_1_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(55621952)))];
            fp16 var_2026_to_fp16 = const()[name = string("op_2026_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = embeddings_1_beta_0_to_fp16, epsilon = var_2026_to_fp16, gamma = embeddings_1_gamma_0_to_fp16, x = inputs6_1_cast_fp16)[name = string("embeddings_1_cast_fp16")];
            tensor<int32, [4]> mlm_embeddings_1_perm_0 = const()[name = string("mlm_embeddings_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string mlm_embeddings_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("mlm_embeddings_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp16, [1, 256, 1, 512]> transpose_0 = transpose(perm = mlm_embeddings_1_perm_0, x = embeddings_1_cast_fp16)[name = string("transpose_0")];
            tensor<fp32, [1, 256, 1, 512]> mlm_embeddings = cast(dtype = mlm_embeddings_1_cast_fp16_to_fp32_dtype_0, x = transpose_0)[name = string("cast_0")];
        } -> (mlm_embeddings);
    func main<ios16>(tensor<int32, [1, 256, 1, 3]> mlm_input) {
            int32 var_6 = const()[name = string("op_6"), val = int32(0)];
            tensor<int32, [3]> var_13 = const()[name = string("op_13"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_14_axis_0 = const()[name = string("op_14_axis_0"), val = int32(-1)];
            tensor<int32, [1, 256, 1, 1]> var_14_0, tensor<int32, [1, 256, 1, 1]> var_14_1, tensor<int32, [1, 256, 1, 1]> var_14_2 = split(axis = var_14_axis_0, split_sizes = var_13, x = mlm_input)[name = string("op_14")];
            tensor<int32, [1]> var_18_axes_0 = const()[name = string("op_18_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_18 = squeeze(axes = var_18_axes_0, x = var_14_0)[name = string("op_18")];
            tensor<int32, [1]> tok_ids_1_axes_0 = const()[name = string("tok_ids_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256]> tok_ids_1 = squeeze(axes = tok_ids_1_axes_0, x = var_18)[name = string("tok_ids_1")];
            tensor<int32, [1]> var_20_axes_0 = const()[name = string("op_20_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_20 = squeeze(axes = var_20_axes_0, x = var_14_1)[name = string("op_20")];
            tensor<int32, [1]> var_22_axes_0 = const()[name = string("op_22_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_22 = squeeze(axes = var_22_axes_0, x = var_14_2)[name = string("op_22")];
            tensor<bool, [1, 256]> var_24 = not_equal(x = tok_ids_1, y = var_6)[name = string("op_24")];
            fp16 var_8_to_fp16 = const()[name = string("op_8_to_fp16"), val = fp16(1)];
            string var_24_to_fp32_to_fp16_dtype_0 = const()[name = string("op_24_to_fp32_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256]> cast_1 = cast(dtype = var_24_to_fp32_to_fp16_dtype_0, x = var_24)[name = string("cast_1")];
            tensor<fp16, [1, 256]> var_29_cast_fp16 = sub(x = var_8_to_fp16, y = cast_1)[name = string("op_29_cast_fp16")];
            fp16 var_30_to_fp16 = const()[name = string("op_30_to_fp16"), val = fp16(-10000)];
            tensor<fp16, [1, 256]> padding_mask0_1_cast_fp16 = mul(x = var_29_cast_fp16, y = var_30_to_fp16)[name = string("padding_mask0_1_cast_fp16")];
            tensor<int32, [4]> var_32 = const()[name = string("op_32"), val = tensor<int32, [4]>([-1, 256, 1, 1])];
            tensor<fp16, [1, 256, 1, 1]> var_33_cast_fp16 = reshape(shape = var_32, x = padding_mask0_1_cast_fp16)[name = string("op_33_cast_fp16")];
            int32 var_46 = const()[name = string("op_46"), val = int32(1)];
            int32 var_47 = const()[name = string("op_47"), val = int32(2)];
            tensor<int32, [1]> input_6_axes_0 = const()[name = string("input_6_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 256, 1]> input_6 = expand_dims(axes = input_6_axes_0, x = tok_ids_1)[name = string("input_6")];
            int32 var_54_axis_0 = const()[name = string("op_54_axis_0"), val = int32(0)];
            tensor<fp16, [50000, 512]> nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [50000, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(150272))), scale = tensor<fp16, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50176))), zero_point = tensor<int8, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(64)))];
            int32 op_54_cast_fp16_batch_dims_0 = const()[name = string("op_54_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_54_cast_fp16 = gather(axis = var_54_axis_0, batch_dims = op_54_cast_fp16_batch_dims_0, indices = input_6, x = nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized)[name = string("op_54_cast_fp16")];
            int32 var_57_axis_0 = const()[name = string("op_57_axis_0"), val = int32(0)];
            tensor<fp16, [256, 512]> nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [256, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25751232))), scale = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750656))), zero_point = tensor<int8, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750336)))];
            int32 op_57_cast_fp16_batch_dims_0 = const()[name = string("op_57_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_57_cast_fp16 = gather(axis = var_57_axis_0, batch_dims = op_57_cast_fp16_batch_dims_0, indices = var_20, x = nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized)[name = string("op_57_cast_fp16")];
            int32 var_60_axis_0 = const()[name = string("op_60_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512]> nlp_net_default_encoder_seg_embed_weight_to_fp16 = const()[name = string("nlp_net_default_encoder_seg_embed_weight_to_fp16"), val = tensor<fp16, [1, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25882368)))];
            int32 op_60_cast_fp16_batch_dims_0 = const()[name = string("op_60_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_60_cast_fp16 = gather(axis = var_60_axis_0, batch_dims = op_60_cast_fp16_batch_dims_0, indices = var_22, x = nlp_net_default_encoder_seg_embed_weight_to_fp16)[name = string("op_60_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_62_cast_fp16 = add(x = var_54_cast_fp16, y = var_57_cast_fp16)[name = string("op_62_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_63_cast_fp16 = add(x = var_62_cast_fp16, y = var_60_cast_fp16)[name = string("op_63_cast_fp16")];
            tensor<int32, [4]> t_1_perm_0 = const()[name = string("t_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_3_gamma_0_to_fp16 = const()[name = string("k_3_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25883456)))];
            tensor<fp16, [512]> k_3_beta_0_to_fp16 = const()[name = string("k_3_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25884544)))];
            fp16 var_94_to_fp16 = const()[name = string("op_94_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> transpose_9 = transpose(perm = t_1_perm_0, x = var_63_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 512, 1, 256]> k_3_cast_fp16 = layer_norm(axes = k_3_axes_0, beta = k_3_beta_0_to_fp16, epsilon = var_94_to_fp16, gamma = k_3_gamma_0_to_fp16, x = transpose_9)[name = string("k_3_cast_fp16")];
            tensor<int32, [2]> var_113 = const()[name = string("op_113"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_115 = const()[name = string("op_115"), val = tensor<int32, [2]>([1, 1])];
            string var_117_pad_type_0 = const()[name = string("op_117_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_117_pad_0 = const()[name = string("op_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25887296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25886208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26149504)))];
            tensor<fp16, [1, 512, 1, 256]> var_117_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16, dilations = var_115, groups = var_46, pad = var_117_pad_0, pad_type = var_117_pad_type_0, strides = var_113, weight = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_117_cast_fp16")];
            tensor<int32, [2]> var_120 = const()[name = string("op_120"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_122 = const()[name = string("op_122"), val = tensor<int32, [2]>([1, 1])];
            string k_5_pad_type_0 = const()[name = string("k_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_5_pad_0 = const()[name = string("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26151680))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26150592))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26413888)))];
            tensor<fp16, [1, 512, 1, 256]> k_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16, dilations = var_122, groups = var_46, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = var_120, weight = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [2]> var_127 = const()[name = string("op_127"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_129 = const()[name = string("op_129"), val = tensor<int32, [2]>([1, 1])];
            string var_131_pad_type_0 = const()[name = string("op_131_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_131_pad_0 = const()[name = string("op_131_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26416064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26414976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26678272)))];
            tensor<fp16, [1, 512, 1, 256]> var_131_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16, dilations = var_129, groups = var_46, pad = var_131_pad_0, pad_type = var_131_pad_type_0, strides = var_127, weight = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_131_cast_fp16")];
            tensor<int32, [8]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_132_axis_0 = const()[name = string("op_132_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_132_cast_fp16_7 = split(axis = var_132_axis_0, split_sizes = tile_0, x = var_117_cast_fp16)[name = string("op_132_cast_fp16")];
            tensor<int32, [4]> var_141_perm_0 = const()[name = string("op_141_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_142_axis_0 = const()[name = string("op_142_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_8 = transpose(perm = var_141_perm_0, x = k_5_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_142_cast_fp16_7 = split(axis = var_142_axis_0, split_sizes = tile_1, x = transpose_8)[name = string("op_142_cast_fp16")];
            tensor<int32, [8]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_151_axis_0 = const()[name = string("op_151_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_151_cast_fp16_7 = split(axis = var_151_axis_0, split_sizes = tile_2, x = var_131_cast_fp16)[name = string("op_151_cast_fp16")];
            string var_161_equation_0 = const()[name = string("op_161_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_161_cast_fp16 = einsum(equation = var_161_equation_0, values = (var_142_cast_fp16_0, var_132_cast_fp16_0))[name = string("op_161_cast_fp16")];
            fp16 var_162_to_fp16 = const()[name = string("op_162_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_163_cast_fp16 = mul(x = var_161_cast_fp16, y = var_162_to_fp16)[name = string("op_163_cast_fp16")];
            string var_165_equation_0 = const()[name = string("op_165_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_165_cast_fp16 = einsum(equation = var_165_equation_0, values = (var_142_cast_fp16_1, var_132_cast_fp16_1))[name = string("op_165_cast_fp16")];
            fp16 var_166_to_fp16 = const()[name = string("op_166_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_167_cast_fp16 = mul(x = var_165_cast_fp16, y = var_166_to_fp16)[name = string("op_167_cast_fp16")];
            string var_169_equation_0 = const()[name = string("op_169_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_169_cast_fp16 = einsum(equation = var_169_equation_0, values = (var_142_cast_fp16_2, var_132_cast_fp16_2))[name = string("op_169_cast_fp16")];
            fp16 var_170_to_fp16 = const()[name = string("op_170_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_171_cast_fp16 = mul(x = var_169_cast_fp16, y = var_170_to_fp16)[name = string("op_171_cast_fp16")];
            string var_173_equation_0 = const()[name = string("op_173_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_173_cast_fp16 = einsum(equation = var_173_equation_0, values = (var_142_cast_fp16_3, var_132_cast_fp16_3))[name = string("op_173_cast_fp16")];
            fp16 var_174_to_fp16 = const()[name = string("op_174_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_175_cast_fp16 = mul(x = var_173_cast_fp16, y = var_174_to_fp16)[name = string("op_175_cast_fp16")];
            string var_177_equation_0 = const()[name = string("op_177_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_177_cast_fp16 = einsum(equation = var_177_equation_0, values = (var_142_cast_fp16_4, var_132_cast_fp16_4))[name = string("op_177_cast_fp16")];
            fp16 var_178_to_fp16 = const()[name = string("op_178_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_179_cast_fp16 = mul(x = var_177_cast_fp16, y = var_178_to_fp16)[name = string("op_179_cast_fp16")];
            string var_181_equation_0 = const()[name = string("op_181_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_181_cast_fp16 = einsum(equation = var_181_equation_0, values = (var_142_cast_fp16_5, var_132_cast_fp16_5))[name = string("op_181_cast_fp16")];
            fp16 var_182_to_fp16 = const()[name = string("op_182_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_183_cast_fp16 = mul(x = var_181_cast_fp16, y = var_182_to_fp16)[name = string("op_183_cast_fp16")];
            string var_185_equation_0 = const()[name = string("op_185_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_185_cast_fp16 = einsum(equation = var_185_equation_0, values = (var_142_cast_fp16_6, var_132_cast_fp16_6))[name = string("op_185_cast_fp16")];
            fp16 var_186_to_fp16 = const()[name = string("op_186_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_187_cast_fp16 = mul(x = var_185_cast_fp16, y = var_186_to_fp16)[name = string("op_187_cast_fp16")];
            string var_189_equation_0 = const()[name = string("op_189_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_189_cast_fp16 = einsum(equation = var_189_equation_0, values = (var_142_cast_fp16_7, var_132_cast_fp16_7))[name = string("op_189_cast_fp16")];
            fp16 var_190_to_fp16 = const()[name = string("op_190_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_191_cast_fp16 = mul(x = var_189_cast_fp16, y = var_190_to_fp16)[name = string("op_191_cast_fp16")];
            bool attn_weights_2_interleave_0 = const()[name = string("attn_weights_2_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_2_cast_fp16 = concat(axis = var_47, interleave = attn_weights_2_interleave_0, values = (var_163_cast_fp16, var_167_cast_fp16, var_171_cast_fp16, var_175_cast_fp16, var_179_cast_fp16, var_183_cast_fp16, var_187_cast_fp16, var_191_cast_fp16))[name = string("attn_weights_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_2_cast_fp16 = add(x = attn_weights_2_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_3_cast_fp16 = softmax(axis = var_46, x = attn_weights0_2_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<int32, [8]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_197_axis_0 = const()[name = string("op_197_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_197_cast_fp16_7 = split(axis = var_197_axis_0, split_sizes = tile_3, x = input_3_cast_fp16)[name = string("op_197_cast_fp16")];
            string var_207_equation_0 = const()[name = string("op_207_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_207_cast_fp16 = einsum(equation = var_207_equation_0, values = (var_151_cast_fp16_0, var_197_cast_fp16_0))[name = string("op_207_cast_fp16")];
            string var_209_equation_0 = const()[name = string("op_209_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_209_cast_fp16 = einsum(equation = var_209_equation_0, values = (var_151_cast_fp16_1, var_197_cast_fp16_1))[name = string("op_209_cast_fp16")];
            string var_211_equation_0 = const()[name = string("op_211_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_211_cast_fp16 = einsum(equation = var_211_equation_0, values = (var_151_cast_fp16_2, var_197_cast_fp16_2))[name = string("op_211_cast_fp16")];
            string var_213_equation_0 = const()[name = string("op_213_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_213_cast_fp16 = einsum(equation = var_213_equation_0, values = (var_151_cast_fp16_3, var_197_cast_fp16_3))[name = string("op_213_cast_fp16")];
            string var_215_equation_0 = const()[name = string("op_215_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_215_cast_fp16 = einsum(equation = var_215_equation_0, values = (var_151_cast_fp16_4, var_197_cast_fp16_4))[name = string("op_215_cast_fp16")];
            string var_217_equation_0 = const()[name = string("op_217_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_217_cast_fp16 = einsum(equation = var_217_equation_0, values = (var_151_cast_fp16_5, var_197_cast_fp16_5))[name = string("op_217_cast_fp16")];
            string var_219_equation_0 = const()[name = string("op_219_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_219_cast_fp16 = einsum(equation = var_219_equation_0, values = (var_151_cast_fp16_6, var_197_cast_fp16_6))[name = string("op_219_cast_fp16")];
            string var_221_equation_0 = const()[name = string("op_221_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_221_cast_fp16 = einsum(equation = var_221_equation_0, values = (var_151_cast_fp16_7, var_197_cast_fp16_7))[name = string("op_221_cast_fp16")];
            bool attn_6_interleave_0 = const()[name = string("attn_6_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_6_cast_fp16 = concat(axis = var_46, interleave = attn_6_interleave_0, values = (var_207_cast_fp16, var_209_cast_fp16, var_211_cast_fp16, var_213_cast_fp16, var_215_cast_fp16, var_217_cast_fp16, var_219_cast_fp16, var_221_cast_fp16))[name = string("attn_6_cast_fp16")];
            tensor<int32, [2]> var_229 = const()[name = string("op_229"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_231 = const()[name = string("op_231"), val = tensor<int32, [2]>([1, 1])];
            string attn_8_pad_type_0 = const()[name = string("attn_8_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_8_pad_0 = const()[name = string("attn_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26680448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26679360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26942656)))];
            tensor<fp16, [1, 512, 1, 256]> attn_8_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16, dilations = var_231, groups = var_46, pad = attn_8_pad_0, pad_type = attn_8_pad_type_0, strides = var_229, weight = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_6_cast_fp16)[name = string("attn_8_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_1_cast_fp16 = add(x = transpose_9, y = attn_8_cast_fp16)[name = string("inputs_1_cast_fp16")];
            tensor<int32, [1]> input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_5_gamma_0_to_fp16 = const()[name = string("input_5_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26943744)))];
            tensor<fp16, [512]> input_5_beta_0_to_fp16 = const()[name = string("input_5_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26944832)))];
            fp16 var_243_to_fp16 = const()[name = string("op_243_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = input_5_beta_0_to_fp16, epsilon = var_243_to_fp16, gamma = input_5_gamma_0_to_fp16, x = inputs_1_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<int32, [2]> var_257 = const()[name = string("op_257"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_259 = const()[name = string("op_259"), val = tensor<int32, [2]>([1, 1])];
            string x_2_pad_type_0 = const()[name = string("x_2_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_2_pad_0 = const()[name = string("x_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26952192))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26948032))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28004992))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28000832))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_2_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_259, groups = var_46, pad = x_2_pad_0, pad_type = x_2_pad_type_0, strides = var_257, weight = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_5_cast_fp16)[name = string("x_2_cast_fp16")];
            fp16 var_262_to_fp16 = const()[name = string("op_262_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_263_cast_fp16 = mul(x = x_2_cast_fp16, y = var_262_to_fp16)[name = string("op_263_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_264_cast_fp16 = sigmoid(x = var_263_cast_fp16)[name = string("op_264_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_7_cast_fp16 = mul(x = x_2_cast_fp16, y = var_264_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<int32, [2]> var_268 = const()[name = string("op_268"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_270 = const()[name = string("op_270"), val = tensor<int32, [2]>([1, 1])];
            string input0_3_pad_type_0 = const()[name = string("input0_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_3_pad_0 = const()[name = string("input0_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28008192))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28007104))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29056832)))];
            tensor<fp16, [1, 512, 1, 256]> input0_3_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16, dilations = var_270, groups = var_46, pad = input0_3_pad_0, pad_type = input0_3_pad_type_0, strides = var_268, weight = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_7_cast_fp16)[name = string("input0_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_274_cast_fp16 = add(x = input0_3_cast_fp16, y = inputs_1_cast_fp16)[name = string("op_274_cast_fp16")];
            fp16 var_275_to_fp16 = const()[name = string("op_275_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_276_cast_fp16 = mul(x = transpose_9, y = var_275_to_fp16)[name = string("op_276_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_2_cast_fp16 = add(x = var_276_cast_fp16, y = var_274_cast_fp16)[name = string("inputs_2_cast_fp16")];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_7_gamma_0_to_fp16 = const()[name = string("k_7_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29057920)))];
            tensor<fp16, [512]> k_7_beta_0_to_fp16 = const()[name = string("k_7_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29059008)))];
            fp16 var_292_to_fp16 = const()[name = string("op_292_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_7_cast_fp16 = layer_norm(axes = k_7_axes_0, beta = k_7_beta_0_to_fp16, epsilon = var_292_to_fp16, gamma = k_7_gamma_0_to_fp16, x = inputs_2_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<int32, [2]> var_311 = const()[name = string("op_311"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_313 = const()[name = string("op_313"), val = tensor<int32, [2]>([1, 1])];
            string var_315_pad_type_0 = const()[name = string("op_315_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_315_pad_0 = const()[name = string("op_315_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29061184))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29060096))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29323392)))];
            tensor<fp16, [1, 512, 1, 256]> var_315_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16, dilations = var_313, groups = var_46, pad = var_315_pad_0, pad_type = var_315_pad_type_0, strides = var_311, weight = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_315_cast_fp16")];
            tensor<int32, [2]> var_318 = const()[name = string("op_318"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_320 = const()[name = string("op_320"), val = tensor<int32, [2]>([1, 1])];
            string k_9_pad_type_0 = const()[name = string("k_9_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_9_pad_0 = const()[name = string("k_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29325568))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29324480))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29587776)))];
            tensor<fp16, [1, 512, 1, 256]> k_9_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16, dilations = var_320, groups = var_46, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = var_318, weight = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [2]> var_325 = const()[name = string("op_325"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_327 = const()[name = string("op_327"), val = tensor<int32, [2]>([1, 1])];
            string var_329_pad_type_0 = const()[name = string("op_329_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_329_pad_0 = const()[name = string("op_329_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29589952))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29588864))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29852160)))];
            tensor<fp16, [1, 512, 1, 256]> var_329_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16, dilations = var_327, groups = var_46, pad = var_329_pad_0, pad_type = var_329_pad_type_0, strides = var_325, weight = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_329_cast_fp16")];
            tensor<int32, [8]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_330_axis_0 = const()[name = string("op_330_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_330_cast_fp16_7 = split(axis = var_330_axis_0, split_sizes = tile_4, x = var_315_cast_fp16)[name = string("op_330_cast_fp16")];
            tensor<int32, [4]> var_339_perm_0 = const()[name = string("op_339_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_340_axis_0 = const()[name = string("op_340_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_7 = transpose(perm = var_339_perm_0, x = k_9_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_340_cast_fp16_7 = split(axis = var_340_axis_0, split_sizes = tile_5, x = transpose_7)[name = string("op_340_cast_fp16")];
            tensor<int32, [8]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_349_axis_0 = const()[name = string("op_349_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_349_cast_fp16_7 = split(axis = var_349_axis_0, split_sizes = tile_6, x = var_329_cast_fp16)[name = string("op_349_cast_fp16")];
            string var_359_equation_0 = const()[name = string("op_359_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_359_cast_fp16 = einsum(equation = var_359_equation_0, values = (var_340_cast_fp16_0, var_330_cast_fp16_0))[name = string("op_359_cast_fp16")];
            fp16 var_360_to_fp16 = const()[name = string("op_360_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_361_cast_fp16 = mul(x = var_359_cast_fp16, y = var_360_to_fp16)[name = string("op_361_cast_fp16")];
            string var_363_equation_0 = const()[name = string("op_363_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_363_cast_fp16 = einsum(equation = var_363_equation_0, values = (var_340_cast_fp16_1, var_330_cast_fp16_1))[name = string("op_363_cast_fp16")];
            fp16 var_364_to_fp16 = const()[name = string("op_364_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_365_cast_fp16 = mul(x = var_363_cast_fp16, y = var_364_to_fp16)[name = string("op_365_cast_fp16")];
            string var_367_equation_0 = const()[name = string("op_367_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_367_cast_fp16 = einsum(equation = var_367_equation_0, values = (var_340_cast_fp16_2, var_330_cast_fp16_2))[name = string("op_367_cast_fp16")];
            fp16 var_368_to_fp16 = const()[name = string("op_368_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_369_cast_fp16 = mul(x = var_367_cast_fp16, y = var_368_to_fp16)[name = string("op_369_cast_fp16")];
            string var_371_equation_0 = const()[name = string("op_371_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_371_cast_fp16 = einsum(equation = var_371_equation_0, values = (var_340_cast_fp16_3, var_330_cast_fp16_3))[name = string("op_371_cast_fp16")];
            fp16 var_372_to_fp16 = const()[name = string("op_372_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_373_cast_fp16 = mul(x = var_371_cast_fp16, y = var_372_to_fp16)[name = string("op_373_cast_fp16")];
            string var_375_equation_0 = const()[name = string("op_375_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_375_cast_fp16 = einsum(equation = var_375_equation_0, values = (var_340_cast_fp16_4, var_330_cast_fp16_4))[name = string("op_375_cast_fp16")];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_377_cast_fp16 = mul(x = var_375_cast_fp16, y = var_376_to_fp16)[name = string("op_377_cast_fp16")];
            string var_379_equation_0 = const()[name = string("op_379_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_379_cast_fp16 = einsum(equation = var_379_equation_0, values = (var_340_cast_fp16_5, var_330_cast_fp16_5))[name = string("op_379_cast_fp16")];
            fp16 var_380_to_fp16 = const()[name = string("op_380_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_381_cast_fp16 = mul(x = var_379_cast_fp16, y = var_380_to_fp16)[name = string("op_381_cast_fp16")];
            string var_383_equation_0 = const()[name = string("op_383_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_383_cast_fp16 = einsum(equation = var_383_equation_0, values = (var_340_cast_fp16_6, var_330_cast_fp16_6))[name = string("op_383_cast_fp16")];
            fp16 var_384_to_fp16 = const()[name = string("op_384_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_385_cast_fp16 = mul(x = var_383_cast_fp16, y = var_384_to_fp16)[name = string("op_385_cast_fp16")];
            string var_387_equation_0 = const()[name = string("op_387_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_387_cast_fp16 = einsum(equation = var_387_equation_0, values = (var_340_cast_fp16_7, var_330_cast_fp16_7))[name = string("op_387_cast_fp16")];
            fp16 var_388_to_fp16 = const()[name = string("op_388_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_389_cast_fp16 = mul(x = var_387_cast_fp16, y = var_388_to_fp16)[name = string("op_389_cast_fp16")];
            bool attn_weights_4_interleave_0 = const()[name = string("attn_weights_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_4_cast_fp16 = concat(axis = var_47, interleave = attn_weights_4_interleave_0, values = (var_361_cast_fp16, var_365_cast_fp16, var_369_cast_fp16, var_373_cast_fp16, var_377_cast_fp16, var_381_cast_fp16, var_385_cast_fp16, var_389_cast_fp16))[name = string("attn_weights_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_4_cast_fp16 = add(x = attn_weights_4_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_9_cast_fp16 = softmax(axis = var_46, x = attn_weights0_4_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<int32, [8]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_395_axis_0 = const()[name = string("op_395_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_395_cast_fp16_7 = split(axis = var_395_axis_0, split_sizes = tile_7, x = input_9_cast_fp16)[name = string("op_395_cast_fp16")];
            string var_405_equation_0 = const()[name = string("op_405_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_405_cast_fp16 = einsum(equation = var_405_equation_0, values = (var_349_cast_fp16_0, var_395_cast_fp16_0))[name = string("op_405_cast_fp16")];
            string var_407_equation_0 = const()[name = string("op_407_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_407_cast_fp16 = einsum(equation = var_407_equation_0, values = (var_349_cast_fp16_1, var_395_cast_fp16_1))[name = string("op_407_cast_fp16")];
            string var_409_equation_0 = const()[name = string("op_409_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_409_cast_fp16 = einsum(equation = var_409_equation_0, values = (var_349_cast_fp16_2, var_395_cast_fp16_2))[name = string("op_409_cast_fp16")];
            string var_411_equation_0 = const()[name = string("op_411_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_411_cast_fp16 = einsum(equation = var_411_equation_0, values = (var_349_cast_fp16_3, var_395_cast_fp16_3))[name = string("op_411_cast_fp16")];
            string var_413_equation_0 = const()[name = string("op_413_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_413_cast_fp16 = einsum(equation = var_413_equation_0, values = (var_349_cast_fp16_4, var_395_cast_fp16_4))[name = string("op_413_cast_fp16")];
            string var_415_equation_0 = const()[name = string("op_415_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_415_cast_fp16 = einsum(equation = var_415_equation_0, values = (var_349_cast_fp16_5, var_395_cast_fp16_5))[name = string("op_415_cast_fp16")];
            string var_417_equation_0 = const()[name = string("op_417_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_417_cast_fp16 = einsum(equation = var_417_equation_0, values = (var_349_cast_fp16_6, var_395_cast_fp16_6))[name = string("op_417_cast_fp16")];
            string var_419_equation_0 = const()[name = string("op_419_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_419_cast_fp16 = einsum(equation = var_419_equation_0, values = (var_349_cast_fp16_7, var_395_cast_fp16_7))[name = string("op_419_cast_fp16")];
            bool attn_12_interleave_0 = const()[name = string("attn_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_12_cast_fp16 = concat(axis = var_46, interleave = attn_12_interleave_0, values = (var_405_cast_fp16, var_407_cast_fp16, var_409_cast_fp16, var_411_cast_fp16, var_413_cast_fp16, var_415_cast_fp16, var_417_cast_fp16, var_419_cast_fp16))[name = string("attn_12_cast_fp16")];
            tensor<int32, [2]> var_427 = const()[name = string("op_427"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_429 = const()[name = string("op_429"), val = tensor<int32, [2]>([1, 1])];
            string attn_14_pad_type_0 = const()[name = string("attn_14_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_14_pad_0 = const()[name = string("attn_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29854336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29853248))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30116544)))];
            tensor<fp16, [1, 512, 1, 256]> attn_14_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16, dilations = var_429, groups = var_46, pad = attn_14_pad_0, pad_type = attn_14_pad_type_0, strides = var_427, weight = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_12_cast_fp16)[name = string("attn_14_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_4_cast_fp16 = add(x = inputs_2_cast_fp16, y = attn_14_cast_fp16)[name = string("inputs0_4_cast_fp16")];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_11_gamma_0_to_fp16 = const()[name = string("input_11_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30117632)))];
            tensor<fp16, [512]> input_11_beta_0_to_fp16 = const()[name = string("input_11_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30118720)))];
            fp16 var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = input_11_beta_0_to_fp16, epsilon = var_441_to_fp16, gamma = input_11_gamma_0_to_fp16, x = inputs0_4_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<int32, [2]> var_455 = const()[name = string("op_455"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_457 = const()[name = string("op_457"), val = tensor<int32, [2]>([1, 1])];
            string x_4_pad_type_0 = const()[name = string("x_4_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_4_pad_0 = const()[name = string("x_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30123968))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30119808))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31176768))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31172608))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_4_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_457, groups = var_46, pad = x_4_pad_0, pad_type = x_4_pad_type_0, strides = var_455, weight = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_11_cast_fp16)[name = string("x_4_cast_fp16")];
            fp16 var_460_to_fp16 = const()[name = string("op_460_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_461_cast_fp16 = mul(x = x_4_cast_fp16, y = var_460_to_fp16)[name = string("op_461_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_462_cast_fp16 = sigmoid(x = var_461_cast_fp16)[name = string("op_462_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_13_cast_fp16 = mul(x = x_4_cast_fp16, y = var_462_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<int32, [2]> var_466 = const()[name = string("op_466"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_468 = const()[name = string("op_468"), val = tensor<int32, [2]>([1, 1])];
            string input0_7_pad_type_0 = const()[name = string("input0_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_7_pad_0 = const()[name = string("input0_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31179968))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31178880))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32228608)))];
            tensor<fp16, [1, 512, 1, 256]> input0_7_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16, dilations = var_468, groups = var_46, pad = input0_7_pad_0, pad_type = input0_7_pad_type_0, strides = var_466, weight = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_13_cast_fp16)[name = string("input0_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_472_cast_fp16 = add(x = input0_7_cast_fp16, y = inputs0_4_cast_fp16)[name = string("op_472_cast_fp16")];
            fp16 var_473_to_fp16 = const()[name = string("op_473_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_474_cast_fp16 = mul(x = inputs_2_cast_fp16, y = var_473_to_fp16)[name = string("op_474_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_2_cast_fp16 = add(x = var_474_cast_fp16, y = var_472_cast_fp16)[name = string("inputs0_2_cast_fp16")];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_11_gamma_0_to_fp16 = const()[name = string("k_11_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32229696)))];
            tensor<fp16, [512]> k_11_beta_0_to_fp16 = const()[name = string("k_11_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32230784)))];
            fp16 var_490_to_fp16 = const()[name = string("op_490_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_11_cast_fp16 = layer_norm(axes = k_11_axes_0, beta = k_11_beta_0_to_fp16, epsilon = var_490_to_fp16, gamma = k_11_gamma_0_to_fp16, x = inputs0_2_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [2]> var_509 = const()[name = string("op_509"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_511 = const()[name = string("op_511"), val = tensor<int32, [2]>([1, 1])];
            string var_513_pad_type_0 = const()[name = string("op_513_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_513_pad_0 = const()[name = string("op_513_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32232960))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32231872))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32495168)))];
            tensor<fp16, [1, 512, 1, 256]> var_513_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16, dilations = var_511, groups = var_46, pad = var_513_pad_0, pad_type = var_513_pad_type_0, strides = var_509, weight = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_513_cast_fp16")];
            tensor<int32, [2]> var_516 = const()[name = string("op_516"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_518 = const()[name = string("op_518"), val = tensor<int32, [2]>([1, 1])];
            string k_13_pad_type_0 = const()[name = string("k_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_13_pad_0 = const()[name = string("k_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32497344))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32496256))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32759552)))];
            tensor<fp16, [1, 512, 1, 256]> k_13_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16, dilations = var_518, groups = var_46, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = var_516, weight = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [2]> var_523 = const()[name = string("op_523"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_525 = const()[name = string("op_525"), val = tensor<int32, [2]>([1, 1])];
            string var_527_pad_type_0 = const()[name = string("op_527_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_527_pad_0 = const()[name = string("op_527_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32761728))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32760640))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33023936)))];
            tensor<fp16, [1, 512, 1, 256]> var_527_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16, dilations = var_525, groups = var_46, pad = var_527_pad_0, pad_type = var_527_pad_type_0, strides = var_523, weight = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_527_cast_fp16")];
            tensor<int32, [8]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_528_axis_0 = const()[name = string("op_528_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_528_cast_fp16_7 = split(axis = var_528_axis_0, split_sizes = tile_8, x = var_513_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<int32, [4]> var_537_perm_0 = const()[name = string("op_537_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_538_axis_0 = const()[name = string("op_538_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_6 = transpose(perm = var_537_perm_0, x = k_13_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_538_cast_fp16_7 = split(axis = var_538_axis_0, split_sizes = tile_9, x = transpose_6)[name = string("op_538_cast_fp16")];
            tensor<int32, [8]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_547_axis_0 = const()[name = string("op_547_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_547_cast_fp16_7 = split(axis = var_547_axis_0, split_sizes = tile_10, x = var_527_cast_fp16)[name = string("op_547_cast_fp16")];
            string var_557_equation_0 = const()[name = string("op_557_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_557_cast_fp16 = einsum(equation = var_557_equation_0, values = (var_538_cast_fp16_0, var_528_cast_fp16_0))[name = string("op_557_cast_fp16")];
            fp16 var_558_to_fp16 = const()[name = string("op_558_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_559_cast_fp16 = mul(x = var_557_cast_fp16, y = var_558_to_fp16)[name = string("op_559_cast_fp16")];
            string var_561_equation_0 = const()[name = string("op_561_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_561_cast_fp16 = einsum(equation = var_561_equation_0, values = (var_538_cast_fp16_1, var_528_cast_fp16_1))[name = string("op_561_cast_fp16")];
            fp16 var_562_to_fp16 = const()[name = string("op_562_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_563_cast_fp16 = mul(x = var_561_cast_fp16, y = var_562_to_fp16)[name = string("op_563_cast_fp16")];
            string var_565_equation_0 = const()[name = string("op_565_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_565_cast_fp16 = einsum(equation = var_565_equation_0, values = (var_538_cast_fp16_2, var_528_cast_fp16_2))[name = string("op_565_cast_fp16")];
            fp16 var_566_to_fp16 = const()[name = string("op_566_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_567_cast_fp16 = mul(x = var_565_cast_fp16, y = var_566_to_fp16)[name = string("op_567_cast_fp16")];
            string var_569_equation_0 = const()[name = string("op_569_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_569_cast_fp16 = einsum(equation = var_569_equation_0, values = (var_538_cast_fp16_3, var_528_cast_fp16_3))[name = string("op_569_cast_fp16")];
            fp16 var_570_to_fp16 = const()[name = string("op_570_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_571_cast_fp16 = mul(x = var_569_cast_fp16, y = var_570_to_fp16)[name = string("op_571_cast_fp16")];
            string var_573_equation_0 = const()[name = string("op_573_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_573_cast_fp16 = einsum(equation = var_573_equation_0, values = (var_538_cast_fp16_4, var_528_cast_fp16_4))[name = string("op_573_cast_fp16")];
            fp16 var_574_to_fp16 = const()[name = string("op_574_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_575_cast_fp16 = mul(x = var_573_cast_fp16, y = var_574_to_fp16)[name = string("op_575_cast_fp16")];
            string var_577_equation_0 = const()[name = string("op_577_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_577_cast_fp16 = einsum(equation = var_577_equation_0, values = (var_538_cast_fp16_5, var_528_cast_fp16_5))[name = string("op_577_cast_fp16")];
            fp16 var_578_to_fp16 = const()[name = string("op_578_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_579_cast_fp16 = mul(x = var_577_cast_fp16, y = var_578_to_fp16)[name = string("op_579_cast_fp16")];
            string var_581_equation_0 = const()[name = string("op_581_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_581_cast_fp16 = einsum(equation = var_581_equation_0, values = (var_538_cast_fp16_6, var_528_cast_fp16_6))[name = string("op_581_cast_fp16")];
            fp16 var_582_to_fp16 = const()[name = string("op_582_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_583_cast_fp16 = mul(x = var_581_cast_fp16, y = var_582_to_fp16)[name = string("op_583_cast_fp16")];
            string var_585_equation_0 = const()[name = string("op_585_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_585_cast_fp16 = einsum(equation = var_585_equation_0, values = (var_538_cast_fp16_7, var_528_cast_fp16_7))[name = string("op_585_cast_fp16")];
            fp16 var_586_to_fp16 = const()[name = string("op_586_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_587_cast_fp16 = mul(x = var_585_cast_fp16, y = var_586_to_fp16)[name = string("op_587_cast_fp16")];
            bool attn_weights_6_interleave_0 = const()[name = string("attn_weights_6_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_6_cast_fp16 = concat(axis = var_47, interleave = attn_weights_6_interleave_0, values = (var_559_cast_fp16, var_563_cast_fp16, var_567_cast_fp16, var_571_cast_fp16, var_575_cast_fp16, var_579_cast_fp16, var_583_cast_fp16, var_587_cast_fp16))[name = string("attn_weights_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_6_cast_fp16 = add(x = attn_weights_6_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_15_cast_fp16 = softmax(axis = var_46, x = attn_weights0_6_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<int32, [8]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_593_axis_0 = const()[name = string("op_593_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_593_cast_fp16_7 = split(axis = var_593_axis_0, split_sizes = tile_11, x = input_15_cast_fp16)[name = string("op_593_cast_fp16")];
            string var_603_equation_0 = const()[name = string("op_603_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_603_cast_fp16 = einsum(equation = var_603_equation_0, values = (var_547_cast_fp16_0, var_593_cast_fp16_0))[name = string("op_603_cast_fp16")];
            string var_605_equation_0 = const()[name = string("op_605_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_605_cast_fp16 = einsum(equation = var_605_equation_0, values = (var_547_cast_fp16_1, var_593_cast_fp16_1))[name = string("op_605_cast_fp16")];
            string var_607_equation_0 = const()[name = string("op_607_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_607_cast_fp16 = einsum(equation = var_607_equation_0, values = (var_547_cast_fp16_2, var_593_cast_fp16_2))[name = string("op_607_cast_fp16")];
            string var_609_equation_0 = const()[name = string("op_609_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_609_cast_fp16 = einsum(equation = var_609_equation_0, values = (var_547_cast_fp16_3, var_593_cast_fp16_3))[name = string("op_609_cast_fp16")];
            string var_611_equation_0 = const()[name = string("op_611_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_611_cast_fp16 = einsum(equation = var_611_equation_0, values = (var_547_cast_fp16_4, var_593_cast_fp16_4))[name = string("op_611_cast_fp16")];
            string var_613_equation_0 = const()[name = string("op_613_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_613_cast_fp16 = einsum(equation = var_613_equation_0, values = (var_547_cast_fp16_5, var_593_cast_fp16_5))[name = string("op_613_cast_fp16")];
            string var_615_equation_0 = const()[name = string("op_615_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_615_cast_fp16 = einsum(equation = var_615_equation_0, values = (var_547_cast_fp16_6, var_593_cast_fp16_6))[name = string("op_615_cast_fp16")];
            string var_617_equation_0 = const()[name = string("op_617_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_617_cast_fp16 = einsum(equation = var_617_equation_0, values = (var_547_cast_fp16_7, var_593_cast_fp16_7))[name = string("op_617_cast_fp16")];
            bool attn_18_interleave_0 = const()[name = string("attn_18_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_18_cast_fp16 = concat(axis = var_46, interleave = attn_18_interleave_0, values = (var_603_cast_fp16, var_605_cast_fp16, var_607_cast_fp16, var_609_cast_fp16, var_611_cast_fp16, var_613_cast_fp16, var_615_cast_fp16, var_617_cast_fp16))[name = string("attn_18_cast_fp16")];
            tensor<int32, [2]> var_625 = const()[name = string("op_625"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_627 = const()[name = string("op_627"), val = tensor<int32, [2]>([1, 1])];
            string attn_20_pad_type_0 = const()[name = string("attn_20_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_20_pad_0 = const()[name = string("attn_20_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33026112))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33025024))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33288320)))];
            tensor<fp16, [1, 512, 1, 256]> attn_20_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16, dilations = var_627, groups = var_46, pad = attn_20_pad_0, pad_type = attn_20_pad_type_0, strides = var_625, weight = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_18_cast_fp16)[name = string("attn_20_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_6_cast_fp16 = add(x = inputs0_2_cast_fp16, y = attn_20_cast_fp16)[name = string("inputs0_6_cast_fp16")];
            tensor<int32, [1]> input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_17_gamma_0_to_fp16 = const()[name = string("input_17_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33289408)))];
            tensor<fp16, [512]> input_17_beta_0_to_fp16 = const()[name = string("input_17_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33290496)))];
            fp16 var_639_to_fp16 = const()[name = string("op_639_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = input_17_beta_0_to_fp16, epsilon = var_639_to_fp16, gamma = input_17_gamma_0_to_fp16, x = inputs0_6_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<int32, [2]> var_653 = const()[name = string("op_653"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_655 = const()[name = string("op_655"), val = tensor<int32, [2]>([1, 1])];
            string x_6_pad_type_0 = const()[name = string("x_6_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_6_pad_0 = const()[name = string("x_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33295744))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33291584))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34348544))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34344384))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_6_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_655, groups = var_46, pad = x_6_pad_0, pad_type = x_6_pad_type_0, strides = var_653, weight = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_17_cast_fp16)[name = string("x_6_cast_fp16")];
            fp16 var_658_to_fp16 = const()[name = string("op_658_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_659_cast_fp16 = mul(x = x_6_cast_fp16, y = var_658_to_fp16)[name = string("op_659_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_660_cast_fp16 = sigmoid(x = var_659_cast_fp16)[name = string("op_660_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_19_cast_fp16 = mul(x = x_6_cast_fp16, y = var_660_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<int32, [2]> var_664 = const()[name = string("op_664"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_666 = const()[name = string("op_666"), val = tensor<int32, [2]>([1, 1])];
            string input0_11_pad_type_0 = const()[name = string("input0_11_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_11_pad_0 = const()[name = string("input0_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34351744))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34350656))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35400384)))];
            tensor<fp16, [1, 512, 1, 256]> input0_11_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16, dilations = var_666, groups = var_46, pad = input0_11_pad_0, pad_type = input0_11_pad_type_0, strides = var_664, weight = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_19_cast_fp16)[name = string("input0_11_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_670_cast_fp16 = add(x = input0_11_cast_fp16, y = inputs0_6_cast_fp16)[name = string("op_670_cast_fp16")];
            fp16 var_671_to_fp16 = const()[name = string("op_671_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_672_cast_fp16 = mul(x = inputs0_2_cast_fp16, y = var_671_to_fp16)[name = string("op_672_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs1_1_cast_fp16 = add(x = var_672_cast_fp16, y = var_670_cast_fp16)[name = string("inputs1_1_cast_fp16")];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_15_gamma_0_to_fp16 = const()[name = string("k_15_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35401472)))];
            tensor<fp16, [512]> k_15_beta_0_to_fp16 = const()[name = string("k_15_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35402560)))];
            fp16 var_688_to_fp16 = const()[name = string("op_688_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_15_cast_fp16 = layer_norm(axes = k_15_axes_0, beta = k_15_beta_0_to_fp16, epsilon = var_688_to_fp16, gamma = k_15_gamma_0_to_fp16, x = inputs1_1_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [2]> var_707 = const()[name = string("op_707"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_709 = const()[name = string("op_709"), val = tensor<int32, [2]>([1, 1])];
            string var_711_pad_type_0 = const()[name = string("op_711_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_711_pad_0 = const()[name = string("op_711_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35404736))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35403648))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35666944)))];
            tensor<fp16, [1, 512, 1, 256]> var_711_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16, dilations = var_709, groups = var_46, pad = var_711_pad_0, pad_type = var_711_pad_type_0, strides = var_707, weight = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_711_cast_fp16")];
            tensor<int32, [2]> var_714 = const()[name = string("op_714"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_716 = const()[name = string("op_716"), val = tensor<int32, [2]>([1, 1])];
            string k_17_pad_type_0 = const()[name = string("k_17_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_17_pad_0 = const()[name = string("k_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35669120))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35668032))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35931328)))];
            tensor<fp16, [1, 512, 1, 256]> k_17_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16, dilations = var_716, groups = var_46, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = var_714, weight = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [2]> var_721 = const()[name = string("op_721"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_723 = const()[name = string("op_723"), val = tensor<int32, [2]>([1, 1])];
            string var_725_pad_type_0 = const()[name = string("op_725_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_725_pad_0 = const()[name = string("op_725_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35933504))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35932416))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36195712)))];
            tensor<fp16, [1, 512, 1, 256]> var_725_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16, dilations = var_723, groups = var_46, pad = var_725_pad_0, pad_type = var_725_pad_type_0, strides = var_721, weight = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<int32, [8]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_726_axis_0 = const()[name = string("op_726_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_726_cast_fp16_7 = split(axis = var_726_axis_0, split_sizes = tile_12, x = var_711_cast_fp16)[name = string("op_726_cast_fp16")];
            tensor<int32, [4]> var_735_perm_0 = const()[name = string("op_735_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_736_axis_0 = const()[name = string("op_736_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_5 = transpose(perm = var_735_perm_0, x = k_17_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_736_cast_fp16_7 = split(axis = var_736_axis_0, split_sizes = tile_13, x = transpose_5)[name = string("op_736_cast_fp16")];
            tensor<int32, [8]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_745_axis_0 = const()[name = string("op_745_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_745_cast_fp16_7 = split(axis = var_745_axis_0, split_sizes = tile_14, x = var_725_cast_fp16)[name = string("op_745_cast_fp16")];
            string var_755_equation_0 = const()[name = string("op_755_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_755_cast_fp16 = einsum(equation = var_755_equation_0, values = (var_736_cast_fp16_0, var_726_cast_fp16_0))[name = string("op_755_cast_fp16")];
            fp16 var_756_to_fp16 = const()[name = string("op_756_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_757_cast_fp16 = mul(x = var_755_cast_fp16, y = var_756_to_fp16)[name = string("op_757_cast_fp16")];
            string var_759_equation_0 = const()[name = string("op_759_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16 = einsum(equation = var_759_equation_0, values = (var_736_cast_fp16_1, var_726_cast_fp16_1))[name = string("op_759_cast_fp16")];
            fp16 var_760_to_fp16 = const()[name = string("op_760_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_761_cast_fp16 = mul(x = var_759_cast_fp16, y = var_760_to_fp16)[name = string("op_761_cast_fp16")];
            string var_763_equation_0 = const()[name = string("op_763_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_763_cast_fp16 = einsum(equation = var_763_equation_0, values = (var_736_cast_fp16_2, var_726_cast_fp16_2))[name = string("op_763_cast_fp16")];
            fp16 var_764_to_fp16 = const()[name = string("op_764_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_765_cast_fp16 = mul(x = var_763_cast_fp16, y = var_764_to_fp16)[name = string("op_765_cast_fp16")];
            string var_767_equation_0 = const()[name = string("op_767_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_767_cast_fp16 = einsum(equation = var_767_equation_0, values = (var_736_cast_fp16_3, var_726_cast_fp16_3))[name = string("op_767_cast_fp16")];
            fp16 var_768_to_fp16 = const()[name = string("op_768_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_769_cast_fp16 = mul(x = var_767_cast_fp16, y = var_768_to_fp16)[name = string("op_769_cast_fp16")];
            string var_771_equation_0 = const()[name = string("op_771_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_771_cast_fp16 = einsum(equation = var_771_equation_0, values = (var_736_cast_fp16_4, var_726_cast_fp16_4))[name = string("op_771_cast_fp16")];
            fp16 var_772_to_fp16 = const()[name = string("op_772_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_773_cast_fp16 = mul(x = var_771_cast_fp16, y = var_772_to_fp16)[name = string("op_773_cast_fp16")];
            string var_775_equation_0 = const()[name = string("op_775_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_775_cast_fp16 = einsum(equation = var_775_equation_0, values = (var_736_cast_fp16_5, var_726_cast_fp16_5))[name = string("op_775_cast_fp16")];
            fp16 var_776_to_fp16 = const()[name = string("op_776_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_777_cast_fp16 = mul(x = var_775_cast_fp16, y = var_776_to_fp16)[name = string("op_777_cast_fp16")];
            string var_779_equation_0 = const()[name = string("op_779_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_779_cast_fp16 = einsum(equation = var_779_equation_0, values = (var_736_cast_fp16_6, var_726_cast_fp16_6))[name = string("op_779_cast_fp16")];
            fp16 var_780_to_fp16 = const()[name = string("op_780_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_781_cast_fp16 = mul(x = var_779_cast_fp16, y = var_780_to_fp16)[name = string("op_781_cast_fp16")];
            string var_783_equation_0 = const()[name = string("op_783_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_783_cast_fp16 = einsum(equation = var_783_equation_0, values = (var_736_cast_fp16_7, var_726_cast_fp16_7))[name = string("op_783_cast_fp16")];
            fp16 var_784_to_fp16 = const()[name = string("op_784_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_785_cast_fp16 = mul(x = var_783_cast_fp16, y = var_784_to_fp16)[name = string("op_785_cast_fp16")];
            bool attn_weights_8_interleave_0 = const()[name = string("attn_weights_8_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_8_cast_fp16 = concat(axis = var_47, interleave = attn_weights_8_interleave_0, values = (var_757_cast_fp16, var_761_cast_fp16, var_765_cast_fp16, var_769_cast_fp16, var_773_cast_fp16, var_777_cast_fp16, var_781_cast_fp16, var_785_cast_fp16))[name = string("attn_weights_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_8_cast_fp16 = add(x = attn_weights_8_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_21_cast_fp16 = softmax(axis = var_46, x = attn_weights0_8_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [8]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_791_axis_0 = const()[name = string("op_791_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_791_cast_fp16_7 = split(axis = var_791_axis_0, split_sizes = tile_15, x = input_21_cast_fp16)[name = string("op_791_cast_fp16")];
            string var_801_equation_0 = const()[name = string("op_801_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_801_cast_fp16 = einsum(equation = var_801_equation_0, values = (var_745_cast_fp16_0, var_791_cast_fp16_0))[name = string("op_801_cast_fp16")];
            string var_803_equation_0 = const()[name = string("op_803_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_803_cast_fp16 = einsum(equation = var_803_equation_0, values = (var_745_cast_fp16_1, var_791_cast_fp16_1))[name = string("op_803_cast_fp16")];
            string var_805_equation_0 = const()[name = string("op_805_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_805_cast_fp16 = einsum(equation = var_805_equation_0, values = (var_745_cast_fp16_2, var_791_cast_fp16_2))[name = string("op_805_cast_fp16")];
            string var_807_equation_0 = const()[name = string("op_807_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_807_cast_fp16 = einsum(equation = var_807_equation_0, values = (var_745_cast_fp16_3, var_791_cast_fp16_3))[name = string("op_807_cast_fp16")];
            string var_809_equation_0 = const()[name = string("op_809_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_809_cast_fp16 = einsum(equation = var_809_equation_0, values = (var_745_cast_fp16_4, var_791_cast_fp16_4))[name = string("op_809_cast_fp16")];
            string var_811_equation_0 = const()[name = string("op_811_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_811_cast_fp16 = einsum(equation = var_811_equation_0, values = (var_745_cast_fp16_5, var_791_cast_fp16_5))[name = string("op_811_cast_fp16")];
            string var_813_equation_0 = const()[name = string("op_813_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_813_cast_fp16 = einsum(equation = var_813_equation_0, values = (var_745_cast_fp16_6, var_791_cast_fp16_6))[name = string("op_813_cast_fp16")];
            string var_815_equation_0 = const()[name = string("op_815_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_815_cast_fp16 = einsum(equation = var_815_equation_0, values = (var_745_cast_fp16_7, var_791_cast_fp16_7))[name = string("op_815_cast_fp16")];
            bool attn_24_interleave_0 = const()[name = string("attn_24_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_24_cast_fp16 = concat(axis = var_46, interleave = attn_24_interleave_0, values = (var_801_cast_fp16, var_803_cast_fp16, var_805_cast_fp16, var_807_cast_fp16, var_809_cast_fp16, var_811_cast_fp16, var_813_cast_fp16, var_815_cast_fp16))[name = string("attn_24_cast_fp16")];
            tensor<int32, [2]> var_823 = const()[name = string("op_823"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_825 = const()[name = string("op_825"), val = tensor<int32, [2]>([1, 1])];
            string attn_26_pad_type_0 = const()[name = string("attn_26_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_26_pad_0 = const()[name = string("attn_26_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36197888))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36196800))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36460096)))];
            tensor<fp16, [1, 512, 1, 256]> attn_26_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16, dilations = var_825, groups = var_46, pad = attn_26_pad_0, pad_type = attn_26_pad_type_0, strides = var_823, weight = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_24_cast_fp16)[name = string("attn_26_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_8_cast_fp16 = add(x = inputs1_1_cast_fp16, y = attn_26_cast_fp16)[name = string("inputs0_8_cast_fp16")];
            tensor<int32, [1]> input_23_axes_0 = const()[name = string("input_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_23_gamma_0_to_fp16 = const()[name = string("input_23_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36461184)))];
            tensor<fp16, [512]> input_23_beta_0_to_fp16 = const()[name = string("input_23_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36462272)))];
            fp16 var_837_to_fp16 = const()[name = string("op_837_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = input_23_beta_0_to_fp16, epsilon = var_837_to_fp16, gamma = input_23_gamma_0_to_fp16, x = inputs0_8_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<int32, [2]> var_851 = const()[name = string("op_851"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_853 = const()[name = string("op_853"), val = tensor<int32, [2]>([1, 1])];
            string x_8_pad_type_0 = const()[name = string("x_8_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_8_pad_0 = const()[name = string("x_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36467520))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36463360))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37520320))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37516160))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_8_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_853, groups = var_46, pad = x_8_pad_0, pad_type = x_8_pad_type_0, strides = var_851, weight = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_23_cast_fp16)[name = string("x_8_cast_fp16")];
            fp16 var_856_to_fp16 = const()[name = string("op_856_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_857_cast_fp16 = mul(x = x_8_cast_fp16, y = var_856_to_fp16)[name = string("op_857_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_858_cast_fp16 = sigmoid(x = var_857_cast_fp16)[name = string("op_858_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_25_cast_fp16 = mul(x = x_8_cast_fp16, y = var_858_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<int32, [2]> var_862 = const()[name = string("op_862"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_864 = const()[name = string("op_864"), val = tensor<int32, [2]>([1, 1])];
            string input0_15_pad_type_0 = const()[name = string("input0_15_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_15_pad_0 = const()[name = string("input0_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37523520))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37522432))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38572160)))];
            tensor<fp16, [1, 512, 1, 256]> input0_15_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16, dilations = var_864, groups = var_46, pad = input0_15_pad_0, pad_type = input0_15_pad_type_0, strides = var_862, weight = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_25_cast_fp16)[name = string("input0_15_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_868_cast_fp16 = add(x = input0_15_cast_fp16, y = inputs0_8_cast_fp16)[name = string("op_868_cast_fp16")];
            fp16 var_869_to_fp16 = const()[name = string("op_869_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_870_cast_fp16 = mul(x = inputs1_1_cast_fp16, y = var_869_to_fp16)[name = string("op_870_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs2_1_cast_fp16 = add(x = var_870_cast_fp16, y = var_868_cast_fp16)[name = string("inputs2_1_cast_fp16")];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_19_gamma_0_to_fp16 = const()[name = string("k_19_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38573248)))];
            tensor<fp16, [512]> k_19_beta_0_to_fp16 = const()[name = string("k_19_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38574336)))];
            fp16 var_886_to_fp16 = const()[name = string("op_886_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_19_cast_fp16 = layer_norm(axes = k_19_axes_0, beta = k_19_beta_0_to_fp16, epsilon = var_886_to_fp16, gamma = k_19_gamma_0_to_fp16, x = inputs2_1_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [2]> var_905 = const()[name = string("op_905"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_907 = const()[name = string("op_907"), val = tensor<int32, [2]>([1, 1])];
            string var_909_pad_type_0 = const()[name = string("op_909_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_909_pad_0 = const()[name = string("op_909_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38576512))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38575424))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38838720)))];
            tensor<fp16, [1, 512, 1, 256]> var_909_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16, dilations = var_907, groups = var_46, pad = var_909_pad_0, pad_type = var_909_pad_type_0, strides = var_905, weight = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_909_cast_fp16")];
            tensor<int32, [2]> var_912 = const()[name = string("op_912"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_914 = const()[name = string("op_914"), val = tensor<int32, [2]>([1, 1])];
            string k_21_pad_type_0 = const()[name = string("k_21_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_21_pad_0 = const()[name = string("k_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38840896))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38839808))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39103104)))];
            tensor<fp16, [1, 512, 1, 256]> k_21_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16, dilations = var_914, groups = var_46, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = var_912, weight = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [2]> var_919 = const()[name = string("op_919"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_921 = const()[name = string("op_921"), val = tensor<int32, [2]>([1, 1])];
            string var_923_pad_type_0 = const()[name = string("op_923_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_923_pad_0 = const()[name = string("op_923_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39105280))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39104192))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39367488)))];
            tensor<fp16, [1, 512, 1, 256]> var_923_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16, dilations = var_921, groups = var_46, pad = var_923_pad_0, pad_type = var_923_pad_type_0, strides = var_919, weight = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_923_cast_fp16")];
            tensor<int32, [8]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_924_axis_0 = const()[name = string("op_924_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_924_cast_fp16_7 = split(axis = var_924_axis_0, split_sizes = tile_16, x = var_909_cast_fp16)[name = string("op_924_cast_fp16")];
            tensor<int32, [4]> var_933_perm_0 = const()[name = string("op_933_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_934_axis_0 = const()[name = string("op_934_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_4 = transpose(perm = var_933_perm_0, x = k_21_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_934_cast_fp16_7 = split(axis = var_934_axis_0, split_sizes = tile_17, x = transpose_4)[name = string("op_934_cast_fp16")];
            tensor<int32, [8]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_943_axis_0 = const()[name = string("op_943_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_943_cast_fp16_7 = split(axis = var_943_axis_0, split_sizes = tile_18, x = var_923_cast_fp16)[name = string("op_943_cast_fp16")];
            string var_953_equation_0 = const()[name = string("op_953_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_953_cast_fp16 = einsum(equation = var_953_equation_0, values = (var_934_cast_fp16_0, var_924_cast_fp16_0))[name = string("op_953_cast_fp16")];
            fp16 var_954_to_fp16 = const()[name = string("op_954_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_955_cast_fp16 = mul(x = var_953_cast_fp16, y = var_954_to_fp16)[name = string("op_955_cast_fp16")];
            string var_957_equation_0 = const()[name = string("op_957_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_957_cast_fp16 = einsum(equation = var_957_equation_0, values = (var_934_cast_fp16_1, var_924_cast_fp16_1))[name = string("op_957_cast_fp16")];
            fp16 var_958_to_fp16 = const()[name = string("op_958_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_959_cast_fp16 = mul(x = var_957_cast_fp16, y = var_958_to_fp16)[name = string("op_959_cast_fp16")];
            string var_961_equation_0 = const()[name = string("op_961_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_961_cast_fp16 = einsum(equation = var_961_equation_0, values = (var_934_cast_fp16_2, var_924_cast_fp16_2))[name = string("op_961_cast_fp16")];
            fp16 var_962_to_fp16 = const()[name = string("op_962_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_963_cast_fp16 = mul(x = var_961_cast_fp16, y = var_962_to_fp16)[name = string("op_963_cast_fp16")];
            string var_965_equation_0 = const()[name = string("op_965_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_965_cast_fp16 = einsum(equation = var_965_equation_0, values = (var_934_cast_fp16_3, var_924_cast_fp16_3))[name = string("op_965_cast_fp16")];
            fp16 var_966_to_fp16 = const()[name = string("op_966_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_967_cast_fp16 = mul(x = var_965_cast_fp16, y = var_966_to_fp16)[name = string("op_967_cast_fp16")];
            string var_969_equation_0 = const()[name = string("op_969_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_969_cast_fp16 = einsum(equation = var_969_equation_0, values = (var_934_cast_fp16_4, var_924_cast_fp16_4))[name = string("op_969_cast_fp16")];
            fp16 var_970_to_fp16 = const()[name = string("op_970_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_971_cast_fp16 = mul(x = var_969_cast_fp16, y = var_970_to_fp16)[name = string("op_971_cast_fp16")];
            string var_973_equation_0 = const()[name = string("op_973_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_973_cast_fp16 = einsum(equation = var_973_equation_0, values = (var_934_cast_fp16_5, var_924_cast_fp16_5))[name = string("op_973_cast_fp16")];
            fp16 var_974_to_fp16 = const()[name = string("op_974_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_975_cast_fp16 = mul(x = var_973_cast_fp16, y = var_974_to_fp16)[name = string("op_975_cast_fp16")];
            string var_977_equation_0 = const()[name = string("op_977_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_977_cast_fp16 = einsum(equation = var_977_equation_0, values = (var_934_cast_fp16_6, var_924_cast_fp16_6))[name = string("op_977_cast_fp16")];
            fp16 var_978_to_fp16 = const()[name = string("op_978_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_979_cast_fp16 = mul(x = var_977_cast_fp16, y = var_978_to_fp16)[name = string("op_979_cast_fp16")];
            string var_981_equation_0 = const()[name = string("op_981_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_981_cast_fp16 = einsum(equation = var_981_equation_0, values = (var_934_cast_fp16_7, var_924_cast_fp16_7))[name = string("op_981_cast_fp16")];
            fp16 var_982_to_fp16 = const()[name = string("op_982_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_983_cast_fp16 = mul(x = var_981_cast_fp16, y = var_982_to_fp16)[name = string("op_983_cast_fp16")];
            bool attn_weights_10_interleave_0 = const()[name = string("attn_weights_10_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_10_cast_fp16 = concat(axis = var_47, interleave = attn_weights_10_interleave_0, values = (var_955_cast_fp16, var_959_cast_fp16, var_963_cast_fp16, var_967_cast_fp16, var_971_cast_fp16, var_975_cast_fp16, var_979_cast_fp16, var_983_cast_fp16))[name = string("attn_weights_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_10_cast_fp16 = add(x = attn_weights_10_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_27_cast_fp16 = softmax(axis = var_46, x = attn_weights0_10_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [8]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_989_axis_0 = const()[name = string("op_989_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_989_cast_fp16_7 = split(axis = var_989_axis_0, split_sizes = tile_19, x = input_27_cast_fp16)[name = string("op_989_cast_fp16")];
            string var_999_equation_0 = const()[name = string("op_999_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_999_cast_fp16 = einsum(equation = var_999_equation_0, values = (var_943_cast_fp16_0, var_989_cast_fp16_0))[name = string("op_999_cast_fp16")];
            string var_1001_equation_0 = const()[name = string("op_1001_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1001_cast_fp16 = einsum(equation = var_1001_equation_0, values = (var_943_cast_fp16_1, var_989_cast_fp16_1))[name = string("op_1001_cast_fp16")];
            string var_1003_equation_0 = const()[name = string("op_1003_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1003_cast_fp16 = einsum(equation = var_1003_equation_0, values = (var_943_cast_fp16_2, var_989_cast_fp16_2))[name = string("op_1003_cast_fp16")];
            string var_1005_equation_0 = const()[name = string("op_1005_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1005_cast_fp16 = einsum(equation = var_1005_equation_0, values = (var_943_cast_fp16_3, var_989_cast_fp16_3))[name = string("op_1005_cast_fp16")];
            string var_1007_equation_0 = const()[name = string("op_1007_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1007_cast_fp16 = einsum(equation = var_1007_equation_0, values = (var_943_cast_fp16_4, var_989_cast_fp16_4))[name = string("op_1007_cast_fp16")];
            string var_1009_equation_0 = const()[name = string("op_1009_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1009_cast_fp16 = einsum(equation = var_1009_equation_0, values = (var_943_cast_fp16_5, var_989_cast_fp16_5))[name = string("op_1009_cast_fp16")];
            string var_1011_equation_0 = const()[name = string("op_1011_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1011_cast_fp16 = einsum(equation = var_1011_equation_0, values = (var_943_cast_fp16_6, var_989_cast_fp16_6))[name = string("op_1011_cast_fp16")];
            string var_1013_equation_0 = const()[name = string("op_1013_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1013_cast_fp16 = einsum(equation = var_1013_equation_0, values = (var_943_cast_fp16_7, var_989_cast_fp16_7))[name = string("op_1013_cast_fp16")];
            bool attn_30_interleave_0 = const()[name = string("attn_30_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_30_cast_fp16 = concat(axis = var_46, interleave = attn_30_interleave_0, values = (var_999_cast_fp16, var_1001_cast_fp16, var_1003_cast_fp16, var_1005_cast_fp16, var_1007_cast_fp16, var_1009_cast_fp16, var_1011_cast_fp16, var_1013_cast_fp16))[name = string("attn_30_cast_fp16")];
            tensor<int32, [2]> var_1021 = const()[name = string("op_1021"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1023 = const()[name = string("op_1023"), val = tensor<int32, [2]>([1, 1])];
            string attn_32_pad_type_0 = const()[name = string("attn_32_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_32_pad_0 = const()[name = string("attn_32_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39369664))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39368576))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39631872)))];
            tensor<fp16, [1, 512, 1, 256]> attn_32_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16, dilations = var_1023, groups = var_46, pad = attn_32_pad_0, pad_type = attn_32_pad_type_0, strides = var_1021, weight = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_30_cast_fp16)[name = string("attn_32_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_10_cast_fp16 = add(x = inputs2_1_cast_fp16, y = attn_32_cast_fp16)[name = string("inputs0_10_cast_fp16")];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_29_gamma_0_to_fp16 = const()[name = string("input_29_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39632960)))];
            tensor<fp16, [512]> input_29_beta_0_to_fp16 = const()[name = string("input_29_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39634048)))];
            fp16 var_1035_to_fp16 = const()[name = string("op_1035_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = input_29_beta_0_to_fp16, epsilon = var_1035_to_fp16, gamma = input_29_gamma_0_to_fp16, x = inputs0_10_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [2]> var_1049 = const()[name = string("op_1049"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1051 = const()[name = string("op_1051"), val = tensor<int32, [2]>([1, 1])];
            string x_10_pad_type_0 = const()[name = string("x_10_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_10_pad_0 = const()[name = string("x_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39639296))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39635136))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40692096))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40687936))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_10_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1051, groups = var_46, pad = x_10_pad_0, pad_type = x_10_pad_type_0, strides = var_1049, weight = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_29_cast_fp16)[name = string("x_10_cast_fp16")];
            fp16 var_1054_to_fp16 = const()[name = string("op_1054_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1055_cast_fp16 = mul(x = x_10_cast_fp16, y = var_1054_to_fp16)[name = string("op_1055_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1056_cast_fp16 = sigmoid(x = var_1055_cast_fp16)[name = string("op_1056_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_31_cast_fp16 = mul(x = x_10_cast_fp16, y = var_1056_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<int32, [2]> var_1060 = const()[name = string("op_1060"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1062 = const()[name = string("op_1062"), val = tensor<int32, [2]>([1, 1])];
            string input0_19_pad_type_0 = const()[name = string("input0_19_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_19_pad_0 = const()[name = string("input0_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40695296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40694208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41743936)))];
            tensor<fp16, [1, 512, 1, 256]> input0_19_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16, dilations = var_1062, groups = var_46, pad = input0_19_pad_0, pad_type = input0_19_pad_type_0, strides = var_1060, weight = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_31_cast_fp16)[name = string("input0_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1066_cast_fp16 = add(x = input0_19_cast_fp16, y = inputs0_10_cast_fp16)[name = string("op_1066_cast_fp16")];
            fp16 var_1067_to_fp16 = const()[name = string("op_1067_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1068_cast_fp16 = mul(x = inputs2_1_cast_fp16, y = var_1067_to_fp16)[name = string("op_1068_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs3_1_cast_fp16 = add(x = var_1068_cast_fp16, y = var_1066_cast_fp16)[name = string("inputs3_1_cast_fp16")];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_23_gamma_0_to_fp16 = const()[name = string("k_23_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41745024)))];
            tensor<fp16, [512]> k_23_beta_0_to_fp16 = const()[name = string("k_23_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41746112)))];
            fp16 var_1084_to_fp16 = const()[name = string("op_1084_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_23_cast_fp16 = layer_norm(axes = k_23_axes_0, beta = k_23_beta_0_to_fp16, epsilon = var_1084_to_fp16, gamma = k_23_gamma_0_to_fp16, x = inputs3_1_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [2]> var_1103 = const()[name = string("op_1103"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1105 = const()[name = string("op_1105"), val = tensor<int32, [2]>([1, 1])];
            string var_1107_pad_type_0 = const()[name = string("op_1107_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1107_pad_0 = const()[name = string("op_1107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41748288))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41747200))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42010496)))];
            tensor<fp16, [1, 512, 1, 256]> var_1107_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16, dilations = var_1105, groups = var_46, pad = var_1107_pad_0, pad_type = var_1107_pad_type_0, strides = var_1103, weight = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1107_cast_fp16")];
            tensor<int32, [2]> var_1110 = const()[name = string("op_1110"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1112 = const()[name = string("op_1112"), val = tensor<int32, [2]>([1, 1])];
            string k_25_pad_type_0 = const()[name = string("k_25_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_25_pad_0 = const()[name = string("k_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42012672))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42011584))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42274880)))];
            tensor<fp16, [1, 512, 1, 256]> k_25_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16, dilations = var_1112, groups = var_46, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = var_1110, weight = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [2]> var_1117 = const()[name = string("op_1117"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1119 = const()[name = string("op_1119"), val = tensor<int32, [2]>([1, 1])];
            string var_1121_pad_type_0 = const()[name = string("op_1121_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1121_pad_0 = const()[name = string("op_1121_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42277056))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42275968))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42539264)))];
            tensor<fp16, [1, 512, 1, 256]> var_1121_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16, dilations = var_1119, groups = var_46, pad = var_1121_pad_0, pad_type = var_1121_pad_type_0, strides = var_1117, weight = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1121_cast_fp16")];
            tensor<int32, [8]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1122_axis_0 = const()[name = string("op_1122_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1122_cast_fp16_7 = split(axis = var_1122_axis_0, split_sizes = tile_20, x = var_1107_cast_fp16)[name = string("op_1122_cast_fp16")];
            tensor<int32, [4]> var_1131_perm_0 = const()[name = string("op_1131_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1132_axis_0 = const()[name = string("op_1132_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_3 = transpose(perm = var_1131_perm_0, x = k_25_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1132_cast_fp16_7 = split(axis = var_1132_axis_0, split_sizes = tile_21, x = transpose_3)[name = string("op_1132_cast_fp16")];
            tensor<int32, [8]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1141_axis_0 = const()[name = string("op_1141_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1141_cast_fp16_7 = split(axis = var_1141_axis_0, split_sizes = tile_22, x = var_1121_cast_fp16)[name = string("op_1141_cast_fp16")];
            string var_1151_equation_0 = const()[name = string("op_1151_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1151_cast_fp16 = einsum(equation = var_1151_equation_0, values = (var_1132_cast_fp16_0, var_1122_cast_fp16_0))[name = string("op_1151_cast_fp16")];
            fp16 var_1152_to_fp16 = const()[name = string("op_1152_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = string("op_1153_cast_fp16")];
            string var_1155_equation_0 = const()[name = string("op_1155_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1155_cast_fp16 = einsum(equation = var_1155_equation_0, values = (var_1132_cast_fp16_1, var_1122_cast_fp16_1))[name = string("op_1155_cast_fp16")];
            fp16 var_1156_to_fp16 = const()[name = string("op_1156_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1157_cast_fp16 = mul(x = var_1155_cast_fp16, y = var_1156_to_fp16)[name = string("op_1157_cast_fp16")];
            string var_1159_equation_0 = const()[name = string("op_1159_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1159_cast_fp16 = einsum(equation = var_1159_equation_0, values = (var_1132_cast_fp16_2, var_1122_cast_fp16_2))[name = string("op_1159_cast_fp16")];
            fp16 var_1160_to_fp16 = const()[name = string("op_1160_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1161_cast_fp16 = mul(x = var_1159_cast_fp16, y = var_1160_to_fp16)[name = string("op_1161_cast_fp16")];
            string var_1163_equation_0 = const()[name = string("op_1163_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1163_cast_fp16 = einsum(equation = var_1163_equation_0, values = (var_1132_cast_fp16_3, var_1122_cast_fp16_3))[name = string("op_1163_cast_fp16")];
            fp16 var_1164_to_fp16 = const()[name = string("op_1164_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1165_cast_fp16 = mul(x = var_1163_cast_fp16, y = var_1164_to_fp16)[name = string("op_1165_cast_fp16")];
            string var_1167_equation_0 = const()[name = string("op_1167_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1167_cast_fp16 = einsum(equation = var_1167_equation_0, values = (var_1132_cast_fp16_4, var_1122_cast_fp16_4))[name = string("op_1167_cast_fp16")];
            fp16 var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1169_cast_fp16 = mul(x = var_1167_cast_fp16, y = var_1168_to_fp16)[name = string("op_1169_cast_fp16")];
            string var_1171_equation_0 = const()[name = string("op_1171_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1171_cast_fp16 = einsum(equation = var_1171_equation_0, values = (var_1132_cast_fp16_5, var_1122_cast_fp16_5))[name = string("op_1171_cast_fp16")];
            fp16 var_1172_to_fp16 = const()[name = string("op_1172_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1173_cast_fp16 = mul(x = var_1171_cast_fp16, y = var_1172_to_fp16)[name = string("op_1173_cast_fp16")];
            string var_1175_equation_0 = const()[name = string("op_1175_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1175_cast_fp16 = einsum(equation = var_1175_equation_0, values = (var_1132_cast_fp16_6, var_1122_cast_fp16_6))[name = string("op_1175_cast_fp16")];
            fp16 var_1176_to_fp16 = const()[name = string("op_1176_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1177_cast_fp16 = mul(x = var_1175_cast_fp16, y = var_1176_to_fp16)[name = string("op_1177_cast_fp16")];
            string var_1179_equation_0 = const()[name = string("op_1179_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1179_cast_fp16 = einsum(equation = var_1179_equation_0, values = (var_1132_cast_fp16_7, var_1122_cast_fp16_7))[name = string("op_1179_cast_fp16")];
            fp16 var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1181_cast_fp16 = mul(x = var_1179_cast_fp16, y = var_1180_to_fp16)[name = string("op_1181_cast_fp16")];
            bool attn_weights_12_interleave_0 = const()[name = string("attn_weights_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_12_cast_fp16 = concat(axis = var_47, interleave = attn_weights_12_interleave_0, values = (var_1153_cast_fp16, var_1157_cast_fp16, var_1161_cast_fp16, var_1165_cast_fp16, var_1169_cast_fp16, var_1173_cast_fp16, var_1177_cast_fp16, var_1181_cast_fp16))[name = string("attn_weights_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_12_cast_fp16 = add(x = attn_weights_12_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_33_cast_fp16 = softmax(axis = var_46, x = attn_weights0_12_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [8]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1187_axis_0 = const()[name = string("op_1187_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1187_cast_fp16_7 = split(axis = var_1187_axis_0, split_sizes = tile_23, x = input_33_cast_fp16)[name = string("op_1187_cast_fp16")];
            string var_1197_equation_0 = const()[name = string("op_1197_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1197_cast_fp16 = einsum(equation = var_1197_equation_0, values = (var_1141_cast_fp16_0, var_1187_cast_fp16_0))[name = string("op_1197_cast_fp16")];
            string var_1199_equation_0 = const()[name = string("op_1199_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1199_cast_fp16 = einsum(equation = var_1199_equation_0, values = (var_1141_cast_fp16_1, var_1187_cast_fp16_1))[name = string("op_1199_cast_fp16")];
            string var_1201_equation_0 = const()[name = string("op_1201_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1201_cast_fp16 = einsum(equation = var_1201_equation_0, values = (var_1141_cast_fp16_2, var_1187_cast_fp16_2))[name = string("op_1201_cast_fp16")];
            string var_1203_equation_0 = const()[name = string("op_1203_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1203_cast_fp16 = einsum(equation = var_1203_equation_0, values = (var_1141_cast_fp16_3, var_1187_cast_fp16_3))[name = string("op_1203_cast_fp16")];
            string var_1205_equation_0 = const()[name = string("op_1205_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1205_cast_fp16 = einsum(equation = var_1205_equation_0, values = (var_1141_cast_fp16_4, var_1187_cast_fp16_4))[name = string("op_1205_cast_fp16")];
            string var_1207_equation_0 = const()[name = string("op_1207_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1207_cast_fp16 = einsum(equation = var_1207_equation_0, values = (var_1141_cast_fp16_5, var_1187_cast_fp16_5))[name = string("op_1207_cast_fp16")];
            string var_1209_equation_0 = const()[name = string("op_1209_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1209_cast_fp16 = einsum(equation = var_1209_equation_0, values = (var_1141_cast_fp16_6, var_1187_cast_fp16_6))[name = string("op_1209_cast_fp16")];
            string var_1211_equation_0 = const()[name = string("op_1211_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1211_cast_fp16 = einsum(equation = var_1211_equation_0, values = (var_1141_cast_fp16_7, var_1187_cast_fp16_7))[name = string("op_1211_cast_fp16")];
            bool attn_36_interleave_0 = const()[name = string("attn_36_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_36_cast_fp16 = concat(axis = var_46, interleave = attn_36_interleave_0, values = (var_1197_cast_fp16, var_1199_cast_fp16, var_1201_cast_fp16, var_1203_cast_fp16, var_1205_cast_fp16, var_1207_cast_fp16, var_1209_cast_fp16, var_1211_cast_fp16))[name = string("attn_36_cast_fp16")];
            tensor<int32, [2]> var_1219 = const()[name = string("op_1219"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [2]>([1, 1])];
            string attn_38_pad_type_0 = const()[name = string("attn_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_38_pad_0 = const()[name = string("attn_38_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42541440))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42540352))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42803648)))];
            tensor<fp16, [1, 512, 1, 256]> attn_38_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16, dilations = var_1221, groups = var_46, pad = attn_38_pad_0, pad_type = attn_38_pad_type_0, strides = var_1219, weight = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_36_cast_fp16)[name = string("attn_38_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_12_cast_fp16 = add(x = inputs3_1_cast_fp16, y = attn_38_cast_fp16)[name = string("inputs0_12_cast_fp16")];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_35_gamma_0_to_fp16 = const()[name = string("input_35_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42804736)))];
            tensor<fp16, [512]> input_35_beta_0_to_fp16 = const()[name = string("input_35_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42805824)))];
            fp16 var_1233_to_fp16 = const()[name = string("op_1233_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = input_35_beta_0_to_fp16, epsilon = var_1233_to_fp16, gamma = input_35_gamma_0_to_fp16, x = inputs0_12_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<int32, [2]> var_1247 = const()[name = string("op_1247"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1249 = const()[name = string("op_1249"), val = tensor<int32, [2]>([1, 1])];
            string x_12_pad_type_0 = const()[name = string("x_12_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_12_pad_0 = const()[name = string("x_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42811072))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42806912))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43863872))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43859712))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_12_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1249, groups = var_46, pad = x_12_pad_0, pad_type = x_12_pad_type_0, strides = var_1247, weight = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_35_cast_fp16)[name = string("x_12_cast_fp16")];
            fp16 var_1252_to_fp16 = const()[name = string("op_1252_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1253_cast_fp16 = mul(x = x_12_cast_fp16, y = var_1252_to_fp16)[name = string("op_1253_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1254_cast_fp16 = sigmoid(x = var_1253_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_37_cast_fp16 = mul(x = x_12_cast_fp16, y = var_1254_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<int32, [2]> var_1258 = const()[name = string("op_1258"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1260 = const()[name = string("op_1260"), val = tensor<int32, [2]>([1, 1])];
            string input0_23_pad_type_0 = const()[name = string("input0_23_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_23_pad_0 = const()[name = string("input0_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43867072))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43865984))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44915712)))];
            tensor<fp16, [1, 512, 1, 256]> input0_23_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16, dilations = var_1260, groups = var_46, pad = input0_23_pad_0, pad_type = input0_23_pad_type_0, strides = var_1258, weight = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_37_cast_fp16)[name = string("input0_23_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1264_cast_fp16 = add(x = input0_23_cast_fp16, y = inputs0_12_cast_fp16)[name = string("op_1264_cast_fp16")];
            fp16 var_1265_to_fp16 = const()[name = string("op_1265_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1266_cast_fp16 = mul(x = inputs3_1_cast_fp16, y = var_1265_to_fp16)[name = string("op_1266_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs4_1_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1264_cast_fp16)[name = string("inputs4_1_cast_fp16")];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_27_gamma_0_to_fp16 = const()[name = string("k_27_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44916800)))];
            tensor<fp16, [512]> k_27_beta_0_to_fp16 = const()[name = string("k_27_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44917888)))];
            fp16 var_1282_to_fp16 = const()[name = string("op_1282_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_27_cast_fp16 = layer_norm(axes = k_27_axes_0, beta = k_27_beta_0_to_fp16, epsilon = var_1282_to_fp16, gamma = k_27_gamma_0_to_fp16, x = inputs4_1_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [2]> var_1301 = const()[name = string("op_1301"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1303 = const()[name = string("op_1303"), val = tensor<int32, [2]>([1, 1])];
            string var_1305_pad_type_0 = const()[name = string("op_1305_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1305_pad_0 = const()[name = string("op_1305_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44920064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44918976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45182272)))];
            tensor<fp16, [1, 512, 1, 256]> var_1305_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16, dilations = var_1303, groups = var_46, pad = var_1305_pad_0, pad_type = var_1305_pad_type_0, strides = var_1301, weight = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1305_cast_fp16")];
            tensor<int32, [2]> var_1308 = const()[name = string("op_1308"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1310 = const()[name = string("op_1310"), val = tensor<int32, [2]>([1, 1])];
            string k_29_pad_type_0 = const()[name = string("k_29_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_29_pad_0 = const()[name = string("k_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45184448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45183360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45446656)))];
            tensor<fp16, [1, 512, 1, 256]> k_29_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16, dilations = var_1310, groups = var_46, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = var_1308, weight = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [2]> var_1315 = const()[name = string("op_1315"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1317 = const()[name = string("op_1317"), val = tensor<int32, [2]>([1, 1])];
            string var_1319_pad_type_0 = const()[name = string("op_1319_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1319_pad_0 = const()[name = string("op_1319_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45448832))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45447744))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45711040)))];
            tensor<fp16, [1, 512, 1, 256]> var_1319_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16, dilations = var_1317, groups = var_46, pad = var_1319_pad_0, pad_type = var_1319_pad_type_0, strides = var_1315, weight = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1319_cast_fp16")];
            tensor<int32, [8]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1320_axis_0 = const()[name = string("op_1320_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1320_cast_fp16_7 = split(axis = var_1320_axis_0, split_sizes = tile_24, x = var_1305_cast_fp16)[name = string("op_1320_cast_fp16")];
            tensor<int32, [4]> var_1329_perm_0 = const()[name = string("op_1329_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1330_axis_0 = const()[name = string("op_1330_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_2 = transpose(perm = var_1329_perm_0, x = k_29_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1330_cast_fp16_7 = split(axis = var_1330_axis_0, split_sizes = tile_25, x = transpose_2)[name = string("op_1330_cast_fp16")];
            tensor<int32, [8]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1339_axis_0 = const()[name = string("op_1339_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16_7 = split(axis = var_1339_axis_0, split_sizes = tile_26, x = var_1319_cast_fp16)[name = string("op_1339_cast_fp16")];
            string var_1349_equation_0 = const()[name = string("op_1349_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1349_cast_fp16 = einsum(equation = var_1349_equation_0, values = (var_1330_cast_fp16_0, var_1320_cast_fp16_0))[name = string("op_1349_cast_fp16")];
            fp16 var_1350_to_fp16 = const()[name = string("op_1350_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1351_cast_fp16 = mul(x = var_1349_cast_fp16, y = var_1350_to_fp16)[name = string("op_1351_cast_fp16")];
            string var_1353_equation_0 = const()[name = string("op_1353_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1353_cast_fp16 = einsum(equation = var_1353_equation_0, values = (var_1330_cast_fp16_1, var_1320_cast_fp16_1))[name = string("op_1353_cast_fp16")];
            fp16 var_1354_to_fp16 = const()[name = string("op_1354_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1355_cast_fp16 = mul(x = var_1353_cast_fp16, y = var_1354_to_fp16)[name = string("op_1355_cast_fp16")];
            string var_1357_equation_0 = const()[name = string("op_1357_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1357_cast_fp16 = einsum(equation = var_1357_equation_0, values = (var_1330_cast_fp16_2, var_1320_cast_fp16_2))[name = string("op_1357_cast_fp16")];
            fp16 var_1358_to_fp16 = const()[name = string("op_1358_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1359_cast_fp16 = mul(x = var_1357_cast_fp16, y = var_1358_to_fp16)[name = string("op_1359_cast_fp16")];
            string var_1361_equation_0 = const()[name = string("op_1361_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1361_cast_fp16 = einsum(equation = var_1361_equation_0, values = (var_1330_cast_fp16_3, var_1320_cast_fp16_3))[name = string("op_1361_cast_fp16")];
            fp16 var_1362_to_fp16 = const()[name = string("op_1362_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1363_cast_fp16 = mul(x = var_1361_cast_fp16, y = var_1362_to_fp16)[name = string("op_1363_cast_fp16")];
            string var_1365_equation_0 = const()[name = string("op_1365_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1365_cast_fp16 = einsum(equation = var_1365_equation_0, values = (var_1330_cast_fp16_4, var_1320_cast_fp16_4))[name = string("op_1365_cast_fp16")];
            fp16 var_1366_to_fp16 = const()[name = string("op_1366_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1367_cast_fp16 = mul(x = var_1365_cast_fp16, y = var_1366_to_fp16)[name = string("op_1367_cast_fp16")];
            string var_1369_equation_0 = const()[name = string("op_1369_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1369_cast_fp16 = einsum(equation = var_1369_equation_0, values = (var_1330_cast_fp16_5, var_1320_cast_fp16_5))[name = string("op_1369_cast_fp16")];
            fp16 var_1370_to_fp16 = const()[name = string("op_1370_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1371_cast_fp16 = mul(x = var_1369_cast_fp16, y = var_1370_to_fp16)[name = string("op_1371_cast_fp16")];
            string var_1373_equation_0 = const()[name = string("op_1373_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1373_cast_fp16 = einsum(equation = var_1373_equation_0, values = (var_1330_cast_fp16_6, var_1320_cast_fp16_6))[name = string("op_1373_cast_fp16")];
            fp16 var_1374_to_fp16 = const()[name = string("op_1374_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1375_cast_fp16 = mul(x = var_1373_cast_fp16, y = var_1374_to_fp16)[name = string("op_1375_cast_fp16")];
            string var_1377_equation_0 = const()[name = string("op_1377_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1377_cast_fp16 = einsum(equation = var_1377_equation_0, values = (var_1330_cast_fp16_7, var_1320_cast_fp16_7))[name = string("op_1377_cast_fp16")];
            fp16 var_1378_to_fp16 = const()[name = string("op_1378_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1379_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = string("op_1379_cast_fp16")];
            bool attn_weights_14_interleave_0 = const()[name = string("attn_weights_14_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_14_cast_fp16 = concat(axis = var_47, interleave = attn_weights_14_interleave_0, values = (var_1351_cast_fp16, var_1355_cast_fp16, var_1359_cast_fp16, var_1363_cast_fp16, var_1367_cast_fp16, var_1371_cast_fp16, var_1375_cast_fp16, var_1379_cast_fp16))[name = string("attn_weights_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_14_cast_fp16 = add(x = attn_weights_14_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_39_cast_fp16 = softmax(axis = var_46, x = attn_weights0_14_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [8]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1385_axis_0 = const()[name = string("op_1385_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1385_cast_fp16_7 = split(axis = var_1385_axis_0, split_sizes = tile_27, x = input_39_cast_fp16)[name = string("op_1385_cast_fp16")];
            string var_1395_equation_0 = const()[name = string("op_1395_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1395_cast_fp16 = einsum(equation = var_1395_equation_0, values = (var_1339_cast_fp16_0, var_1385_cast_fp16_0))[name = string("op_1395_cast_fp16")];
            string var_1397_equation_0 = const()[name = string("op_1397_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1397_cast_fp16 = einsum(equation = var_1397_equation_0, values = (var_1339_cast_fp16_1, var_1385_cast_fp16_1))[name = string("op_1397_cast_fp16")];
            string var_1399_equation_0 = const()[name = string("op_1399_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1399_cast_fp16 = einsum(equation = var_1399_equation_0, values = (var_1339_cast_fp16_2, var_1385_cast_fp16_2))[name = string("op_1399_cast_fp16")];
            string var_1401_equation_0 = const()[name = string("op_1401_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1401_cast_fp16 = einsum(equation = var_1401_equation_0, values = (var_1339_cast_fp16_3, var_1385_cast_fp16_3))[name = string("op_1401_cast_fp16")];
            string var_1403_equation_0 = const()[name = string("op_1403_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1403_cast_fp16 = einsum(equation = var_1403_equation_0, values = (var_1339_cast_fp16_4, var_1385_cast_fp16_4))[name = string("op_1403_cast_fp16")];
            string var_1405_equation_0 = const()[name = string("op_1405_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1405_cast_fp16 = einsum(equation = var_1405_equation_0, values = (var_1339_cast_fp16_5, var_1385_cast_fp16_5))[name = string("op_1405_cast_fp16")];
            string var_1407_equation_0 = const()[name = string("op_1407_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1407_cast_fp16 = einsum(equation = var_1407_equation_0, values = (var_1339_cast_fp16_6, var_1385_cast_fp16_6))[name = string("op_1407_cast_fp16")];
            string var_1409_equation_0 = const()[name = string("op_1409_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1409_cast_fp16 = einsum(equation = var_1409_equation_0, values = (var_1339_cast_fp16_7, var_1385_cast_fp16_7))[name = string("op_1409_cast_fp16")];
            bool attn_42_interleave_0 = const()[name = string("attn_42_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_42_cast_fp16 = concat(axis = var_46, interleave = attn_42_interleave_0, values = (var_1395_cast_fp16, var_1397_cast_fp16, var_1399_cast_fp16, var_1401_cast_fp16, var_1403_cast_fp16, var_1405_cast_fp16, var_1407_cast_fp16, var_1409_cast_fp16))[name = string("attn_42_cast_fp16")];
            tensor<int32, [2]> var_1417 = const()[name = string("op_1417"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1419 = const()[name = string("op_1419"), val = tensor<int32, [2]>([1, 1])];
            string attn_44_pad_type_0 = const()[name = string("attn_44_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_44_pad_0 = const()[name = string("attn_44_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45713216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45712128))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45975424)))];
            tensor<fp16, [1, 512, 1, 256]> attn_44_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16, dilations = var_1419, groups = var_46, pad = attn_44_pad_0, pad_type = attn_44_pad_type_0, strides = var_1417, weight = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_42_cast_fp16)[name = string("attn_44_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_14_cast_fp16 = add(x = inputs4_1_cast_fp16, y = attn_44_cast_fp16)[name = string("inputs0_14_cast_fp16")];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_41_gamma_0_to_fp16 = const()[name = string("input_41_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45976512)))];
            tensor<fp16, [512]> input_41_beta_0_to_fp16 = const()[name = string("input_41_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45977600)))];
            fp16 var_1431_to_fp16 = const()[name = string("op_1431_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = input_41_beta_0_to_fp16, epsilon = var_1431_to_fp16, gamma = input_41_gamma_0_to_fp16, x = inputs0_14_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<int32, [2]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [2]>([1, 1])];
            string x_14_pad_type_0 = const()[name = string("x_14_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_14_pad_0 = const()[name = string("x_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45982848))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45978688))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47035648))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47031488))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_14_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1447, groups = var_46, pad = x_14_pad_0, pad_type = x_14_pad_type_0, strides = var_1445, weight = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_41_cast_fp16)[name = string("x_14_cast_fp16")];
            fp16 var_1450_to_fp16 = const()[name = string("op_1450_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1451_cast_fp16 = mul(x = x_14_cast_fp16, y = var_1450_to_fp16)[name = string("op_1451_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1452_cast_fp16 = sigmoid(x = var_1451_cast_fp16)[name = string("op_1452_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_43_cast_fp16 = mul(x = x_14_cast_fp16, y = var_1452_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<int32, [2]> var_1456 = const()[name = string("op_1456"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1458 = const()[name = string("op_1458"), val = tensor<int32, [2]>([1, 1])];
            string input0_27_pad_type_0 = const()[name = string("input0_27_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_27_pad_0 = const()[name = string("input0_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47038848))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47037760))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48087488)))];
            tensor<fp16, [1, 512, 1, 256]> input0_27_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16, dilations = var_1458, groups = var_46, pad = input0_27_pad_0, pad_type = input0_27_pad_type_0, strides = var_1456, weight = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_43_cast_fp16)[name = string("input0_27_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1462_cast_fp16 = add(x = input0_27_cast_fp16, y = inputs0_14_cast_fp16)[name = string("op_1462_cast_fp16")];
            fp16 var_1463_to_fp16 = const()[name = string("op_1463_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1464_cast_fp16 = mul(x = inputs4_1_cast_fp16, y = var_1463_to_fp16)[name = string("op_1464_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs5_1_cast_fp16 = add(x = var_1464_cast_fp16, y = var_1462_cast_fp16)[name = string("inputs5_1_cast_fp16")];
            tensor<int32, [1]> k_2_axes_0 = const()[name = string("k_2_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_2_gamma_0_to_fp16 = const()[name = string("k_2_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48088576)))];
            tensor<fp16, [512]> k_2_beta_0_to_fp16 = const()[name = string("k_2_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48089664)))];
            fp16 var_1480_to_fp16 = const()[name = string("op_1480_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_2_cast_fp16 = layer_norm(axes = k_2_axes_0, beta = k_2_beta_0_to_fp16, epsilon = var_1480_to_fp16, gamma = k_2_gamma_0_to_fp16, x = inputs5_1_cast_fp16)[name = string("k_2_cast_fp16")];
            tensor<int32, [2]> var_1499 = const()[name = string("op_1499"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1501 = const()[name = string("op_1501"), val = tensor<int32, [2]>([1, 1])];
            string var_1503_pad_type_0 = const()[name = string("op_1503_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1503_pad_0 = const()[name = string("op_1503_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48091840))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48090752))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48354048)))];
            tensor<fp16, [1, 512, 1, 256]> var_1503_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16, dilations = var_1501, groups = var_46, pad = var_1503_pad_0, pad_type = var_1503_pad_type_0, strides = var_1499, weight = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_1503_cast_fp16")];
            tensor<int32, [2]> var_1506 = const()[name = string("op_1506"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1508 = const()[name = string("op_1508"), val = tensor<int32, [2]>([1, 1])];
            string k_1_pad_type_0 = const()[name = string("k_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_1_pad_0 = const()[name = string("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48356224))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48355136))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48618432)))];
            tensor<fp16, [1, 512, 1, 256]> k_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16, dilations = var_1508, groups = var_46, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = var_1506, weight = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [2]> var_1513 = const()[name = string("op_1513"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [2]>([1, 1])];
            string var_1517_pad_type_0 = const()[name = string("op_1517_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1517_pad_0 = const()[name = string("op_1517_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48620608))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48619520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48882816)))];
            tensor<fp16, [1, 512, 1, 256]> var_1517_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16, dilations = var_1515, groups = var_46, pad = var_1517_pad_0, pad_type = var_1517_pad_type_0, strides = var_1513, weight = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<int32, [8]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1518_axis_0 = const()[name = string("op_1518_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1518_cast_fp16_7 = split(axis = var_1518_axis_0, split_sizes = tile_28, x = var_1503_cast_fp16)[name = string("op_1518_cast_fp16")];
            tensor<int32, [4]> var_1527_perm_0 = const()[name = string("op_1527_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1528_axis_0 = const()[name = string("op_1528_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_1 = transpose(perm = var_1527_perm_0, x = k_1_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1528_cast_fp16_7 = split(axis = var_1528_axis_0, split_sizes = tile_29, x = transpose_1)[name = string("op_1528_cast_fp16")];
            tensor<int32, [8]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1537_axis_0 = const()[name = string("op_1537_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1537_cast_fp16_7 = split(axis = var_1537_axis_0, split_sizes = tile_30, x = var_1517_cast_fp16)[name = string("op_1537_cast_fp16")];
            string var_1547_equation_0 = const()[name = string("op_1547_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1547_cast_fp16 = einsum(equation = var_1547_equation_0, values = (var_1528_cast_fp16_0, var_1518_cast_fp16_0))[name = string("op_1547_cast_fp16")];
            fp16 var_1548_to_fp16 = const()[name = string("op_1548_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1549_cast_fp16 = mul(x = var_1547_cast_fp16, y = var_1548_to_fp16)[name = string("op_1549_cast_fp16")];
            string var_1551_equation_0 = const()[name = string("op_1551_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1551_cast_fp16 = einsum(equation = var_1551_equation_0, values = (var_1528_cast_fp16_1, var_1518_cast_fp16_1))[name = string("op_1551_cast_fp16")];
            fp16 var_1552_to_fp16 = const()[name = string("op_1552_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1553_cast_fp16 = mul(x = var_1551_cast_fp16, y = var_1552_to_fp16)[name = string("op_1553_cast_fp16")];
            string var_1555_equation_0 = const()[name = string("op_1555_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1555_cast_fp16 = einsum(equation = var_1555_equation_0, values = (var_1528_cast_fp16_2, var_1518_cast_fp16_2))[name = string("op_1555_cast_fp16")];
            fp16 var_1556_to_fp16 = const()[name = string("op_1556_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1557_cast_fp16 = mul(x = var_1555_cast_fp16, y = var_1556_to_fp16)[name = string("op_1557_cast_fp16")];
            string var_1559_equation_0 = const()[name = string("op_1559_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1559_cast_fp16 = einsum(equation = var_1559_equation_0, values = (var_1528_cast_fp16_3, var_1518_cast_fp16_3))[name = string("op_1559_cast_fp16")];
            fp16 var_1560_to_fp16 = const()[name = string("op_1560_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1561_cast_fp16 = mul(x = var_1559_cast_fp16, y = var_1560_to_fp16)[name = string("op_1561_cast_fp16")];
            string var_1563_equation_0 = const()[name = string("op_1563_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1563_cast_fp16 = einsum(equation = var_1563_equation_0, values = (var_1528_cast_fp16_4, var_1518_cast_fp16_4))[name = string("op_1563_cast_fp16")];
            fp16 var_1564_to_fp16 = const()[name = string("op_1564_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1565_cast_fp16 = mul(x = var_1563_cast_fp16, y = var_1564_to_fp16)[name = string("op_1565_cast_fp16")];
            string var_1567_equation_0 = const()[name = string("op_1567_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1567_cast_fp16 = einsum(equation = var_1567_equation_0, values = (var_1528_cast_fp16_5, var_1518_cast_fp16_5))[name = string("op_1567_cast_fp16")];
            fp16 var_1568_to_fp16 = const()[name = string("op_1568_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1569_cast_fp16 = mul(x = var_1567_cast_fp16, y = var_1568_to_fp16)[name = string("op_1569_cast_fp16")];
            string var_1571_equation_0 = const()[name = string("op_1571_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1571_cast_fp16 = einsum(equation = var_1571_equation_0, values = (var_1528_cast_fp16_6, var_1518_cast_fp16_6))[name = string("op_1571_cast_fp16")];
            fp16 var_1572_to_fp16 = const()[name = string("op_1572_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1573_cast_fp16 = mul(x = var_1571_cast_fp16, y = var_1572_to_fp16)[name = string("op_1573_cast_fp16")];
            string var_1575_equation_0 = const()[name = string("op_1575_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1575_cast_fp16 = einsum(equation = var_1575_equation_0, values = (var_1528_cast_fp16_7, var_1518_cast_fp16_7))[name = string("op_1575_cast_fp16")];
            fp16 var_1576_to_fp16 = const()[name = string("op_1576_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1577_cast_fp16 = mul(x = var_1575_cast_fp16, y = var_1576_to_fp16)[name = string("op_1577_cast_fp16")];
            bool attn_weights_1_interleave_0 = const()[name = string("attn_weights_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_1_cast_fp16 = concat(axis = var_47, interleave = attn_weights_1_interleave_0, values = (var_1549_cast_fp16, var_1553_cast_fp16, var_1557_cast_fp16, var_1561_cast_fp16, var_1565_cast_fp16, var_1569_cast_fp16, var_1573_cast_fp16, var_1577_cast_fp16))[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_1_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_2_cast_fp16 = softmax(axis = var_46, x = attn_weights0_1_cast_fp16)[name = string("input_2_cast_fp16")];
            tensor<int32, [8]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1583_axis_0 = const()[name = string("op_1583_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16_7 = split(axis = var_1583_axis_0, split_sizes = tile_31, x = input_2_cast_fp16)[name = string("op_1583_cast_fp16")];
            string var_1593_equation_0 = const()[name = string("op_1593_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1593_cast_fp16 = einsum(equation = var_1593_equation_0, values = (var_1537_cast_fp16_0, var_1583_cast_fp16_0))[name = string("op_1593_cast_fp16")];
            string var_1595_equation_0 = const()[name = string("op_1595_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1595_cast_fp16 = einsum(equation = var_1595_equation_0, values = (var_1537_cast_fp16_1, var_1583_cast_fp16_1))[name = string("op_1595_cast_fp16")];
            string var_1597_equation_0 = const()[name = string("op_1597_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1597_cast_fp16 = einsum(equation = var_1597_equation_0, values = (var_1537_cast_fp16_2, var_1583_cast_fp16_2))[name = string("op_1597_cast_fp16")];
            string var_1599_equation_0 = const()[name = string("op_1599_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1599_cast_fp16 = einsum(equation = var_1599_equation_0, values = (var_1537_cast_fp16_3, var_1583_cast_fp16_3))[name = string("op_1599_cast_fp16")];
            string var_1601_equation_0 = const()[name = string("op_1601_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1601_cast_fp16 = einsum(equation = var_1601_equation_0, values = (var_1537_cast_fp16_4, var_1583_cast_fp16_4))[name = string("op_1601_cast_fp16")];
            string var_1603_equation_0 = const()[name = string("op_1603_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1603_cast_fp16 = einsum(equation = var_1603_equation_0, values = (var_1537_cast_fp16_5, var_1583_cast_fp16_5))[name = string("op_1603_cast_fp16")];
            string var_1605_equation_0 = const()[name = string("op_1605_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1605_cast_fp16 = einsum(equation = var_1605_equation_0, values = (var_1537_cast_fp16_6, var_1583_cast_fp16_6))[name = string("op_1605_cast_fp16")];
            string var_1607_equation_0 = const()[name = string("op_1607_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1607_cast_fp16 = einsum(equation = var_1607_equation_0, values = (var_1537_cast_fp16_7, var_1583_cast_fp16_7))[name = string("op_1607_cast_fp16")];
            bool attn_2_interleave_0 = const()[name = string("attn_2_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_2_cast_fp16 = concat(axis = var_46, interleave = attn_2_interleave_0, values = (var_1593_cast_fp16, var_1595_cast_fp16, var_1597_cast_fp16, var_1599_cast_fp16, var_1601_cast_fp16, var_1603_cast_fp16, var_1605_cast_fp16, var_1607_cast_fp16))[name = string("attn_2_cast_fp16")];
            tensor<int32, [2]> var_1615 = const()[name = string("op_1615"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1617 = const()[name = string("op_1617"), val = tensor<int32, [2]>([1, 1])];
            string attn_1_pad_type_0 = const()[name = string("attn_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> attn_1_pad_0 = const()[name = string("attn_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48884992))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48883904))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49147200)))];
            tensor<fp16, [1, 512, 1, 256]> attn_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16, dilations = var_1617, groups = var_46, pad = attn_1_pad_0, pad_type = attn_1_pad_type_0, strides = var_1615, weight = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_2_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_1_cast_fp16 = add(x = inputs5_1_cast_fp16, y = attn_1_cast_fp16)[name = string("inputs0_1_cast_fp16")];
            tensor<int32, [1]> input_4_axes_0 = const()[name = string("input_4_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_4_gamma_0_to_fp16 = const()[name = string("input_4_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49148288)))];
            tensor<fp16, [512]> input_4_beta_0_to_fp16 = const()[name = string("input_4_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49149376)))];
            fp16 var_1629_to_fp16 = const()[name = string("op_1629_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_4_cast_fp16 = layer_norm(axes = input_4_axes_0, beta = input_4_beta_0_to_fp16, epsilon = var_1629_to_fp16, gamma = input_4_gamma_0_to_fp16, x = inputs0_1_cast_fp16)[name = string("input_4_cast_fp16")];
            tensor<int32, [2]> var_1643 = const()[name = string("op_1643"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1645 = const()[name = string("op_1645"), val = tensor<int32, [2]>([1, 1])];
            string x_1_pad_type_0 = const()[name = string("x_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_1_pad_0 = const()[name = string("x_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49154624))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49150464))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50207424))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50203264))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1645, groups = var_46, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = var_1643, weight = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_4_cast_fp16)[name = string("x_1_cast_fp16")];
            fp16 var_1648_to_fp16 = const()[name = string("op_1648_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1649_cast_fp16 = mul(x = x_1_cast_fp16, y = var_1648_to_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1650_cast_fp16 = sigmoid(x = var_1649_cast_fp16)[name = string("op_1650_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_1_cast_fp16 = mul(x = x_1_cast_fp16, y = var_1650_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [2]> var_1654 = const()[name = string("op_1654"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1656 = const()[name = string("op_1656"), val = tensor<int32, [2]>([1, 1])];
            string input0_1_pad_type_0 = const()[name = string("input0_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_1_pad_0 = const()[name = string("input0_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50210624))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50209536))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51259264)))];
            tensor<fp16, [1, 512, 1, 256]> input0_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16, dilations = var_1656, groups = var_46, pad = input0_1_pad_0, pad_type = input0_1_pad_type_0, strides = var_1654, weight = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_1_cast_fp16)[name = string("input0_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> var_1660_cast_fp16 = add(x = input0_1_cast_fp16, y = inputs0_1_cast_fp16)[name = string("op_1660_cast_fp16")];
            fp16 var_1661_to_fp16 = const()[name = string("op_1661_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1662_cast_fp16 = mul(x = inputs5_1_cast_fp16, y = var_1661_to_fp16)[name = string("op_1662_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs6_1_cast_fp16 = add(x = var_1662_cast_fp16, y = var_1660_cast_fp16)[name = string("inputs6_1_cast_fp16")];
            tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> embeddings_1_gamma_0_to_fp16 = const()[name = string("embeddings_1_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51260352)))];
            tensor<fp16, [512]> embeddings_1_beta_0_to_fp16 = const()[name = string("embeddings_1_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51261440)))];
            fp16 var_1674_to_fp16 = const()[name = string("op_1674_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = embeddings_1_beta_0_to_fp16, epsilon = var_1674_to_fp16, gamma = embeddings_1_gamma_0_to_fp16, x = inputs6_1_cast_fp16)[name = string("embeddings_1_cast_fp16")];
            tensor<int32, [4]> mlm_embeddings_1_perm_0 = const()[name = string("mlm_embeddings_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string mlm_embeddings_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("mlm_embeddings_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp16, [1, 256, 1, 512]> transpose_0 = transpose(perm = mlm_embeddings_1_perm_0, x = embeddings_1_cast_fp16)[name = string("transpose_0")];
            tensor<fp32, [1, 256, 1, 512]> mlm_embeddings = cast(dtype = mlm_embeddings_1_cast_fp16_to_fp32_dtype_0, x = transpose_0)[name = string("cast_0")];
        } -> (mlm_embeddings);
    func sentence_embed<ios16>(tensor<int32, [1, 256, 1, 3]> mlm_input) {
            int32 var_6 = const()[name = string("op_6"), val = int32(0)];
            tensor<int32, [3]> var_13 = const()[name = string("op_13"), val = tensor<int32, [3]>([1, 1, 1])];
            int32 var_14_axis_0 = const()[name = string("op_14_axis_0"), val = int32(-1)];
            tensor<int32, [1, 256, 1, 1]> var_14_0, tensor<int32, [1, 256, 1, 1]> var_14_1, tensor<int32, [1, 256, 1, 1]> var_14_2 = split(axis = var_14_axis_0, split_sizes = var_13, x = mlm_input)[name = string("op_14")];
            tensor<int32, [1]> var_18_axes_0 = const()[name = string("op_18_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_18 = squeeze(axes = var_18_axes_0, x = var_14_0)[name = string("op_18")];
            tensor<int32, [1]> tok_ids_1_axes_0 = const()[name = string("tok_ids_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256]> tok_ids_1 = squeeze(axes = tok_ids_1_axes_0, x = var_18)[name = string("tok_ids_1")];
            tensor<int32, [1]> var_20_axes_0 = const()[name = string("op_20_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_20 = squeeze(axes = var_20_axes_0, x = var_14_1)[name = string("op_20")];
            tensor<int32, [1]> var_22_axes_0 = const()[name = string("op_22_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [1, 256, 1]> var_22 = squeeze(axes = var_22_axes_0, x = var_14_2)[name = string("op_22")];
            tensor<bool, [1, 256]> var_24 = not_equal(x = tok_ids_1, y = var_6)[name = string("op_24")];
            fp16 var_8_to_fp16 = const()[name = string("op_8_to_fp16"), val = fp16(1)];
            string var_24_to_fp32_to_fp16_dtype_0 = const()[name = string("op_24_to_fp32_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 256]> cast_1 = cast(dtype = var_24_to_fp32_to_fp16_dtype_0, x = var_24)[name = string("cast_1")];
            tensor<fp16, [1, 256]> var_29_cast_fp16 = sub(x = var_8_to_fp16, y = cast_1)[name = string("op_29_cast_fp16")];
            fp16 var_30_to_fp16 = const()[name = string("op_30_to_fp16"), val = fp16(-10000)];
            tensor<fp16, [1, 256]> padding_mask0_1_cast_fp16 = mul(x = var_29_cast_fp16, y = var_30_to_fp16)[name = string("padding_mask0_1_cast_fp16")];
            tensor<int32, [4]> var_32 = const()[name = string("op_32"), val = tensor<int32, [4]>([-1, 256, 1, 1])];
            tensor<fp16, [1, 256, 1, 1]> var_33_cast_fp16 = reshape(shape = var_32, x = padding_mask0_1_cast_fp16)[name = string("op_33_cast_fp16")];
            int32 var_46 = const()[name = string("op_46"), val = int32(1)];
            int32 var_47 = const()[name = string("op_47"), val = int32(2)];
            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<int32, [1, 256, 1]> input_9 = expand_dims(axes = input_9_axes_0, x = tok_ids_1)[name = string("input_9")];
            int32 var_54_axis_0 = const()[name = string("op_54_axis_0"), val = int32(0)];
            tensor<fp16, [50000, 512]> nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [50000, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(150272))), scale = tensor<fp16, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50176))), zero_point = tensor<int8, [50000]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(64)))];
            int32 op_54_cast_fp16_batch_dims_0 = const()[name = string("op_54_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_54_cast_fp16 = gather(axis = var_54_axis_0, batch_dims = op_54_cast_fp16_batch_dims_0, indices = input_9, x = nlp_net_default_encoder_tok_embed_weight_to_fp16_affine_quantized)[name = string("op_54_cast_fp16")];
            int32 var_57_axis_0 = const()[name = string("op_57_axis_0"), val = int32(0)];
            tensor<fp16, [256, 512]> nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [256, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25751232))), scale = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750656))), zero_point = tensor<int8, [256]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25750336)))];
            int32 op_57_cast_fp16_batch_dims_0 = const()[name = string("op_57_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_57_cast_fp16 = gather(axis = var_57_axis_0, batch_dims = op_57_cast_fp16_batch_dims_0, indices = var_20, x = nlp_net_default_encoder_pos_embed_weight_to_fp16_affine_quantized)[name = string("op_57_cast_fp16")];
            int32 var_60_axis_0 = const()[name = string("op_60_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512]> nlp_net_default_encoder_seg_embed_weight_to_fp16 = const()[name = string("nlp_net_default_encoder_seg_embed_weight_to_fp16"), val = tensor<fp16, [1, 512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25882368)))];
            int32 op_60_cast_fp16_batch_dims_0 = const()[name = string("op_60_cast_fp16_batch_dims_0"), val = int32(0)];
            tensor<fp16, [1, 256, 1, 512]> var_60_cast_fp16 = gather(axis = var_60_axis_0, batch_dims = op_60_cast_fp16_batch_dims_0, indices = var_22, x = nlp_net_default_encoder_seg_embed_weight_to_fp16)[name = string("op_60_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_62_cast_fp16 = add(x = var_54_cast_fp16, y = var_57_cast_fp16)[name = string("op_62_cast_fp16")];
            tensor<fp16, [1, 256, 1, 512]> var_63_cast_fp16 = add(x = var_62_cast_fp16, y = var_60_cast_fp16)[name = string("op_63_cast_fp16")];
            tensor<int32, [4]> t_1_perm_0 = const()[name = string("t_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [1]> k_3_axes_0 = const()[name = string("k_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_3_gamma_0_to_fp16 = const()[name = string("k_3_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25883456)))];
            tensor<fp16, [512]> k_3_beta_0_to_fp16 = const()[name = string("k_3_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25884544)))];
            fp16 var_96_to_fp16 = const()[name = string("op_96_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> transpose_9 = transpose(perm = t_1_perm_0, x = var_63_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 512, 1, 256]> k_3_cast_fp16 = layer_norm(axes = k_3_axes_0, beta = k_3_beta_0_to_fp16, epsilon = var_96_to_fp16, gamma = k_3_gamma_0_to_fp16, x = transpose_9)[name = string("k_3_cast_fp16")];
            tensor<int32, [2]> var_115 = const()[name = string("op_115"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_117 = const()[name = string("op_117"), val = tensor<int32, [2]>([1, 1])];
            string var_119_pad_type_0 = const()[name = string("op_119_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_119_pad_0 = const()[name = string("op_119_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25887296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25886208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26149504)))];
            tensor<fp16, [1, 512, 1, 256]> var_119_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_bias_to_fp16, dilations = var_117, groups = var_46, pad = var_119_pad_0, pad_type = var_119_pad_type_0, strides = var_115, weight = nlp_net_default_encoder_transformer_layers_0_attn_q_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_119_cast_fp16")];
            tensor<int32, [2]> var_122 = const()[name = string("op_122"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_124 = const()[name = string("op_124"), val = tensor<int32, [2]>([1, 1])];
            string k_5_pad_type_0 = const()[name = string("k_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_5_pad_0 = const()[name = string("k_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26151680))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26150592))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26413888)))];
            tensor<fp16, [1, 512, 1, 256]> k_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_bias_to_fp16, dilations = var_124, groups = var_46, pad = k_5_pad_0, pad_type = k_5_pad_type_0, strides = var_122, weight = nlp_net_default_encoder_transformer_layers_0_attn_k_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<int32, [2]> var_129 = const()[name = string("op_129"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_131 = const()[name = string("op_131"), val = tensor<int32, [2]>([1, 1])];
            string var_133_pad_type_0 = const()[name = string("op_133_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_133_pad_0 = const()[name = string("op_133_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26416064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26414976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26678272)))];
            tensor<fp16, [1, 512, 1, 256]> var_133_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_bias_to_fp16, dilations = var_131, groups = var_46, pad = var_133_pad_0, pad_type = var_133_pad_type_0, strides = var_129, weight = nlp_net_default_encoder_transformer_layers_0_attn_v_proj_weight_to_fp16_affine_quantized, x = k_3_cast_fp16)[name = string("op_133_cast_fp16")];
            tensor<int32, [8]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_134_axis_0 = const()[name = string("op_134_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_134_cast_fp16_7 = split(axis = var_134_axis_0, split_sizes = tile_0, x = var_119_cast_fp16)[name = string("op_134_cast_fp16")];
            tensor<int32, [4]> var_143_perm_0 = const()[name = string("op_143_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_144_axis_0 = const()[name = string("op_144_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_8 = transpose(perm = var_143_perm_0, x = k_5_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_144_cast_fp16_7 = split(axis = var_144_axis_0, split_sizes = tile_1, x = transpose_8)[name = string("op_144_cast_fp16")];
            tensor<int32, [8]> tile_2 = const()[name = string("tile_2"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_153_axis_0 = const()[name = string("op_153_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_153_cast_fp16_7 = split(axis = var_153_axis_0, split_sizes = tile_2, x = var_133_cast_fp16)[name = string("op_153_cast_fp16")];
            string var_163_equation_0 = const()[name = string("op_163_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_163_cast_fp16 = einsum(equation = var_163_equation_0, values = (var_144_cast_fp16_0, var_134_cast_fp16_0))[name = string("op_163_cast_fp16")];
            fp16 var_164_to_fp16 = const()[name = string("op_164_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_165_cast_fp16 = mul(x = var_163_cast_fp16, y = var_164_to_fp16)[name = string("op_165_cast_fp16")];
            string var_167_equation_0 = const()[name = string("op_167_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_167_cast_fp16 = einsum(equation = var_167_equation_0, values = (var_144_cast_fp16_1, var_134_cast_fp16_1))[name = string("op_167_cast_fp16")];
            fp16 var_168_to_fp16 = const()[name = string("op_168_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_169_cast_fp16 = mul(x = var_167_cast_fp16, y = var_168_to_fp16)[name = string("op_169_cast_fp16")];
            string var_171_equation_0 = const()[name = string("op_171_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_171_cast_fp16 = einsum(equation = var_171_equation_0, values = (var_144_cast_fp16_2, var_134_cast_fp16_2))[name = string("op_171_cast_fp16")];
            fp16 var_172_to_fp16 = const()[name = string("op_172_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_173_cast_fp16 = mul(x = var_171_cast_fp16, y = var_172_to_fp16)[name = string("op_173_cast_fp16")];
            string var_175_equation_0 = const()[name = string("op_175_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_175_cast_fp16 = einsum(equation = var_175_equation_0, values = (var_144_cast_fp16_3, var_134_cast_fp16_3))[name = string("op_175_cast_fp16")];
            fp16 var_176_to_fp16 = const()[name = string("op_176_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_177_cast_fp16 = mul(x = var_175_cast_fp16, y = var_176_to_fp16)[name = string("op_177_cast_fp16")];
            string var_179_equation_0 = const()[name = string("op_179_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_179_cast_fp16 = einsum(equation = var_179_equation_0, values = (var_144_cast_fp16_4, var_134_cast_fp16_4))[name = string("op_179_cast_fp16")];
            fp16 var_180_to_fp16 = const()[name = string("op_180_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_181_cast_fp16 = mul(x = var_179_cast_fp16, y = var_180_to_fp16)[name = string("op_181_cast_fp16")];
            string var_183_equation_0 = const()[name = string("op_183_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_183_cast_fp16 = einsum(equation = var_183_equation_0, values = (var_144_cast_fp16_5, var_134_cast_fp16_5))[name = string("op_183_cast_fp16")];
            fp16 var_184_to_fp16 = const()[name = string("op_184_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_185_cast_fp16 = mul(x = var_183_cast_fp16, y = var_184_to_fp16)[name = string("op_185_cast_fp16")];
            string var_187_equation_0 = const()[name = string("op_187_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_187_cast_fp16 = einsum(equation = var_187_equation_0, values = (var_144_cast_fp16_6, var_134_cast_fp16_6))[name = string("op_187_cast_fp16")];
            fp16 var_188_to_fp16 = const()[name = string("op_188_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_189_cast_fp16 = mul(x = var_187_cast_fp16, y = var_188_to_fp16)[name = string("op_189_cast_fp16")];
            string var_191_equation_0 = const()[name = string("op_191_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_191_cast_fp16 = einsum(equation = var_191_equation_0, values = (var_144_cast_fp16_7, var_134_cast_fp16_7))[name = string("op_191_cast_fp16")];
            fp16 var_192_to_fp16 = const()[name = string("op_192_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_193_cast_fp16 = mul(x = var_191_cast_fp16, y = var_192_to_fp16)[name = string("op_193_cast_fp16")];
            bool attn_weights_2_interleave_0 = const()[name = string("attn_weights_2_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_2_cast_fp16 = concat(axis = var_47, interleave = attn_weights_2_interleave_0, values = (var_165_cast_fp16, var_169_cast_fp16, var_173_cast_fp16, var_177_cast_fp16, var_181_cast_fp16, var_185_cast_fp16, var_189_cast_fp16, var_193_cast_fp16))[name = string("attn_weights_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_2_cast_fp16 = add(x = attn_weights_2_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_2_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_5_cast_fp16 = softmax(axis = var_46, x = attn_weights0_2_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<int32, [8]> tile_3 = const()[name = string("tile_3"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_199_axis_0 = const()[name = string("op_199_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_199_cast_fp16_7 = split(axis = var_199_axis_0, split_sizes = tile_3, x = input_5_cast_fp16)[name = string("op_199_cast_fp16")];
            string var_209_equation_0 = const()[name = string("op_209_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_209_cast_fp16 = einsum(equation = var_209_equation_0, values = (var_153_cast_fp16_0, var_199_cast_fp16_0))[name = string("op_209_cast_fp16")];
            string var_211_equation_0 = const()[name = string("op_211_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_211_cast_fp16 = einsum(equation = var_211_equation_0, values = (var_153_cast_fp16_1, var_199_cast_fp16_1))[name = string("op_211_cast_fp16")];
            string var_213_equation_0 = const()[name = string("op_213_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_213_cast_fp16 = einsum(equation = var_213_equation_0, values = (var_153_cast_fp16_2, var_199_cast_fp16_2))[name = string("op_213_cast_fp16")];
            string var_215_equation_0 = const()[name = string("op_215_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_215_cast_fp16 = einsum(equation = var_215_equation_0, values = (var_153_cast_fp16_3, var_199_cast_fp16_3))[name = string("op_215_cast_fp16")];
            string var_217_equation_0 = const()[name = string("op_217_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_217_cast_fp16 = einsum(equation = var_217_equation_0, values = (var_153_cast_fp16_4, var_199_cast_fp16_4))[name = string("op_217_cast_fp16")];
            string var_219_equation_0 = const()[name = string("op_219_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_219_cast_fp16 = einsum(equation = var_219_equation_0, values = (var_153_cast_fp16_5, var_199_cast_fp16_5))[name = string("op_219_cast_fp16")];
            string var_221_equation_0 = const()[name = string("op_221_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_221_cast_fp16 = einsum(equation = var_221_equation_0, values = (var_153_cast_fp16_6, var_199_cast_fp16_6))[name = string("op_221_cast_fp16")];
            string var_223_equation_0 = const()[name = string("op_223_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_223_cast_fp16 = einsum(equation = var_223_equation_0, values = (var_153_cast_fp16_7, var_199_cast_fp16_7))[name = string("op_223_cast_fp16")];
            bool attn_5_interleave_0 = const()[name = string("attn_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_5_cast_fp16 = concat(axis = var_46, interleave = attn_5_interleave_0, values = (var_209_cast_fp16, var_211_cast_fp16, var_213_cast_fp16, var_215_cast_fp16, var_217_cast_fp16, var_219_cast_fp16, var_221_cast_fp16, var_223_cast_fp16))[name = string("attn_5_cast_fp16")];
            tensor<int32, [2]> var_231 = const()[name = string("op_231"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_233 = const()[name = string("op_233"), val = tensor<int32, [2]>([1, 1])];
            string inputs_5_pad_type_0 = const()[name = string("inputs_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_5_pad_0 = const()[name = string("inputs_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26680448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26679360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26942656)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_bias_to_fp16, dilations = var_233, groups = var_46, pad = inputs_5_pad_0, pad_type = inputs_5_pad_type_0, strides = var_231, weight = nlp_net_default_encoder_transformer_layers_0_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_5_cast_fp16)[name = string("inputs_5_cast_fp16")];
            tensor<int32, [1]> input_7_axes_0 = const()[name = string("input_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_7_gamma_0_to_fp16 = const()[name = string("input_7_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51262528)))];
            tensor<fp16, [512]> input_7_beta_0_to_fp16 = const()[name = string("input_7_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51263616)))];
            fp16 var_247_to_fp16 = const()[name = string("op_247_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = input_7_beta_0_to_fp16, epsilon = var_247_to_fp16, gamma = input_7_gamma_0_to_fp16, x = inputs_5_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<int32, [2]> var_259 = const()[name = string("op_259"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_261 = const()[name = string("op_261"), val = tensor<int32, [2]>([1, 1])];
            string x_6_pad_type_0 = const()[name = string("x_6_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_6_pad_0 = const()[name = string("x_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51265216))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264896))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51330816)))];
            tensor<fp16, [1, 128, 1, 256]> x_6_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_bias_to_fp16, dilations = var_261, groups = var_46, pad = x_6_pad_0, pad_type = x_6_pad_type_0, strides = var_259, weight = nlp_net_default_encoder_transformer_layers_0_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_7_cast_fp16)[name = string("x_6_cast_fp16")];
            fp16 var_264_to_fp16 = const()[name = string("op_264_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_265_cast_fp16 = mul(x = x_6_cast_fp16, y = var_264_to_fp16)[name = string("op_265_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_266_cast_fp16 = sigmoid(x = var_265_cast_fp16)[name = string("op_266_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_13_cast_fp16 = mul(x = x_6_cast_fp16, y = var_266_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<int32, [2]> var_270 = const()[name = string("op_270"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_272 = const()[name = string("op_272"), val = tensor<int32, [2]>([1, 1])];
            string x_8_pad_type_0 = const()[name = string("x_8_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_8_pad_0 = const()[name = string("x_8_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51332224))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51331136))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51397824)))];
            tensor<fp16, [1, 512, 1, 256]> x_8_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_bias_to_fp16, dilations = var_272, groups = var_46, pad = x_8_pad_0, pad_type = x_8_pad_type_0, strides = var_270, weight = nlp_net_default_encoder_transformer_layers_0_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_13_cast_fp16)[name = string("x_8_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_7_cast_fp16 = add(x = x_8_cast_fp16, y = inputs_5_cast_fp16)[name = string("attn_7_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_2_cast_fp16 = add(x = transpose_9, y = attn_7_cast_fp16)[name = string("inputs_2_cast_fp16")];
            tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_15_gamma_0_to_fp16 = const()[name = string("input_15_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26943744)))];
            tensor<fp16, [512]> input_15_beta_0_to_fp16 = const()[name = string("input_15_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26944832)))];
            fp16 var_285_to_fp16 = const()[name = string("op_285_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = input_15_beta_0_to_fp16, epsilon = var_285_to_fp16, gamma = input_15_gamma_0_to_fp16, x = inputs_2_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<int32, [2]> var_299 = const()[name = string("op_299"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_301 = const()[name = string("op_301"), val = tensor<int32, [2]>([1, 1])];
            string x_10_pad_type_0 = const()[name = string("x_10_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_10_pad_0 = const()[name = string("x_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26952192))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26948032))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28004992))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28000832))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_10_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_301, groups = var_46, pad = x_10_pad_0, pad_type = x_10_pad_type_0, strides = var_299, weight = nlp_net_default_encoder_transformer_layers_0_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_15_cast_fp16)[name = string("x_10_cast_fp16")];
            fp16 var_304_to_fp16 = const()[name = string("op_304_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_305_cast_fp16 = mul(x = x_10_cast_fp16, y = var_304_to_fp16)[name = string("op_305_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_306_cast_fp16 = sigmoid(x = var_305_cast_fp16)[name = string("op_306_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_17_cast_fp16 = mul(x = x_10_cast_fp16, y = var_306_cast_fp16)[name = string("input_17_cast_fp16")];
            tensor<int32, [2]> var_310 = const()[name = string("op_310"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_312 = const()[name = string("op_312"), val = tensor<int32, [2]>([1, 1])];
            string input0_3_pad_type_0 = const()[name = string("input0_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_3_pad_0 = const()[name = string("input0_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28008192))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(28007104))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29056832)))];
            tensor<fp16, [1, 512, 1, 256]> input0_3_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_bias_to_fp16, dilations = var_312, groups = var_46, pad = input0_3_pad_0, pad_type = input0_3_pad_type_0, strides = var_310, weight = nlp_net_default_encoder_transformer_layers_0_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_17_cast_fp16)[name = string("input0_3_cast_fp16")];
            tensor<int32, [1]> input_19_axes_0 = const()[name = string("input_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_19_gamma_0_to_fp16 = const()[name = string("input_19_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51398912)))];
            tensor<fp16, [512]> input_19_beta_0_to_fp16 = const()[name = string("input_19_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51400000)))];
            fp16 var_327_to_fp16 = const()[name = string("op_327_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = input_19_beta_0_to_fp16, epsilon = var_327_to_fp16, gamma = input_19_gamma_0_to_fp16, x = input0_3_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<int32, [2]> var_339 = const()[name = string("op_339"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_341 = const()[name = string("op_341"), val = tensor<int32, [2]>([1, 1])];
            string x_12_pad_type_0 = const()[name = string("x_12_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_12_pad_0 = const()[name = string("x_12_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51401408))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51401088))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51467008)))];
            tensor<fp16, [1, 128, 1, 256]> x_12_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_341, groups = var_46, pad = x_12_pad_0, pad_type = x_12_pad_type_0, strides = var_339, weight = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_19_cast_fp16)[name = string("x_12_cast_fp16")];
            fp16 var_344_to_fp16 = const()[name = string("op_344_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_345_cast_fp16 = mul(x = x_12_cast_fp16, y = var_344_to_fp16)[name = string("op_345_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_346_cast_fp16 = sigmoid(x = var_345_cast_fp16)[name = string("op_346_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_21_cast_fp16 = mul(x = x_12_cast_fp16, y = var_346_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<int32, [2]> var_350 = const()[name = string("op_350"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_352 = const()[name = string("op_352"), val = tensor<int32, [2]>([1, 1])];
            string x_14_pad_type_0 = const()[name = string("x_14_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_14_pad_0 = const()[name = string("x_14_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51468416))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51467328))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51534016)))];
            tensor<fp16, [1, 512, 1, 256]> x_14_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_352, groups = var_46, pad = x_14_pad_0, pad_type = x_14_pad_type_0, strides = var_350, weight = nlp_net_default_encoder_transformer_layers_0_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_21_cast_fp16)[name = string("x_14_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_2_cast_fp16 = add(x = x_14_cast_fp16, y = input0_3_cast_fp16)[name = string("f_2_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_2_cast_fp16 = add(x = f_2_cast_fp16, y = inputs_2_cast_fp16)[name = string("x1_2_cast_fp16")];
            fp16 var_357_to_fp16 = const()[name = string("op_357_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_358_cast_fp16 = mul(x = transpose_9, y = var_357_to_fp16)[name = string("op_358_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs_3_cast_fp16 = add(x = var_358_cast_fp16, y = x1_2_cast_fp16)[name = string("inputs_3_cast_fp16")];
            tensor<int32, [1]> k_7_axes_0 = const()[name = string("k_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_7_gamma_0_to_fp16 = const()[name = string("k_7_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29057920)))];
            tensor<fp16, [512]> k_7_beta_0_to_fp16 = const()[name = string("k_7_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29059008)))];
            fp16 var_376_to_fp16 = const()[name = string("op_376_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_7_cast_fp16 = layer_norm(axes = k_7_axes_0, beta = k_7_beta_0_to_fp16, epsilon = var_376_to_fp16, gamma = k_7_gamma_0_to_fp16, x = inputs_3_cast_fp16)[name = string("k_7_cast_fp16")];
            tensor<int32, [2]> var_395 = const()[name = string("op_395"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_397 = const()[name = string("op_397"), val = tensor<int32, [2]>([1, 1])];
            string var_399_pad_type_0 = const()[name = string("op_399_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_399_pad_0 = const()[name = string("op_399_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29061184))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29060096))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29323392)))];
            tensor<fp16, [1, 512, 1, 256]> var_399_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_bias_to_fp16, dilations = var_397, groups = var_46, pad = var_399_pad_0, pad_type = var_399_pad_type_0, strides = var_395, weight = nlp_net_default_encoder_transformer_layers_1_attn_q_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_399_cast_fp16")];
            tensor<int32, [2]> var_402 = const()[name = string("op_402"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_404 = const()[name = string("op_404"), val = tensor<int32, [2]>([1, 1])];
            string k_9_pad_type_0 = const()[name = string("k_9_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_9_pad_0 = const()[name = string("k_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29325568))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29324480))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29587776)))];
            tensor<fp16, [1, 512, 1, 256]> k_9_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_bias_to_fp16, dilations = var_404, groups = var_46, pad = k_9_pad_0, pad_type = k_9_pad_type_0, strides = var_402, weight = nlp_net_default_encoder_transformer_layers_1_attn_k_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<int32, [2]> var_409 = const()[name = string("op_409"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_411 = const()[name = string("op_411"), val = tensor<int32, [2]>([1, 1])];
            string var_413_pad_type_0 = const()[name = string("op_413_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_413_pad_0 = const()[name = string("op_413_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29589952))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29588864))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29852160)))];
            tensor<fp16, [1, 512, 1, 256]> var_413_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_bias_to_fp16, dilations = var_411, groups = var_46, pad = var_413_pad_0, pad_type = var_413_pad_type_0, strides = var_409, weight = nlp_net_default_encoder_transformer_layers_1_attn_v_proj_weight_to_fp16_affine_quantized, x = k_7_cast_fp16)[name = string("op_413_cast_fp16")];
            tensor<int32, [8]> tile_4 = const()[name = string("tile_4"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_414_axis_0 = const()[name = string("op_414_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_414_cast_fp16_7 = split(axis = var_414_axis_0, split_sizes = tile_4, x = var_399_cast_fp16)[name = string("op_414_cast_fp16")];
            tensor<int32, [4]> var_423_perm_0 = const()[name = string("op_423_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_5 = const()[name = string("tile_5"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_424_axis_0 = const()[name = string("op_424_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_7 = transpose(perm = var_423_perm_0, x = k_9_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_424_cast_fp16_7 = split(axis = var_424_axis_0, split_sizes = tile_5, x = transpose_7)[name = string("op_424_cast_fp16")];
            tensor<int32, [8]> tile_6 = const()[name = string("tile_6"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_433_axis_0 = const()[name = string("op_433_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_433_cast_fp16_7 = split(axis = var_433_axis_0, split_sizes = tile_6, x = var_413_cast_fp16)[name = string("op_433_cast_fp16")];
            string var_443_equation_0 = const()[name = string("op_443_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_443_cast_fp16 = einsum(equation = var_443_equation_0, values = (var_424_cast_fp16_0, var_414_cast_fp16_0))[name = string("op_443_cast_fp16")];
            fp16 var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_445_cast_fp16 = mul(x = var_443_cast_fp16, y = var_444_to_fp16)[name = string("op_445_cast_fp16")];
            string var_447_equation_0 = const()[name = string("op_447_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_447_cast_fp16 = einsum(equation = var_447_equation_0, values = (var_424_cast_fp16_1, var_414_cast_fp16_1))[name = string("op_447_cast_fp16")];
            fp16 var_448_to_fp16 = const()[name = string("op_448_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_449_cast_fp16 = mul(x = var_447_cast_fp16, y = var_448_to_fp16)[name = string("op_449_cast_fp16")];
            string var_451_equation_0 = const()[name = string("op_451_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_451_cast_fp16 = einsum(equation = var_451_equation_0, values = (var_424_cast_fp16_2, var_414_cast_fp16_2))[name = string("op_451_cast_fp16")];
            fp16 var_452_to_fp16 = const()[name = string("op_452_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_453_cast_fp16 = mul(x = var_451_cast_fp16, y = var_452_to_fp16)[name = string("op_453_cast_fp16")];
            string var_455_equation_0 = const()[name = string("op_455_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_455_cast_fp16 = einsum(equation = var_455_equation_0, values = (var_424_cast_fp16_3, var_414_cast_fp16_3))[name = string("op_455_cast_fp16")];
            fp16 var_456_to_fp16 = const()[name = string("op_456_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_457_cast_fp16 = mul(x = var_455_cast_fp16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")];
            string var_459_equation_0 = const()[name = string("op_459_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_459_cast_fp16 = einsum(equation = var_459_equation_0, values = (var_424_cast_fp16_4, var_414_cast_fp16_4))[name = string("op_459_cast_fp16")];
            fp16 var_460_to_fp16 = const()[name = string("op_460_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_461_cast_fp16 = mul(x = var_459_cast_fp16, y = var_460_to_fp16)[name = string("op_461_cast_fp16")];
            string var_463_equation_0 = const()[name = string("op_463_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_463_cast_fp16 = einsum(equation = var_463_equation_0, values = (var_424_cast_fp16_5, var_414_cast_fp16_5))[name = string("op_463_cast_fp16")];
            fp16 var_464_to_fp16 = const()[name = string("op_464_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_465_cast_fp16 = mul(x = var_463_cast_fp16, y = var_464_to_fp16)[name = string("op_465_cast_fp16")];
            string var_467_equation_0 = const()[name = string("op_467_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_467_cast_fp16 = einsum(equation = var_467_equation_0, values = (var_424_cast_fp16_6, var_414_cast_fp16_6))[name = string("op_467_cast_fp16")];
            fp16 var_468_to_fp16 = const()[name = string("op_468_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_469_cast_fp16 = mul(x = var_467_cast_fp16, y = var_468_to_fp16)[name = string("op_469_cast_fp16")];
            string var_471_equation_0 = const()[name = string("op_471_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_471_cast_fp16 = einsum(equation = var_471_equation_0, values = (var_424_cast_fp16_7, var_414_cast_fp16_7))[name = string("op_471_cast_fp16")];
            fp16 var_472_to_fp16 = const()[name = string("op_472_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_473_cast_fp16 = mul(x = var_471_cast_fp16, y = var_472_to_fp16)[name = string("op_473_cast_fp16")];
            bool attn_weights_4_interleave_0 = const()[name = string("attn_weights_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_4_cast_fp16 = concat(axis = var_47, interleave = attn_weights_4_interleave_0, values = (var_445_cast_fp16, var_449_cast_fp16, var_453_cast_fp16, var_457_cast_fp16, var_461_cast_fp16, var_465_cast_fp16, var_469_cast_fp16, var_473_cast_fp16))[name = string("attn_weights_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_4_cast_fp16 = add(x = attn_weights_4_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_4_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_23_cast_fp16 = softmax(axis = var_46, x = attn_weights0_4_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<int32, [8]> tile_7 = const()[name = string("tile_7"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_479_axis_0 = const()[name = string("op_479_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_479_cast_fp16_7 = split(axis = var_479_axis_0, split_sizes = tile_7, x = input_23_cast_fp16)[name = string("op_479_cast_fp16")];
            string var_489_equation_0 = const()[name = string("op_489_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_489_cast_fp16 = einsum(equation = var_489_equation_0, values = (var_433_cast_fp16_0, var_479_cast_fp16_0))[name = string("op_489_cast_fp16")];
            string var_491_equation_0 = const()[name = string("op_491_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_491_cast_fp16 = einsum(equation = var_491_equation_0, values = (var_433_cast_fp16_1, var_479_cast_fp16_1))[name = string("op_491_cast_fp16")];
            string var_493_equation_0 = const()[name = string("op_493_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_493_cast_fp16 = einsum(equation = var_493_equation_0, values = (var_433_cast_fp16_2, var_479_cast_fp16_2))[name = string("op_493_cast_fp16")];
            string var_495_equation_0 = const()[name = string("op_495_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_495_cast_fp16 = einsum(equation = var_495_equation_0, values = (var_433_cast_fp16_3, var_479_cast_fp16_3))[name = string("op_495_cast_fp16")];
            string var_497_equation_0 = const()[name = string("op_497_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_497_cast_fp16 = einsum(equation = var_497_equation_0, values = (var_433_cast_fp16_4, var_479_cast_fp16_4))[name = string("op_497_cast_fp16")];
            string var_499_equation_0 = const()[name = string("op_499_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_499_cast_fp16 = einsum(equation = var_499_equation_0, values = (var_433_cast_fp16_5, var_479_cast_fp16_5))[name = string("op_499_cast_fp16")];
            string var_501_equation_0 = const()[name = string("op_501_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_501_cast_fp16 = einsum(equation = var_501_equation_0, values = (var_433_cast_fp16_6, var_479_cast_fp16_6))[name = string("op_501_cast_fp16")];
            string var_503_equation_0 = const()[name = string("op_503_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_503_cast_fp16 = einsum(equation = var_503_equation_0, values = (var_433_cast_fp16_7, var_479_cast_fp16_7))[name = string("op_503_cast_fp16")];
            bool attn_11_interleave_0 = const()[name = string("attn_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_11_cast_fp16 = concat(axis = var_46, interleave = attn_11_interleave_0, values = (var_489_cast_fp16, var_491_cast_fp16, var_493_cast_fp16, var_495_cast_fp16, var_497_cast_fp16, var_499_cast_fp16, var_501_cast_fp16, var_503_cast_fp16))[name = string("attn_11_cast_fp16")];
            tensor<int32, [2]> var_511 = const()[name = string("op_511"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_513 = const()[name = string("op_513"), val = tensor<int32, [2]>([1, 1])];
            string inputs_9_pad_type_0 = const()[name = string("inputs_9_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_9_pad_0 = const()[name = string("inputs_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29854336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(29853248))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30116544)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_9_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_bias_to_fp16, dilations = var_513, groups = var_46, pad = inputs_9_pad_0, pad_type = inputs_9_pad_type_0, strides = var_511, weight = nlp_net_default_encoder_transformer_layers_1_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_11_cast_fp16)[name = string("inputs_9_cast_fp16")];
            tensor<int32, [1]> input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_25_gamma_0_to_fp16 = const()[name = string("input_25_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51535104)))];
            tensor<fp16, [512]> input_25_beta_0_to_fp16 = const()[name = string("input_25_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51536192)))];
            fp16 var_527_to_fp16 = const()[name = string("op_527_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = input_25_beta_0_to_fp16, epsilon = var_527_to_fp16, gamma = input_25_gamma_0_to_fp16, x = inputs_9_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<int32, [2]> var_539 = const()[name = string("op_539"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_541 = const()[name = string("op_541"), val = tensor<int32, [2]>([1, 1])];
            string x_16_pad_type_0 = const()[name = string("x_16_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_16_pad_0 = const()[name = string("x_16_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51537600))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51537280))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51603200)))];
            tensor<fp16, [1, 128, 1, 256]> x_16_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_bias_to_fp16, dilations = var_541, groups = var_46, pad = x_16_pad_0, pad_type = x_16_pad_type_0, strides = var_539, weight = nlp_net_default_encoder_transformer_layers_1_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_25_cast_fp16)[name = string("x_16_cast_fp16")];
            fp16 var_544_to_fp16 = const()[name = string("op_544_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_545_cast_fp16 = mul(x = x_16_cast_fp16, y = var_544_to_fp16)[name = string("op_545_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_546_cast_fp16 = sigmoid(x = var_545_cast_fp16)[name = string("op_546_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_27_cast_fp16 = mul(x = x_16_cast_fp16, y = var_546_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<int32, [2]> var_550 = const()[name = string("op_550"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_552 = const()[name = string("op_552"), val = tensor<int32, [2]>([1, 1])];
            string x_18_pad_type_0 = const()[name = string("x_18_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_18_pad_0 = const()[name = string("x_18_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51604608))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51603520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51670208)))];
            tensor<fp16, [1, 512, 1, 256]> x_18_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_bias_to_fp16, dilations = var_552, groups = var_46, pad = x_18_pad_0, pad_type = x_18_pad_type_0, strides = var_550, weight = nlp_net_default_encoder_transformer_layers_1_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_27_cast_fp16)[name = string("x_18_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_13_cast_fp16 = add(x = x_18_cast_fp16, y = inputs_9_cast_fp16)[name = string("attn_13_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_4_cast_fp16 = add(x = inputs_3_cast_fp16, y = attn_13_cast_fp16)[name = string("inputs0_4_cast_fp16")];
            tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_29_gamma_0_to_fp16 = const()[name = string("input_29_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30117632)))];
            tensor<fp16, [512]> input_29_beta_0_to_fp16 = const()[name = string("input_29_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30118720)))];
            fp16 var_565_to_fp16 = const()[name = string("op_565_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = input_29_beta_0_to_fp16, epsilon = var_565_to_fp16, gamma = input_29_gamma_0_to_fp16, x = inputs0_4_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<int32, [2]> var_579 = const()[name = string("op_579"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_581 = const()[name = string("op_581"), val = tensor<int32, [2]>([1, 1])];
            string x_20_pad_type_0 = const()[name = string("x_20_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_20_pad_0 = const()[name = string("x_20_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30123968))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(30119808))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31176768))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31172608))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_20_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_581, groups = var_46, pad = x_20_pad_0, pad_type = x_20_pad_type_0, strides = var_579, weight = nlp_net_default_encoder_transformer_layers_1_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_29_cast_fp16)[name = string("x_20_cast_fp16")];
            fp16 var_584_to_fp16 = const()[name = string("op_584_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_585_cast_fp16 = mul(x = x_20_cast_fp16, y = var_584_to_fp16)[name = string("op_585_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_586_cast_fp16 = sigmoid(x = var_585_cast_fp16)[name = string("op_586_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_31_cast_fp16 = mul(x = x_20_cast_fp16, y = var_586_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<int32, [2]> var_590 = const()[name = string("op_590"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_592 = const()[name = string("op_592"), val = tensor<int32, [2]>([1, 1])];
            string input0_7_pad_type_0 = const()[name = string("input0_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_7_pad_0 = const()[name = string("input0_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31179968))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(31178880))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32228608)))];
            tensor<fp16, [1, 512, 1, 256]> input0_7_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_bias_to_fp16, dilations = var_592, groups = var_46, pad = input0_7_pad_0, pad_type = input0_7_pad_type_0, strides = var_590, weight = nlp_net_default_encoder_transformer_layers_1_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_31_cast_fp16)[name = string("input0_7_cast_fp16")];
            tensor<int32, [1]> input_33_axes_0 = const()[name = string("input_33_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_33_gamma_0_to_fp16 = const()[name = string("input_33_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51671296)))];
            tensor<fp16, [512]> input_33_beta_0_to_fp16 = const()[name = string("input_33_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51672384)))];
            fp16 var_607_to_fp16 = const()[name = string("op_607_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = input_33_beta_0_to_fp16, epsilon = var_607_to_fp16, gamma = input_33_gamma_0_to_fp16, x = input0_7_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<int32, [2]> var_619 = const()[name = string("op_619"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_621 = const()[name = string("op_621"), val = tensor<int32, [2]>([1, 1])];
            string x_22_pad_type_0 = const()[name = string("x_22_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_22_pad_0 = const()[name = string("x_22_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51673792))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51673472))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51739392)))];
            tensor<fp16, [1, 128, 1, 256]> x_22_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_621, groups = var_46, pad = x_22_pad_0, pad_type = x_22_pad_type_0, strides = var_619, weight = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_33_cast_fp16)[name = string("x_22_cast_fp16")];
            fp16 var_624_to_fp16 = const()[name = string("op_624_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_625_cast_fp16 = mul(x = x_22_cast_fp16, y = var_624_to_fp16)[name = string("op_625_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_626_cast_fp16 = sigmoid(x = var_625_cast_fp16)[name = string("op_626_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_35_cast_fp16 = mul(x = x_22_cast_fp16, y = var_626_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<int32, [2]> var_630 = const()[name = string("op_630"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_632 = const()[name = string("op_632"), val = tensor<int32, [2]>([1, 1])];
            string x_24_pad_type_0 = const()[name = string("x_24_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_24_pad_0 = const()[name = string("x_24_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51740800))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51739712))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51806400)))];
            tensor<fp16, [1, 512, 1, 256]> x_24_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_632, groups = var_46, pad = x_24_pad_0, pad_type = x_24_pad_type_0, strides = var_630, weight = nlp_net_default_encoder_transformer_layers_1_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_35_cast_fp16)[name = string("x_24_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_4_cast_fp16 = add(x = x_24_cast_fp16, y = input0_7_cast_fp16)[name = string("f_4_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_4_cast_fp16 = add(x = f_4_cast_fp16, y = inputs0_4_cast_fp16)[name = string("x1_4_cast_fp16")];
            fp16 var_637_to_fp16 = const()[name = string("op_637_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_638_cast_fp16 = mul(x = inputs_3_cast_fp16, y = var_637_to_fp16)[name = string("op_638_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_2_cast_fp16 = add(x = var_638_cast_fp16, y = x1_4_cast_fp16)[name = string("inputs0_2_cast_fp16")];
            tensor<int32, [1]> k_11_axes_0 = const()[name = string("k_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_11_gamma_0_to_fp16 = const()[name = string("k_11_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32229696)))];
            tensor<fp16, [512]> k_11_beta_0_to_fp16 = const()[name = string("k_11_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32230784)))];
            fp16 var_656_to_fp16 = const()[name = string("op_656_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_11_cast_fp16 = layer_norm(axes = k_11_axes_0, beta = k_11_beta_0_to_fp16, epsilon = var_656_to_fp16, gamma = k_11_gamma_0_to_fp16, x = inputs0_2_cast_fp16)[name = string("k_11_cast_fp16")];
            tensor<int32, [2]> var_675 = const()[name = string("op_675"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_677 = const()[name = string("op_677"), val = tensor<int32, [2]>([1, 1])];
            string var_679_pad_type_0 = const()[name = string("op_679_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_679_pad_0 = const()[name = string("op_679_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32232960))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32231872))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32495168)))];
            tensor<fp16, [1, 512, 1, 256]> var_679_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_bias_to_fp16, dilations = var_677, groups = var_46, pad = var_679_pad_0, pad_type = var_679_pad_type_0, strides = var_675, weight = nlp_net_default_encoder_transformer_layers_2_attn_q_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_679_cast_fp16")];
            tensor<int32, [2]> var_682 = const()[name = string("op_682"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_684 = const()[name = string("op_684"), val = tensor<int32, [2]>([1, 1])];
            string k_13_pad_type_0 = const()[name = string("k_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_13_pad_0 = const()[name = string("k_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32497344))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32496256))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32759552)))];
            tensor<fp16, [1, 512, 1, 256]> k_13_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_bias_to_fp16, dilations = var_684, groups = var_46, pad = k_13_pad_0, pad_type = k_13_pad_type_0, strides = var_682, weight = nlp_net_default_encoder_transformer_layers_2_attn_k_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<int32, [2]> var_689 = const()[name = string("op_689"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_691 = const()[name = string("op_691"), val = tensor<int32, [2]>([1, 1])];
            string var_693_pad_type_0 = const()[name = string("op_693_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_693_pad_0 = const()[name = string("op_693_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32761728))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(32760640))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33023936)))];
            tensor<fp16, [1, 512, 1, 256]> var_693_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_bias_to_fp16, dilations = var_691, groups = var_46, pad = var_693_pad_0, pad_type = var_693_pad_type_0, strides = var_689, weight = nlp_net_default_encoder_transformer_layers_2_attn_v_proj_weight_to_fp16_affine_quantized, x = k_11_cast_fp16)[name = string("op_693_cast_fp16")];
            tensor<int32, [8]> tile_8 = const()[name = string("tile_8"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_694_axis_0 = const()[name = string("op_694_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_694_cast_fp16_7 = split(axis = var_694_axis_0, split_sizes = tile_8, x = var_679_cast_fp16)[name = string("op_694_cast_fp16")];
            tensor<int32, [4]> var_703_perm_0 = const()[name = string("op_703_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_9 = const()[name = string("tile_9"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_704_axis_0 = const()[name = string("op_704_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_6 = transpose(perm = var_703_perm_0, x = k_13_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_704_cast_fp16_7 = split(axis = var_704_axis_0, split_sizes = tile_9, x = transpose_6)[name = string("op_704_cast_fp16")];
            tensor<int32, [8]> tile_10 = const()[name = string("tile_10"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_713_axis_0 = const()[name = string("op_713_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_713_cast_fp16_7 = split(axis = var_713_axis_0, split_sizes = tile_10, x = var_693_cast_fp16)[name = string("op_713_cast_fp16")];
            string var_723_equation_0 = const()[name = string("op_723_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_723_cast_fp16 = einsum(equation = var_723_equation_0, values = (var_704_cast_fp16_0, var_694_cast_fp16_0))[name = string("op_723_cast_fp16")];
            fp16 var_724_to_fp16 = const()[name = string("op_724_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_725_cast_fp16 = mul(x = var_723_cast_fp16, y = var_724_to_fp16)[name = string("op_725_cast_fp16")];
            string var_727_equation_0 = const()[name = string("op_727_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_727_cast_fp16 = einsum(equation = var_727_equation_0, values = (var_704_cast_fp16_1, var_694_cast_fp16_1))[name = string("op_727_cast_fp16")];
            fp16 var_728_to_fp16 = const()[name = string("op_728_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_729_cast_fp16 = mul(x = var_727_cast_fp16, y = var_728_to_fp16)[name = string("op_729_cast_fp16")];
            string var_731_equation_0 = const()[name = string("op_731_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_731_cast_fp16 = einsum(equation = var_731_equation_0, values = (var_704_cast_fp16_2, var_694_cast_fp16_2))[name = string("op_731_cast_fp16")];
            fp16 var_732_to_fp16 = const()[name = string("op_732_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_733_cast_fp16 = mul(x = var_731_cast_fp16, y = var_732_to_fp16)[name = string("op_733_cast_fp16")];
            string var_735_equation_0 = const()[name = string("op_735_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_735_cast_fp16 = einsum(equation = var_735_equation_0, values = (var_704_cast_fp16_3, var_694_cast_fp16_3))[name = string("op_735_cast_fp16")];
            fp16 var_736_to_fp16 = const()[name = string("op_736_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_737_cast_fp16 = mul(x = var_735_cast_fp16, y = var_736_to_fp16)[name = string("op_737_cast_fp16")];
            string var_739_equation_0 = const()[name = string("op_739_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_739_cast_fp16 = einsum(equation = var_739_equation_0, values = (var_704_cast_fp16_4, var_694_cast_fp16_4))[name = string("op_739_cast_fp16")];
            fp16 var_740_to_fp16 = const()[name = string("op_740_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_741_cast_fp16 = mul(x = var_739_cast_fp16, y = var_740_to_fp16)[name = string("op_741_cast_fp16")];
            string var_743_equation_0 = const()[name = string("op_743_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_743_cast_fp16 = einsum(equation = var_743_equation_0, values = (var_704_cast_fp16_5, var_694_cast_fp16_5))[name = string("op_743_cast_fp16")];
            fp16 var_744_to_fp16 = const()[name = string("op_744_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_745_cast_fp16 = mul(x = var_743_cast_fp16, y = var_744_to_fp16)[name = string("op_745_cast_fp16")];
            string var_747_equation_0 = const()[name = string("op_747_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_747_cast_fp16 = einsum(equation = var_747_equation_0, values = (var_704_cast_fp16_6, var_694_cast_fp16_6))[name = string("op_747_cast_fp16")];
            fp16 var_748_to_fp16 = const()[name = string("op_748_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_749_cast_fp16 = mul(x = var_747_cast_fp16, y = var_748_to_fp16)[name = string("op_749_cast_fp16")];
            string var_751_equation_0 = const()[name = string("op_751_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_751_cast_fp16 = einsum(equation = var_751_equation_0, values = (var_704_cast_fp16_7, var_694_cast_fp16_7))[name = string("op_751_cast_fp16")];
            fp16 var_752_to_fp16 = const()[name = string("op_752_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_753_cast_fp16 = mul(x = var_751_cast_fp16, y = var_752_to_fp16)[name = string("op_753_cast_fp16")];
            bool attn_weights_6_interleave_0 = const()[name = string("attn_weights_6_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_6_cast_fp16 = concat(axis = var_47, interleave = attn_weights_6_interleave_0, values = (var_725_cast_fp16, var_729_cast_fp16, var_733_cast_fp16, var_737_cast_fp16, var_741_cast_fp16, var_745_cast_fp16, var_749_cast_fp16, var_753_cast_fp16))[name = string("attn_weights_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_6_cast_fp16 = add(x = attn_weights_6_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_6_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_37_cast_fp16 = softmax(axis = var_46, x = attn_weights0_6_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<int32, [8]> tile_11 = const()[name = string("tile_11"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_759_axis_0 = const()[name = string("op_759_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_759_cast_fp16_7 = split(axis = var_759_axis_0, split_sizes = tile_11, x = input_37_cast_fp16)[name = string("op_759_cast_fp16")];
            string var_769_equation_0 = const()[name = string("op_769_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_769_cast_fp16 = einsum(equation = var_769_equation_0, values = (var_713_cast_fp16_0, var_759_cast_fp16_0))[name = string("op_769_cast_fp16")];
            string var_771_equation_0 = const()[name = string("op_771_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_771_cast_fp16 = einsum(equation = var_771_equation_0, values = (var_713_cast_fp16_1, var_759_cast_fp16_1))[name = string("op_771_cast_fp16")];
            string var_773_equation_0 = const()[name = string("op_773_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_773_cast_fp16 = einsum(equation = var_773_equation_0, values = (var_713_cast_fp16_2, var_759_cast_fp16_2))[name = string("op_773_cast_fp16")];
            string var_775_equation_0 = const()[name = string("op_775_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_775_cast_fp16 = einsum(equation = var_775_equation_0, values = (var_713_cast_fp16_3, var_759_cast_fp16_3))[name = string("op_775_cast_fp16")];
            string var_777_equation_0 = const()[name = string("op_777_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_777_cast_fp16 = einsum(equation = var_777_equation_0, values = (var_713_cast_fp16_4, var_759_cast_fp16_4))[name = string("op_777_cast_fp16")];
            string var_779_equation_0 = const()[name = string("op_779_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_779_cast_fp16 = einsum(equation = var_779_equation_0, values = (var_713_cast_fp16_5, var_759_cast_fp16_5))[name = string("op_779_cast_fp16")];
            string var_781_equation_0 = const()[name = string("op_781_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_781_cast_fp16 = einsum(equation = var_781_equation_0, values = (var_713_cast_fp16_6, var_759_cast_fp16_6))[name = string("op_781_cast_fp16")];
            string var_783_equation_0 = const()[name = string("op_783_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_783_cast_fp16 = einsum(equation = var_783_equation_0, values = (var_713_cast_fp16_7, var_759_cast_fp16_7))[name = string("op_783_cast_fp16")];
            bool attn_17_interleave_0 = const()[name = string("attn_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_17_cast_fp16 = concat(axis = var_46, interleave = attn_17_interleave_0, values = (var_769_cast_fp16, var_771_cast_fp16, var_773_cast_fp16, var_775_cast_fp16, var_777_cast_fp16, var_779_cast_fp16, var_781_cast_fp16, var_783_cast_fp16))[name = string("attn_17_cast_fp16")];
            tensor<int32, [2]> var_791 = const()[name = string("op_791"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_793 = const()[name = string("op_793"), val = tensor<int32, [2]>([1, 1])];
            string inputs_13_pad_type_0 = const()[name = string("inputs_13_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_13_pad_0 = const()[name = string("inputs_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33026112))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33025024))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33288320)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_13_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_bias_to_fp16, dilations = var_793, groups = var_46, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = var_791, weight = nlp_net_default_encoder_transformer_layers_2_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_17_cast_fp16)[name = string("inputs_13_cast_fp16")];
            tensor<int32, [1]> input_39_axes_0 = const()[name = string("input_39_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_39_gamma_0_to_fp16 = const()[name = string("input_39_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51807488)))];
            tensor<fp16, [512]> input_39_beta_0_to_fp16 = const()[name = string("input_39_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51808576)))];
            fp16 var_807_to_fp16 = const()[name = string("op_807_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = input_39_beta_0_to_fp16, epsilon = var_807_to_fp16, gamma = input_39_gamma_0_to_fp16, x = inputs_13_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<int32, [2]> var_819 = const()[name = string("op_819"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_821 = const()[name = string("op_821"), val = tensor<int32, [2]>([1, 1])];
            string x_26_pad_type_0 = const()[name = string("x_26_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_26_pad_0 = const()[name = string("x_26_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51809984))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51809664))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51875584)))];
            tensor<fp16, [1, 128, 1, 256]> x_26_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_bias_to_fp16, dilations = var_821, groups = var_46, pad = x_26_pad_0, pad_type = x_26_pad_type_0, strides = var_819, weight = nlp_net_default_encoder_transformer_layers_2_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_39_cast_fp16)[name = string("x_26_cast_fp16")];
            fp16 var_824_to_fp16 = const()[name = string("op_824_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_825_cast_fp16 = mul(x = x_26_cast_fp16, y = var_824_to_fp16)[name = string("op_825_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_826_cast_fp16 = sigmoid(x = var_825_cast_fp16)[name = string("op_826_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_41_cast_fp16 = mul(x = x_26_cast_fp16, y = var_826_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<int32, [2]> var_830 = const()[name = string("op_830"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_832 = const()[name = string("op_832"), val = tensor<int32, [2]>([1, 1])];
            string x_28_pad_type_0 = const()[name = string("x_28_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_28_pad_0 = const()[name = string("x_28_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51876992))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51875904))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51942592)))];
            tensor<fp16, [1, 512, 1, 256]> x_28_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_bias_to_fp16, dilations = var_832, groups = var_46, pad = x_28_pad_0, pad_type = x_28_pad_type_0, strides = var_830, weight = nlp_net_default_encoder_transformer_layers_2_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_41_cast_fp16)[name = string("x_28_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_19_cast_fp16 = add(x = x_28_cast_fp16, y = inputs_13_cast_fp16)[name = string("attn_19_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_6_cast_fp16 = add(x = inputs0_2_cast_fp16, y = attn_19_cast_fp16)[name = string("inputs0_6_cast_fp16")];
            tensor<int32, [1]> input_43_axes_0 = const()[name = string("input_43_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_43_gamma_0_to_fp16 = const()[name = string("input_43_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33289408)))];
            tensor<fp16, [512]> input_43_beta_0_to_fp16 = const()[name = string("input_43_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33290496)))];
            fp16 var_845_to_fp16 = const()[name = string("op_845_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = input_43_beta_0_to_fp16, epsilon = var_845_to_fp16, gamma = input_43_gamma_0_to_fp16, x = inputs0_6_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<int32, [2]> var_859 = const()[name = string("op_859"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_861 = const()[name = string("op_861"), val = tensor<int32, [2]>([1, 1])];
            string x_30_pad_type_0 = const()[name = string("x_30_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_30_pad_0 = const()[name = string("x_30_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33295744))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(33291584))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34348544))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34344384))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_30_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_861, groups = var_46, pad = x_30_pad_0, pad_type = x_30_pad_type_0, strides = var_859, weight = nlp_net_default_encoder_transformer_layers_2_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_43_cast_fp16)[name = string("x_30_cast_fp16")];
            fp16 var_864_to_fp16 = const()[name = string("op_864_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_865_cast_fp16 = mul(x = x_30_cast_fp16, y = var_864_to_fp16)[name = string("op_865_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_866_cast_fp16 = sigmoid(x = var_865_cast_fp16)[name = string("op_866_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_45_cast_fp16 = mul(x = x_30_cast_fp16, y = var_866_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<int32, [2]> var_870 = const()[name = string("op_870"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_872 = const()[name = string("op_872"), val = tensor<int32, [2]>([1, 1])];
            string input0_11_pad_type_0 = const()[name = string("input0_11_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_11_pad_0 = const()[name = string("input0_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34351744))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(34350656))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35400384)))];
            tensor<fp16, [1, 512, 1, 256]> input0_11_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_bias_to_fp16, dilations = var_872, groups = var_46, pad = input0_11_pad_0, pad_type = input0_11_pad_type_0, strides = var_870, weight = nlp_net_default_encoder_transformer_layers_2_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_45_cast_fp16)[name = string("input0_11_cast_fp16")];
            tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_47_gamma_0_to_fp16 = const()[name = string("input_47_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51943680)))];
            tensor<fp16, [512]> input_47_beta_0_to_fp16 = const()[name = string("input_47_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51944768)))];
            fp16 var_887_to_fp16 = const()[name = string("op_887_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = input_47_beta_0_to_fp16, epsilon = var_887_to_fp16, gamma = input_47_gamma_0_to_fp16, x = input0_11_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<int32, [2]> var_899 = const()[name = string("op_899"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_901 = const()[name = string("op_901"), val = tensor<int32, [2]>([1, 1])];
            string x_32_pad_type_0 = const()[name = string("x_32_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_32_pad_0 = const()[name = string("x_32_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51946176))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51945856))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52011776)))];
            tensor<fp16, [1, 128, 1, 256]> x_32_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_901, groups = var_46, pad = x_32_pad_0, pad_type = x_32_pad_type_0, strides = var_899, weight = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_47_cast_fp16)[name = string("x_32_cast_fp16")];
            fp16 var_904_to_fp16 = const()[name = string("op_904_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_905_cast_fp16 = mul(x = x_32_cast_fp16, y = var_904_to_fp16)[name = string("op_905_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_906_cast_fp16 = sigmoid(x = var_905_cast_fp16)[name = string("op_906_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_49_cast_fp16 = mul(x = x_32_cast_fp16, y = var_906_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<int32, [2]> var_910 = const()[name = string("op_910"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_912 = const()[name = string("op_912"), val = tensor<int32, [2]>([1, 1])];
            string x_34_pad_type_0 = const()[name = string("x_34_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_34_pad_0 = const()[name = string("x_34_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52013184))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52012096))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52078784)))];
            tensor<fp16, [1, 512, 1, 256]> x_34_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_912, groups = var_46, pad = x_34_pad_0, pad_type = x_34_pad_type_0, strides = var_910, weight = nlp_net_default_encoder_transformer_layers_2_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_49_cast_fp16)[name = string("x_34_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_6_cast_fp16 = add(x = x_34_cast_fp16, y = input0_11_cast_fp16)[name = string("f_6_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_6_cast_fp16 = add(x = f_6_cast_fp16, y = inputs0_6_cast_fp16)[name = string("x1_6_cast_fp16")];
            fp16 var_917_to_fp16 = const()[name = string("op_917_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_918_cast_fp16 = mul(x = inputs0_2_cast_fp16, y = var_917_to_fp16)[name = string("op_918_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs1_1_cast_fp16 = add(x = var_918_cast_fp16, y = x1_6_cast_fp16)[name = string("inputs1_1_cast_fp16")];
            tensor<int32, [1]> k_15_axes_0 = const()[name = string("k_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_15_gamma_0_to_fp16 = const()[name = string("k_15_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35401472)))];
            tensor<fp16, [512]> k_15_beta_0_to_fp16 = const()[name = string("k_15_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35402560)))];
            fp16 var_936_to_fp16 = const()[name = string("op_936_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_15_cast_fp16 = layer_norm(axes = k_15_axes_0, beta = k_15_beta_0_to_fp16, epsilon = var_936_to_fp16, gamma = k_15_gamma_0_to_fp16, x = inputs1_1_cast_fp16)[name = string("k_15_cast_fp16")];
            tensor<int32, [2]> var_955 = const()[name = string("op_955"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_957 = const()[name = string("op_957"), val = tensor<int32, [2]>([1, 1])];
            string var_959_pad_type_0 = const()[name = string("op_959_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_959_pad_0 = const()[name = string("op_959_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35404736))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35403648))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35666944)))];
            tensor<fp16, [1, 512, 1, 256]> var_959_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_bias_to_fp16, dilations = var_957, groups = var_46, pad = var_959_pad_0, pad_type = var_959_pad_type_0, strides = var_955, weight = nlp_net_default_encoder_transformer_layers_3_attn_q_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_959_cast_fp16")];
            tensor<int32, [2]> var_962 = const()[name = string("op_962"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_964 = const()[name = string("op_964"), val = tensor<int32, [2]>([1, 1])];
            string k_17_pad_type_0 = const()[name = string("k_17_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_17_pad_0 = const()[name = string("k_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35669120))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35668032))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35931328)))];
            tensor<fp16, [1, 512, 1, 256]> k_17_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_bias_to_fp16, dilations = var_964, groups = var_46, pad = k_17_pad_0, pad_type = k_17_pad_type_0, strides = var_962, weight = nlp_net_default_encoder_transformer_layers_3_attn_k_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<int32, [2]> var_969 = const()[name = string("op_969"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_971 = const()[name = string("op_971"), val = tensor<int32, [2]>([1, 1])];
            string var_973_pad_type_0 = const()[name = string("op_973_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_973_pad_0 = const()[name = string("op_973_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35933504))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(35932416))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36195712)))];
            tensor<fp16, [1, 512, 1, 256]> var_973_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_bias_to_fp16, dilations = var_971, groups = var_46, pad = var_973_pad_0, pad_type = var_973_pad_type_0, strides = var_969, weight = nlp_net_default_encoder_transformer_layers_3_attn_v_proj_weight_to_fp16_affine_quantized, x = k_15_cast_fp16)[name = string("op_973_cast_fp16")];
            tensor<int32, [8]> tile_12 = const()[name = string("tile_12"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_974_axis_0 = const()[name = string("op_974_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_974_cast_fp16_7 = split(axis = var_974_axis_0, split_sizes = tile_12, x = var_959_cast_fp16)[name = string("op_974_cast_fp16")];
            tensor<int32, [4]> var_983_perm_0 = const()[name = string("op_983_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_13 = const()[name = string("tile_13"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_984_axis_0 = const()[name = string("op_984_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_5 = transpose(perm = var_983_perm_0, x = k_17_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_984_cast_fp16_7 = split(axis = var_984_axis_0, split_sizes = tile_13, x = transpose_5)[name = string("op_984_cast_fp16")];
            tensor<int32, [8]> tile_14 = const()[name = string("tile_14"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_993_axis_0 = const()[name = string("op_993_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_993_cast_fp16_7 = split(axis = var_993_axis_0, split_sizes = tile_14, x = var_973_cast_fp16)[name = string("op_993_cast_fp16")];
            string var_1003_equation_0 = const()[name = string("op_1003_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1003_cast_fp16 = einsum(equation = var_1003_equation_0, values = (var_984_cast_fp16_0, var_974_cast_fp16_0))[name = string("op_1003_cast_fp16")];
            fp16 var_1004_to_fp16 = const()[name = string("op_1004_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1005_cast_fp16 = mul(x = var_1003_cast_fp16, y = var_1004_to_fp16)[name = string("op_1005_cast_fp16")];
            string var_1007_equation_0 = const()[name = string("op_1007_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1007_cast_fp16 = einsum(equation = var_1007_equation_0, values = (var_984_cast_fp16_1, var_974_cast_fp16_1))[name = string("op_1007_cast_fp16")];
            fp16 var_1008_to_fp16 = const()[name = string("op_1008_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1009_cast_fp16 = mul(x = var_1007_cast_fp16, y = var_1008_to_fp16)[name = string("op_1009_cast_fp16")];
            string var_1011_equation_0 = const()[name = string("op_1011_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1011_cast_fp16 = einsum(equation = var_1011_equation_0, values = (var_984_cast_fp16_2, var_974_cast_fp16_2))[name = string("op_1011_cast_fp16")];
            fp16 var_1012_to_fp16 = const()[name = string("op_1012_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1013_cast_fp16 = mul(x = var_1011_cast_fp16, y = var_1012_to_fp16)[name = string("op_1013_cast_fp16")];
            string var_1015_equation_0 = const()[name = string("op_1015_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1015_cast_fp16 = einsum(equation = var_1015_equation_0, values = (var_984_cast_fp16_3, var_974_cast_fp16_3))[name = string("op_1015_cast_fp16")];
            fp16 var_1016_to_fp16 = const()[name = string("op_1016_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1017_cast_fp16 = mul(x = var_1015_cast_fp16, y = var_1016_to_fp16)[name = string("op_1017_cast_fp16")];
            string var_1019_equation_0 = const()[name = string("op_1019_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1019_cast_fp16 = einsum(equation = var_1019_equation_0, values = (var_984_cast_fp16_4, var_974_cast_fp16_4))[name = string("op_1019_cast_fp16")];
            fp16 var_1020_to_fp16 = const()[name = string("op_1020_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1021_cast_fp16 = mul(x = var_1019_cast_fp16, y = var_1020_to_fp16)[name = string("op_1021_cast_fp16")];
            string var_1023_equation_0 = const()[name = string("op_1023_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1023_cast_fp16 = einsum(equation = var_1023_equation_0, values = (var_984_cast_fp16_5, var_974_cast_fp16_5))[name = string("op_1023_cast_fp16")];
            fp16 var_1024_to_fp16 = const()[name = string("op_1024_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1025_cast_fp16 = mul(x = var_1023_cast_fp16, y = var_1024_to_fp16)[name = string("op_1025_cast_fp16")];
            string var_1027_equation_0 = const()[name = string("op_1027_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1027_cast_fp16 = einsum(equation = var_1027_equation_0, values = (var_984_cast_fp16_6, var_974_cast_fp16_6))[name = string("op_1027_cast_fp16")];
            fp16 var_1028_to_fp16 = const()[name = string("op_1028_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1029_cast_fp16 = mul(x = var_1027_cast_fp16, y = var_1028_to_fp16)[name = string("op_1029_cast_fp16")];
            string var_1031_equation_0 = const()[name = string("op_1031_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1031_cast_fp16 = einsum(equation = var_1031_equation_0, values = (var_984_cast_fp16_7, var_974_cast_fp16_7))[name = string("op_1031_cast_fp16")];
            fp16 var_1032_to_fp16 = const()[name = string("op_1032_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1033_cast_fp16 = mul(x = var_1031_cast_fp16, y = var_1032_to_fp16)[name = string("op_1033_cast_fp16")];
            bool attn_weights_8_interleave_0 = const()[name = string("attn_weights_8_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_8_cast_fp16 = concat(axis = var_47, interleave = attn_weights_8_interleave_0, values = (var_1005_cast_fp16, var_1009_cast_fp16, var_1013_cast_fp16, var_1017_cast_fp16, var_1021_cast_fp16, var_1025_cast_fp16, var_1029_cast_fp16, var_1033_cast_fp16))[name = string("attn_weights_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_8_cast_fp16 = add(x = attn_weights_8_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_8_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_51_cast_fp16 = softmax(axis = var_46, x = attn_weights0_8_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<int32, [8]> tile_15 = const()[name = string("tile_15"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1039_axis_0 = const()[name = string("op_1039_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1039_cast_fp16_7 = split(axis = var_1039_axis_0, split_sizes = tile_15, x = input_51_cast_fp16)[name = string("op_1039_cast_fp16")];
            string var_1049_equation_0 = const()[name = string("op_1049_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1049_cast_fp16 = einsum(equation = var_1049_equation_0, values = (var_993_cast_fp16_0, var_1039_cast_fp16_0))[name = string("op_1049_cast_fp16")];
            string var_1051_equation_0 = const()[name = string("op_1051_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1051_cast_fp16 = einsum(equation = var_1051_equation_0, values = (var_993_cast_fp16_1, var_1039_cast_fp16_1))[name = string("op_1051_cast_fp16")];
            string var_1053_equation_0 = const()[name = string("op_1053_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1053_cast_fp16 = einsum(equation = var_1053_equation_0, values = (var_993_cast_fp16_2, var_1039_cast_fp16_2))[name = string("op_1053_cast_fp16")];
            string var_1055_equation_0 = const()[name = string("op_1055_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1055_cast_fp16 = einsum(equation = var_1055_equation_0, values = (var_993_cast_fp16_3, var_1039_cast_fp16_3))[name = string("op_1055_cast_fp16")];
            string var_1057_equation_0 = const()[name = string("op_1057_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1057_cast_fp16 = einsum(equation = var_1057_equation_0, values = (var_993_cast_fp16_4, var_1039_cast_fp16_4))[name = string("op_1057_cast_fp16")];
            string var_1059_equation_0 = const()[name = string("op_1059_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1059_cast_fp16 = einsum(equation = var_1059_equation_0, values = (var_993_cast_fp16_5, var_1039_cast_fp16_5))[name = string("op_1059_cast_fp16")];
            string var_1061_equation_0 = const()[name = string("op_1061_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1061_cast_fp16 = einsum(equation = var_1061_equation_0, values = (var_993_cast_fp16_6, var_1039_cast_fp16_6))[name = string("op_1061_cast_fp16")];
            string var_1063_equation_0 = const()[name = string("op_1063_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1063_cast_fp16 = einsum(equation = var_1063_equation_0, values = (var_993_cast_fp16_7, var_1039_cast_fp16_7))[name = string("op_1063_cast_fp16")];
            bool attn_23_interleave_0 = const()[name = string("attn_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_23_cast_fp16 = concat(axis = var_46, interleave = attn_23_interleave_0, values = (var_1049_cast_fp16, var_1051_cast_fp16, var_1053_cast_fp16, var_1055_cast_fp16, var_1057_cast_fp16, var_1059_cast_fp16, var_1061_cast_fp16, var_1063_cast_fp16))[name = string("attn_23_cast_fp16")];
            tensor<int32, [2]> var_1071 = const()[name = string("op_1071"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1073 = const()[name = string("op_1073"), val = tensor<int32, [2]>([1, 1])];
            string inputs_17_pad_type_0 = const()[name = string("inputs_17_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_17_pad_0 = const()[name = string("inputs_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36197888))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36196800))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36460096)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_17_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_bias_to_fp16, dilations = var_1073, groups = var_46, pad = inputs_17_pad_0, pad_type = inputs_17_pad_type_0, strides = var_1071, weight = nlp_net_default_encoder_transformer_layers_3_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_23_cast_fp16)[name = string("inputs_17_cast_fp16")];
            tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_53_gamma_0_to_fp16 = const()[name = string("input_53_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52079872)))];
            tensor<fp16, [512]> input_53_beta_0_to_fp16 = const()[name = string("input_53_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52080960)))];
            fp16 var_1087_to_fp16 = const()[name = string("op_1087_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = input_53_beta_0_to_fp16, epsilon = var_1087_to_fp16, gamma = input_53_gamma_0_to_fp16, x = inputs_17_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<int32, [2]> var_1099 = const()[name = string("op_1099"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1101 = const()[name = string("op_1101"), val = tensor<int32, [2]>([1, 1])];
            string x_36_pad_type_0 = const()[name = string("x_36_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_36_pad_0 = const()[name = string("x_36_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52082368))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52082048))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52147968)))];
            tensor<fp16, [1, 128, 1, 256]> x_36_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1101, groups = var_46, pad = x_36_pad_0, pad_type = x_36_pad_type_0, strides = var_1099, weight = nlp_net_default_encoder_transformer_layers_3_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_53_cast_fp16)[name = string("x_36_cast_fp16")];
            fp16 var_1104_to_fp16 = const()[name = string("op_1104_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1105_cast_fp16 = mul(x = x_36_cast_fp16, y = var_1104_to_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1106_cast_fp16 = sigmoid(x = var_1105_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_55_cast_fp16 = mul(x = x_36_cast_fp16, y = var_1106_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<int32, [2]> var_1110 = const()[name = string("op_1110"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1112 = const()[name = string("op_1112"), val = tensor<int32, [2]>([1, 1])];
            string x_38_pad_type_0 = const()[name = string("x_38_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_38_pad_0 = const()[name = string("x_38_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52149376))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52148288))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52214976)))];
            tensor<fp16, [1, 512, 1, 256]> x_38_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1112, groups = var_46, pad = x_38_pad_0, pad_type = x_38_pad_type_0, strides = var_1110, weight = nlp_net_default_encoder_transformer_layers_3_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_55_cast_fp16)[name = string("x_38_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_25_cast_fp16 = add(x = x_38_cast_fp16, y = inputs_17_cast_fp16)[name = string("attn_25_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_8_cast_fp16 = add(x = inputs1_1_cast_fp16, y = attn_25_cast_fp16)[name = string("inputs0_8_cast_fp16")];
            tensor<int32, [1]> input_57_axes_0 = const()[name = string("input_57_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_57_gamma_0_to_fp16 = const()[name = string("input_57_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36461184)))];
            tensor<fp16, [512]> input_57_beta_0_to_fp16 = const()[name = string("input_57_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36462272)))];
            fp16 var_1125_to_fp16 = const()[name = string("op_1125_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = input_57_beta_0_to_fp16, epsilon = var_1125_to_fp16, gamma = input_57_gamma_0_to_fp16, x = inputs0_8_cast_fp16)[name = string("input_57_cast_fp16")];
            tensor<int32, [2]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1141 = const()[name = string("op_1141"), val = tensor<int32, [2]>([1, 1])];
            string x_40_pad_type_0 = const()[name = string("x_40_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_40_pad_0 = const()[name = string("x_40_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36467520))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(36463360))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37520320))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37516160))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_40_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1141, groups = var_46, pad = x_40_pad_0, pad_type = x_40_pad_type_0, strides = var_1139, weight = nlp_net_default_encoder_transformer_layers_3_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_57_cast_fp16)[name = string("x_40_cast_fp16")];
            fp16 var_1144_to_fp16 = const()[name = string("op_1144_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1145_cast_fp16 = mul(x = x_40_cast_fp16, y = var_1144_to_fp16)[name = string("op_1145_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1146_cast_fp16 = sigmoid(x = var_1145_cast_fp16)[name = string("op_1146_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_59_cast_fp16 = mul(x = x_40_cast_fp16, y = var_1146_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<int32, [2]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1152 = const()[name = string("op_1152"), val = tensor<int32, [2]>([1, 1])];
            string input0_15_pad_type_0 = const()[name = string("input0_15_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_15_pad_0 = const()[name = string("input0_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37523520))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(37522432))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38572160)))];
            tensor<fp16, [1, 512, 1, 256]> input0_15_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_bias_to_fp16, dilations = var_1152, groups = var_46, pad = input0_15_pad_0, pad_type = input0_15_pad_type_0, strides = var_1150, weight = nlp_net_default_encoder_transformer_layers_3_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_59_cast_fp16)[name = string("input0_15_cast_fp16")];
            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_61_gamma_0_to_fp16 = const()[name = string("input_61_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52216064)))];
            tensor<fp16, [512]> input_61_beta_0_to_fp16 = const()[name = string("input_61_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52217152)))];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = input_61_beta_0_to_fp16, epsilon = var_1167_to_fp16, gamma = input_61_gamma_0_to_fp16, x = input0_15_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<int32, [2]> var_1179 = const()[name = string("op_1179"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1181 = const()[name = string("op_1181"), val = tensor<int32, [2]>([1, 1])];
            string x_42_pad_type_0 = const()[name = string("x_42_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_42_pad_0 = const()[name = string("x_42_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52218560))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52218240))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52284160)))];
            tensor<fp16, [1, 128, 1, 256]> x_42_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1181, groups = var_46, pad = x_42_pad_0, pad_type = x_42_pad_type_0, strides = var_1179, weight = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_61_cast_fp16)[name = string("x_42_cast_fp16")];
            fp16 var_1184_to_fp16 = const()[name = string("op_1184_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1185_cast_fp16 = mul(x = x_42_cast_fp16, y = var_1184_to_fp16)[name = string("op_1185_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1186_cast_fp16 = sigmoid(x = var_1185_cast_fp16)[name = string("op_1186_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_63_cast_fp16 = mul(x = x_42_cast_fp16, y = var_1186_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<int32, [2]> var_1190 = const()[name = string("op_1190"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1192 = const()[name = string("op_1192"), val = tensor<int32, [2]>([1, 1])];
            string x_44_pad_type_0 = const()[name = string("x_44_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_44_pad_0 = const()[name = string("x_44_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52285568))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52284480))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52351168)))];
            tensor<fp16, [1, 512, 1, 256]> x_44_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1192, groups = var_46, pad = x_44_pad_0, pad_type = x_44_pad_type_0, strides = var_1190, weight = nlp_net_default_encoder_transformer_layers_3_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_63_cast_fp16)[name = string("x_44_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_8_cast_fp16 = add(x = x_44_cast_fp16, y = input0_15_cast_fp16)[name = string("f_8_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_8_cast_fp16 = add(x = f_8_cast_fp16, y = inputs0_8_cast_fp16)[name = string("x1_8_cast_fp16")];
            fp16 var_1197_to_fp16 = const()[name = string("op_1197_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1198_cast_fp16 = mul(x = inputs1_1_cast_fp16, y = var_1197_to_fp16)[name = string("op_1198_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs2_1_cast_fp16 = add(x = var_1198_cast_fp16, y = x1_8_cast_fp16)[name = string("inputs2_1_cast_fp16")];
            tensor<int32, [1]> k_19_axes_0 = const()[name = string("k_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_19_gamma_0_to_fp16 = const()[name = string("k_19_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38573248)))];
            tensor<fp16, [512]> k_19_beta_0_to_fp16 = const()[name = string("k_19_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38574336)))];
            fp16 var_1216_to_fp16 = const()[name = string("op_1216_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_19_cast_fp16 = layer_norm(axes = k_19_axes_0, beta = k_19_beta_0_to_fp16, epsilon = var_1216_to_fp16, gamma = k_19_gamma_0_to_fp16, x = inputs2_1_cast_fp16)[name = string("k_19_cast_fp16")];
            tensor<int32, [2]> var_1235 = const()[name = string("op_1235"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1237 = const()[name = string("op_1237"), val = tensor<int32, [2]>([1, 1])];
            string var_1239_pad_type_0 = const()[name = string("op_1239_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1239_pad_0 = const()[name = string("op_1239_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38576512))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38575424))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38838720)))];
            tensor<fp16, [1, 512, 1, 256]> var_1239_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_bias_to_fp16, dilations = var_1237, groups = var_46, pad = var_1239_pad_0, pad_type = var_1239_pad_type_0, strides = var_1235, weight = nlp_net_default_encoder_transformer_layers_4_attn_q_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_1239_cast_fp16")];
            tensor<int32, [2]> var_1242 = const()[name = string("op_1242"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1244 = const()[name = string("op_1244"), val = tensor<int32, [2]>([1, 1])];
            string k_21_pad_type_0 = const()[name = string("k_21_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_21_pad_0 = const()[name = string("k_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38840896))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(38839808))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39103104)))];
            tensor<fp16, [1, 512, 1, 256]> k_21_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_bias_to_fp16, dilations = var_1244, groups = var_46, pad = k_21_pad_0, pad_type = k_21_pad_type_0, strides = var_1242, weight = nlp_net_default_encoder_transformer_layers_4_attn_k_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<int32, [2]> var_1249 = const()[name = string("op_1249"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1251 = const()[name = string("op_1251"), val = tensor<int32, [2]>([1, 1])];
            string var_1253_pad_type_0 = const()[name = string("op_1253_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1253_pad_0 = const()[name = string("op_1253_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39105280))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39104192))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39367488)))];
            tensor<fp16, [1, 512, 1, 256]> var_1253_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_bias_to_fp16, dilations = var_1251, groups = var_46, pad = var_1253_pad_0, pad_type = var_1253_pad_type_0, strides = var_1249, weight = nlp_net_default_encoder_transformer_layers_4_attn_v_proj_weight_to_fp16_affine_quantized, x = k_19_cast_fp16)[name = string("op_1253_cast_fp16")];
            tensor<int32, [8]> tile_16 = const()[name = string("tile_16"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1254_axis_0 = const()[name = string("op_1254_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1254_cast_fp16_7 = split(axis = var_1254_axis_0, split_sizes = tile_16, x = var_1239_cast_fp16)[name = string("op_1254_cast_fp16")];
            tensor<int32, [4]> var_1263_perm_0 = const()[name = string("op_1263_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_17 = const()[name = string("tile_17"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1264_axis_0 = const()[name = string("op_1264_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_4 = transpose(perm = var_1263_perm_0, x = k_21_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1264_cast_fp16_7 = split(axis = var_1264_axis_0, split_sizes = tile_17, x = transpose_4)[name = string("op_1264_cast_fp16")];
            tensor<int32, [8]> tile_18 = const()[name = string("tile_18"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1273_axis_0 = const()[name = string("op_1273_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1273_cast_fp16_7 = split(axis = var_1273_axis_0, split_sizes = tile_18, x = var_1253_cast_fp16)[name = string("op_1273_cast_fp16")];
            string var_1283_equation_0 = const()[name = string("op_1283_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1283_cast_fp16 = einsum(equation = var_1283_equation_0, values = (var_1264_cast_fp16_0, var_1254_cast_fp16_0))[name = string("op_1283_cast_fp16")];
            fp16 var_1284_to_fp16 = const()[name = string("op_1284_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1285_cast_fp16 = mul(x = var_1283_cast_fp16, y = var_1284_to_fp16)[name = string("op_1285_cast_fp16")];
            string var_1287_equation_0 = const()[name = string("op_1287_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1287_cast_fp16 = einsum(equation = var_1287_equation_0, values = (var_1264_cast_fp16_1, var_1254_cast_fp16_1))[name = string("op_1287_cast_fp16")];
            fp16 var_1288_to_fp16 = const()[name = string("op_1288_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1289_cast_fp16 = mul(x = var_1287_cast_fp16, y = var_1288_to_fp16)[name = string("op_1289_cast_fp16")];
            string var_1291_equation_0 = const()[name = string("op_1291_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1291_cast_fp16 = einsum(equation = var_1291_equation_0, values = (var_1264_cast_fp16_2, var_1254_cast_fp16_2))[name = string("op_1291_cast_fp16")];
            fp16 var_1292_to_fp16 = const()[name = string("op_1292_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1293_cast_fp16 = mul(x = var_1291_cast_fp16, y = var_1292_to_fp16)[name = string("op_1293_cast_fp16")];
            string var_1295_equation_0 = const()[name = string("op_1295_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1295_cast_fp16 = einsum(equation = var_1295_equation_0, values = (var_1264_cast_fp16_3, var_1254_cast_fp16_3))[name = string("op_1295_cast_fp16")];
            fp16 var_1296_to_fp16 = const()[name = string("op_1296_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = string("op_1297_cast_fp16")];
            string var_1299_equation_0 = const()[name = string("op_1299_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1299_cast_fp16 = einsum(equation = var_1299_equation_0, values = (var_1264_cast_fp16_4, var_1254_cast_fp16_4))[name = string("op_1299_cast_fp16")];
            fp16 var_1300_to_fp16 = const()[name = string("op_1300_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1301_cast_fp16 = mul(x = var_1299_cast_fp16, y = var_1300_to_fp16)[name = string("op_1301_cast_fp16")];
            string var_1303_equation_0 = const()[name = string("op_1303_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1303_cast_fp16 = einsum(equation = var_1303_equation_0, values = (var_1264_cast_fp16_5, var_1254_cast_fp16_5))[name = string("op_1303_cast_fp16")];
            fp16 var_1304_to_fp16 = const()[name = string("op_1304_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1305_cast_fp16 = mul(x = var_1303_cast_fp16, y = var_1304_to_fp16)[name = string("op_1305_cast_fp16")];
            string var_1307_equation_0 = const()[name = string("op_1307_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1307_cast_fp16 = einsum(equation = var_1307_equation_0, values = (var_1264_cast_fp16_6, var_1254_cast_fp16_6))[name = string("op_1307_cast_fp16")];
            fp16 var_1308_to_fp16 = const()[name = string("op_1308_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1309_cast_fp16 = mul(x = var_1307_cast_fp16, y = var_1308_to_fp16)[name = string("op_1309_cast_fp16")];
            string var_1311_equation_0 = const()[name = string("op_1311_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1311_cast_fp16 = einsum(equation = var_1311_equation_0, values = (var_1264_cast_fp16_7, var_1254_cast_fp16_7))[name = string("op_1311_cast_fp16")];
            fp16 var_1312_to_fp16 = const()[name = string("op_1312_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1313_cast_fp16 = mul(x = var_1311_cast_fp16, y = var_1312_to_fp16)[name = string("op_1313_cast_fp16")];
            bool attn_weights_10_interleave_0 = const()[name = string("attn_weights_10_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_10_cast_fp16 = concat(axis = var_47, interleave = attn_weights_10_interleave_0, values = (var_1285_cast_fp16, var_1289_cast_fp16, var_1293_cast_fp16, var_1297_cast_fp16, var_1301_cast_fp16, var_1305_cast_fp16, var_1309_cast_fp16, var_1313_cast_fp16))[name = string("attn_weights_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_10_cast_fp16 = add(x = attn_weights_10_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_10_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_65_cast_fp16 = softmax(axis = var_46, x = attn_weights0_10_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<int32, [8]> tile_19 = const()[name = string("tile_19"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1319_axis_0 = const()[name = string("op_1319_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1319_cast_fp16_7 = split(axis = var_1319_axis_0, split_sizes = tile_19, x = input_65_cast_fp16)[name = string("op_1319_cast_fp16")];
            string var_1329_equation_0 = const()[name = string("op_1329_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1329_cast_fp16 = einsum(equation = var_1329_equation_0, values = (var_1273_cast_fp16_0, var_1319_cast_fp16_0))[name = string("op_1329_cast_fp16")];
            string var_1331_equation_0 = const()[name = string("op_1331_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1331_cast_fp16 = einsum(equation = var_1331_equation_0, values = (var_1273_cast_fp16_1, var_1319_cast_fp16_1))[name = string("op_1331_cast_fp16")];
            string var_1333_equation_0 = const()[name = string("op_1333_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1333_cast_fp16 = einsum(equation = var_1333_equation_0, values = (var_1273_cast_fp16_2, var_1319_cast_fp16_2))[name = string("op_1333_cast_fp16")];
            string var_1335_equation_0 = const()[name = string("op_1335_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1335_cast_fp16 = einsum(equation = var_1335_equation_0, values = (var_1273_cast_fp16_3, var_1319_cast_fp16_3))[name = string("op_1335_cast_fp16")];
            string var_1337_equation_0 = const()[name = string("op_1337_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1337_cast_fp16 = einsum(equation = var_1337_equation_0, values = (var_1273_cast_fp16_4, var_1319_cast_fp16_4))[name = string("op_1337_cast_fp16")];
            string var_1339_equation_0 = const()[name = string("op_1339_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1339_cast_fp16 = einsum(equation = var_1339_equation_0, values = (var_1273_cast_fp16_5, var_1319_cast_fp16_5))[name = string("op_1339_cast_fp16")];
            string var_1341_equation_0 = const()[name = string("op_1341_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1341_cast_fp16 = einsum(equation = var_1341_equation_0, values = (var_1273_cast_fp16_6, var_1319_cast_fp16_6))[name = string("op_1341_cast_fp16")];
            string var_1343_equation_0 = const()[name = string("op_1343_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1343_cast_fp16 = einsum(equation = var_1343_equation_0, values = (var_1273_cast_fp16_7, var_1319_cast_fp16_7))[name = string("op_1343_cast_fp16")];
            bool attn_29_interleave_0 = const()[name = string("attn_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_29_cast_fp16 = concat(axis = var_46, interleave = attn_29_interleave_0, values = (var_1329_cast_fp16, var_1331_cast_fp16, var_1333_cast_fp16, var_1335_cast_fp16, var_1337_cast_fp16, var_1339_cast_fp16, var_1341_cast_fp16, var_1343_cast_fp16))[name = string("attn_29_cast_fp16")];
            tensor<int32, [2]> var_1351 = const()[name = string("op_1351"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1353 = const()[name = string("op_1353"), val = tensor<int32, [2]>([1, 1])];
            string inputs_21_pad_type_0 = const()[name = string("inputs_21_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_21_pad_0 = const()[name = string("inputs_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39369664))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39368576))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39631872)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_21_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_bias_to_fp16, dilations = var_1353, groups = var_46, pad = inputs_21_pad_0, pad_type = inputs_21_pad_type_0, strides = var_1351, weight = nlp_net_default_encoder_transformer_layers_4_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_29_cast_fp16)[name = string("inputs_21_cast_fp16")];
            tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_67_gamma_0_to_fp16 = const()[name = string("input_67_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52352256)))];
            tensor<fp16, [512]> input_67_beta_0_to_fp16 = const()[name = string("input_67_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52353344)))];
            fp16 var_1367_to_fp16 = const()[name = string("op_1367_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = input_67_beta_0_to_fp16, epsilon = var_1367_to_fp16, gamma = input_67_gamma_0_to_fp16, x = inputs_21_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<int32, [2]> var_1379 = const()[name = string("op_1379"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1381 = const()[name = string("op_1381"), val = tensor<int32, [2]>([1, 1])];
            string x_46_pad_type_0 = const()[name = string("x_46_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_46_pad_0 = const()[name = string("x_46_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52354752))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52354432))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52420352)))];
            tensor<fp16, [1, 128, 1, 256]> x_46_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1381, groups = var_46, pad = x_46_pad_0, pad_type = x_46_pad_type_0, strides = var_1379, weight = nlp_net_default_encoder_transformer_layers_4_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_67_cast_fp16)[name = string("x_46_cast_fp16")];
            fp16 var_1384_to_fp16 = const()[name = string("op_1384_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1385_cast_fp16 = mul(x = x_46_cast_fp16, y = var_1384_to_fp16)[name = string("op_1385_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1386_cast_fp16 = sigmoid(x = var_1385_cast_fp16)[name = string("op_1386_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_69_cast_fp16 = mul(x = x_46_cast_fp16, y = var_1386_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<int32, [2]> var_1390 = const()[name = string("op_1390"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1392 = const()[name = string("op_1392"), val = tensor<int32, [2]>([1, 1])];
            string x_48_pad_type_0 = const()[name = string("x_48_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_48_pad_0 = const()[name = string("x_48_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52421760))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52420672))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52487360)))];
            tensor<fp16, [1, 512, 1, 256]> x_48_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1392, groups = var_46, pad = x_48_pad_0, pad_type = x_48_pad_type_0, strides = var_1390, weight = nlp_net_default_encoder_transformer_layers_4_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_69_cast_fp16)[name = string("x_48_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_31_cast_fp16 = add(x = x_48_cast_fp16, y = inputs_21_cast_fp16)[name = string("attn_31_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_10_cast_fp16 = add(x = inputs2_1_cast_fp16, y = attn_31_cast_fp16)[name = string("inputs0_10_cast_fp16")];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_71_gamma_0_to_fp16 = const()[name = string("input_71_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39632960)))];
            tensor<fp16, [512]> input_71_beta_0_to_fp16 = const()[name = string("input_71_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39634048)))];
            fp16 var_1405_to_fp16 = const()[name = string("op_1405_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = input_71_beta_0_to_fp16, epsilon = var_1405_to_fp16, gamma = input_71_gamma_0_to_fp16, x = inputs0_10_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<int32, [2]> var_1419 = const()[name = string("op_1419"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1421 = const()[name = string("op_1421"), val = tensor<int32, [2]>([1, 1])];
            string x_50_pad_type_0 = const()[name = string("x_50_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_50_pad_0 = const()[name = string("x_50_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39639296))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(39635136))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40692096))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40687936))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_50_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1421, groups = var_46, pad = x_50_pad_0, pad_type = x_50_pad_type_0, strides = var_1419, weight = nlp_net_default_encoder_transformer_layers_4_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_71_cast_fp16)[name = string("x_50_cast_fp16")];
            fp16 var_1424_to_fp16 = const()[name = string("op_1424_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1425_cast_fp16 = mul(x = x_50_cast_fp16, y = var_1424_to_fp16)[name = string("op_1425_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1426_cast_fp16 = sigmoid(x = var_1425_cast_fp16)[name = string("op_1426_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_73_cast_fp16 = mul(x = x_50_cast_fp16, y = var_1426_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<int32, [2]> var_1430 = const()[name = string("op_1430"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1432 = const()[name = string("op_1432"), val = tensor<int32, [2]>([1, 1])];
            string input0_19_pad_type_0 = const()[name = string("input0_19_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_19_pad_0 = const()[name = string("input0_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40695296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(40694208))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41743936)))];
            tensor<fp16, [1, 512, 1, 256]> input0_19_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_bias_to_fp16, dilations = var_1432, groups = var_46, pad = input0_19_pad_0, pad_type = input0_19_pad_type_0, strides = var_1430, weight = nlp_net_default_encoder_transformer_layers_4_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_73_cast_fp16)[name = string("input0_19_cast_fp16")];
            tensor<int32, [1]> input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_75_gamma_0_to_fp16 = const()[name = string("input_75_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52488448)))];
            tensor<fp16, [512]> input_75_beta_0_to_fp16 = const()[name = string("input_75_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52489536)))];
            fp16 var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = input_75_beta_0_to_fp16, epsilon = var_1447_to_fp16, gamma = input_75_gamma_0_to_fp16, x = input0_19_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<int32, [2]> var_1459 = const()[name = string("op_1459"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1461 = const()[name = string("op_1461"), val = tensor<int32, [2]>([1, 1])];
            string x_52_pad_type_0 = const()[name = string("x_52_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_52_pad_0 = const()[name = string("x_52_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52490944))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52490624))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52556544)))];
            tensor<fp16, [1, 128, 1, 256]> x_52_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1461, groups = var_46, pad = x_52_pad_0, pad_type = x_52_pad_type_0, strides = var_1459, weight = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_75_cast_fp16)[name = string("x_52_cast_fp16")];
            fp16 var_1464_to_fp16 = const()[name = string("op_1464_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1465_cast_fp16 = mul(x = x_52_cast_fp16, y = var_1464_to_fp16)[name = string("op_1465_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1466_cast_fp16 = sigmoid(x = var_1465_cast_fp16)[name = string("op_1466_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_77_cast_fp16 = mul(x = x_52_cast_fp16, y = var_1466_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<int32, [2]> var_1470 = const()[name = string("op_1470"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1472 = const()[name = string("op_1472"), val = tensor<int32, [2]>([1, 1])];
            string x_54_pad_type_0 = const()[name = string("x_54_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_54_pad_0 = const()[name = string("x_54_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52557952))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52556864))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52623552)))];
            tensor<fp16, [1, 512, 1, 256]> x_54_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1472, groups = var_46, pad = x_54_pad_0, pad_type = x_54_pad_type_0, strides = var_1470, weight = nlp_net_default_encoder_transformer_layers_4_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_77_cast_fp16)[name = string("x_54_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_10_cast_fp16 = add(x = x_54_cast_fp16, y = input0_19_cast_fp16)[name = string("f_10_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_10_cast_fp16 = add(x = f_10_cast_fp16, y = inputs0_10_cast_fp16)[name = string("x1_10_cast_fp16")];
            fp16 var_1477_to_fp16 = const()[name = string("op_1477_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1478_cast_fp16 = mul(x = inputs2_1_cast_fp16, y = var_1477_to_fp16)[name = string("op_1478_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs3_1_cast_fp16 = add(x = var_1478_cast_fp16, y = x1_10_cast_fp16)[name = string("inputs3_1_cast_fp16")];
            tensor<int32, [1]> k_23_axes_0 = const()[name = string("k_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_23_gamma_0_to_fp16 = const()[name = string("k_23_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41745024)))];
            tensor<fp16, [512]> k_23_beta_0_to_fp16 = const()[name = string("k_23_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41746112)))];
            fp16 var_1496_to_fp16 = const()[name = string("op_1496_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_23_cast_fp16 = layer_norm(axes = k_23_axes_0, beta = k_23_beta_0_to_fp16, epsilon = var_1496_to_fp16, gamma = k_23_gamma_0_to_fp16, x = inputs3_1_cast_fp16)[name = string("k_23_cast_fp16")];
            tensor<int32, [2]> var_1515 = const()[name = string("op_1515"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [2]>([1, 1])];
            string var_1519_pad_type_0 = const()[name = string("op_1519_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1519_pad_0 = const()[name = string("op_1519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41748288))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(41747200))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42010496)))];
            tensor<fp16, [1, 512, 1, 256]> var_1519_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_bias_to_fp16, dilations = var_1517, groups = var_46, pad = var_1519_pad_0, pad_type = var_1519_pad_type_0, strides = var_1515, weight = nlp_net_default_encoder_transformer_layers_5_attn_q_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1519_cast_fp16")];
            tensor<int32, [2]> var_1522 = const()[name = string("op_1522"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [2]>([1, 1])];
            string k_25_pad_type_0 = const()[name = string("k_25_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_25_pad_0 = const()[name = string("k_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42012672))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42011584))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42274880)))];
            tensor<fp16, [1, 512, 1, 256]> k_25_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_bias_to_fp16, dilations = var_1524, groups = var_46, pad = k_25_pad_0, pad_type = k_25_pad_type_0, strides = var_1522, weight = nlp_net_default_encoder_transformer_layers_5_attn_k_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<int32, [2]> var_1529 = const()[name = string("op_1529"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [2]>([1, 1])];
            string var_1533_pad_type_0 = const()[name = string("op_1533_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1533_pad_0 = const()[name = string("op_1533_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42277056))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42275968))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42539264)))];
            tensor<fp16, [1, 512, 1, 256]> var_1533_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_bias_to_fp16, dilations = var_1531, groups = var_46, pad = var_1533_pad_0, pad_type = var_1533_pad_type_0, strides = var_1529, weight = nlp_net_default_encoder_transformer_layers_5_attn_v_proj_weight_to_fp16_affine_quantized, x = k_23_cast_fp16)[name = string("op_1533_cast_fp16")];
            tensor<int32, [8]> tile_20 = const()[name = string("tile_20"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1534_axis_0 = const()[name = string("op_1534_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1534_cast_fp16_7 = split(axis = var_1534_axis_0, split_sizes = tile_20, x = var_1519_cast_fp16)[name = string("op_1534_cast_fp16")];
            tensor<int32, [4]> var_1543_perm_0 = const()[name = string("op_1543_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_21 = const()[name = string("tile_21"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_3 = transpose(perm = var_1543_perm_0, x = k_25_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1544_cast_fp16_7 = split(axis = var_1544_axis_0, split_sizes = tile_21, x = transpose_3)[name = string("op_1544_cast_fp16")];
            tensor<int32, [8]> tile_22 = const()[name = string("tile_22"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1553_axis_0 = const()[name = string("op_1553_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1553_cast_fp16_7 = split(axis = var_1553_axis_0, split_sizes = tile_22, x = var_1533_cast_fp16)[name = string("op_1553_cast_fp16")];
            string var_1563_equation_0 = const()[name = string("op_1563_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1563_cast_fp16 = einsum(equation = var_1563_equation_0, values = (var_1544_cast_fp16_0, var_1534_cast_fp16_0))[name = string("op_1563_cast_fp16")];
            fp16 var_1564_to_fp16 = const()[name = string("op_1564_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1565_cast_fp16 = mul(x = var_1563_cast_fp16, y = var_1564_to_fp16)[name = string("op_1565_cast_fp16")];
            string var_1567_equation_0 = const()[name = string("op_1567_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1567_cast_fp16 = einsum(equation = var_1567_equation_0, values = (var_1544_cast_fp16_1, var_1534_cast_fp16_1))[name = string("op_1567_cast_fp16")];
            fp16 var_1568_to_fp16 = const()[name = string("op_1568_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1569_cast_fp16 = mul(x = var_1567_cast_fp16, y = var_1568_to_fp16)[name = string("op_1569_cast_fp16")];
            string var_1571_equation_0 = const()[name = string("op_1571_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1571_cast_fp16 = einsum(equation = var_1571_equation_0, values = (var_1544_cast_fp16_2, var_1534_cast_fp16_2))[name = string("op_1571_cast_fp16")];
            fp16 var_1572_to_fp16 = const()[name = string("op_1572_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1573_cast_fp16 = mul(x = var_1571_cast_fp16, y = var_1572_to_fp16)[name = string("op_1573_cast_fp16")];
            string var_1575_equation_0 = const()[name = string("op_1575_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1575_cast_fp16 = einsum(equation = var_1575_equation_0, values = (var_1544_cast_fp16_3, var_1534_cast_fp16_3))[name = string("op_1575_cast_fp16")];
            fp16 var_1576_to_fp16 = const()[name = string("op_1576_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1577_cast_fp16 = mul(x = var_1575_cast_fp16, y = var_1576_to_fp16)[name = string("op_1577_cast_fp16")];
            string var_1579_equation_0 = const()[name = string("op_1579_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1579_cast_fp16 = einsum(equation = var_1579_equation_0, values = (var_1544_cast_fp16_4, var_1534_cast_fp16_4))[name = string("op_1579_cast_fp16")];
            fp16 var_1580_to_fp16 = const()[name = string("op_1580_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1581_cast_fp16 = mul(x = var_1579_cast_fp16, y = var_1580_to_fp16)[name = string("op_1581_cast_fp16")];
            string var_1583_equation_0 = const()[name = string("op_1583_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1583_cast_fp16 = einsum(equation = var_1583_equation_0, values = (var_1544_cast_fp16_5, var_1534_cast_fp16_5))[name = string("op_1583_cast_fp16")];
            fp16 var_1584_to_fp16 = const()[name = string("op_1584_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1585_cast_fp16 = mul(x = var_1583_cast_fp16, y = var_1584_to_fp16)[name = string("op_1585_cast_fp16")];
            string var_1587_equation_0 = const()[name = string("op_1587_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1587_cast_fp16 = einsum(equation = var_1587_equation_0, values = (var_1544_cast_fp16_6, var_1534_cast_fp16_6))[name = string("op_1587_cast_fp16")];
            fp16 var_1588_to_fp16 = const()[name = string("op_1588_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1589_cast_fp16 = mul(x = var_1587_cast_fp16, y = var_1588_to_fp16)[name = string("op_1589_cast_fp16")];
            string var_1591_equation_0 = const()[name = string("op_1591_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1591_cast_fp16 = einsum(equation = var_1591_equation_0, values = (var_1544_cast_fp16_7, var_1534_cast_fp16_7))[name = string("op_1591_cast_fp16")];
            fp16 var_1592_to_fp16 = const()[name = string("op_1592_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1593_cast_fp16 = mul(x = var_1591_cast_fp16, y = var_1592_to_fp16)[name = string("op_1593_cast_fp16")];
            bool attn_weights_12_interleave_0 = const()[name = string("attn_weights_12_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_12_cast_fp16 = concat(axis = var_47, interleave = attn_weights_12_interleave_0, values = (var_1565_cast_fp16, var_1569_cast_fp16, var_1573_cast_fp16, var_1577_cast_fp16, var_1581_cast_fp16, var_1585_cast_fp16, var_1589_cast_fp16, var_1593_cast_fp16))[name = string("attn_weights_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_12_cast_fp16 = add(x = attn_weights_12_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_12_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_79_cast_fp16 = softmax(axis = var_46, x = attn_weights0_12_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<int32, [8]> tile_23 = const()[name = string("tile_23"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1599_axis_0 = const()[name = string("op_1599_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1599_cast_fp16_7 = split(axis = var_1599_axis_0, split_sizes = tile_23, x = input_79_cast_fp16)[name = string("op_1599_cast_fp16")];
            string var_1609_equation_0 = const()[name = string("op_1609_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1609_cast_fp16 = einsum(equation = var_1609_equation_0, values = (var_1553_cast_fp16_0, var_1599_cast_fp16_0))[name = string("op_1609_cast_fp16")];
            string var_1611_equation_0 = const()[name = string("op_1611_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1611_cast_fp16 = einsum(equation = var_1611_equation_0, values = (var_1553_cast_fp16_1, var_1599_cast_fp16_1))[name = string("op_1611_cast_fp16")];
            string var_1613_equation_0 = const()[name = string("op_1613_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1613_cast_fp16 = einsum(equation = var_1613_equation_0, values = (var_1553_cast_fp16_2, var_1599_cast_fp16_2))[name = string("op_1613_cast_fp16")];
            string var_1615_equation_0 = const()[name = string("op_1615_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1615_cast_fp16 = einsum(equation = var_1615_equation_0, values = (var_1553_cast_fp16_3, var_1599_cast_fp16_3))[name = string("op_1615_cast_fp16")];
            string var_1617_equation_0 = const()[name = string("op_1617_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1617_cast_fp16 = einsum(equation = var_1617_equation_0, values = (var_1553_cast_fp16_4, var_1599_cast_fp16_4))[name = string("op_1617_cast_fp16")];
            string var_1619_equation_0 = const()[name = string("op_1619_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1619_cast_fp16 = einsum(equation = var_1619_equation_0, values = (var_1553_cast_fp16_5, var_1599_cast_fp16_5))[name = string("op_1619_cast_fp16")];
            string var_1621_equation_0 = const()[name = string("op_1621_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1621_cast_fp16 = einsum(equation = var_1621_equation_0, values = (var_1553_cast_fp16_6, var_1599_cast_fp16_6))[name = string("op_1621_cast_fp16")];
            string var_1623_equation_0 = const()[name = string("op_1623_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1623_cast_fp16 = einsum(equation = var_1623_equation_0, values = (var_1553_cast_fp16_7, var_1599_cast_fp16_7))[name = string("op_1623_cast_fp16")];
            bool attn_35_interleave_0 = const()[name = string("attn_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_35_cast_fp16 = concat(axis = var_46, interleave = attn_35_interleave_0, values = (var_1609_cast_fp16, var_1611_cast_fp16, var_1613_cast_fp16, var_1615_cast_fp16, var_1617_cast_fp16, var_1619_cast_fp16, var_1621_cast_fp16, var_1623_cast_fp16))[name = string("attn_35_cast_fp16")];
            tensor<int32, [2]> var_1631 = const()[name = string("op_1631"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1633 = const()[name = string("op_1633"), val = tensor<int32, [2]>([1, 1])];
            string inputs_25_pad_type_0 = const()[name = string("inputs_25_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_25_pad_0 = const()[name = string("inputs_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42541440))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42540352))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42803648)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_25_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_bias_to_fp16, dilations = var_1633, groups = var_46, pad = inputs_25_pad_0, pad_type = inputs_25_pad_type_0, strides = var_1631, weight = nlp_net_default_encoder_transformer_layers_5_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_35_cast_fp16)[name = string("inputs_25_cast_fp16")];
            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_81_gamma_0_to_fp16 = const()[name = string("input_81_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52624640)))];
            tensor<fp16, [512]> input_81_beta_0_to_fp16 = const()[name = string("input_81_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52625728)))];
            fp16 var_1647_to_fp16 = const()[name = string("op_1647_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = input_81_beta_0_to_fp16, epsilon = var_1647_to_fp16, gamma = input_81_gamma_0_to_fp16, x = inputs_25_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<int32, [2]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1661 = const()[name = string("op_1661"), val = tensor<int32, [2]>([1, 1])];
            string x_56_pad_type_0 = const()[name = string("x_56_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_56_pad_0 = const()[name = string("x_56_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52627136))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52626816))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52692736)))];
            tensor<fp16, [1, 128, 1, 256]> x_56_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1661, groups = var_46, pad = x_56_pad_0, pad_type = x_56_pad_type_0, strides = var_1659, weight = nlp_net_default_encoder_transformer_layers_5_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_81_cast_fp16)[name = string("x_56_cast_fp16")];
            fp16 var_1664_to_fp16 = const()[name = string("op_1664_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1665_cast_fp16 = mul(x = x_56_cast_fp16, y = var_1664_to_fp16)[name = string("op_1665_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1666_cast_fp16 = sigmoid(x = var_1665_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_83_cast_fp16 = mul(x = x_56_cast_fp16, y = var_1666_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<int32, [2]> var_1670 = const()[name = string("op_1670"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1672 = const()[name = string("op_1672"), val = tensor<int32, [2]>([1, 1])];
            string x_58_pad_type_0 = const()[name = string("x_58_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_58_pad_0 = const()[name = string("x_58_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52694144))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52693056))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52759744)))];
            tensor<fp16, [1, 512, 1, 256]> x_58_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1672, groups = var_46, pad = x_58_pad_0, pad_type = x_58_pad_type_0, strides = var_1670, weight = nlp_net_default_encoder_transformer_layers_5_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_83_cast_fp16)[name = string("x_58_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_37_cast_fp16 = add(x = x_58_cast_fp16, y = inputs_25_cast_fp16)[name = string("attn_37_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_12_cast_fp16 = add(x = inputs3_1_cast_fp16, y = attn_37_cast_fp16)[name = string("inputs0_12_cast_fp16")];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_85_gamma_0_to_fp16 = const()[name = string("input_85_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42804736)))];
            tensor<fp16, [512]> input_85_beta_0_to_fp16 = const()[name = string("input_85_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42805824)))];
            fp16 var_1685_to_fp16 = const()[name = string("op_1685_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_85_cast_fp16 = layer_norm(axes = input_85_axes_0, beta = input_85_beta_0_to_fp16, epsilon = var_1685_to_fp16, gamma = input_85_gamma_0_to_fp16, x = inputs0_12_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<int32, [2]> var_1699 = const()[name = string("op_1699"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1701 = const()[name = string("op_1701"), val = tensor<int32, [2]>([1, 1])];
            string x_60_pad_type_0 = const()[name = string("x_60_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_60_pad_0 = const()[name = string("x_60_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42811072))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(42806912))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43863872))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43859712))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_60_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1701, groups = var_46, pad = x_60_pad_0, pad_type = x_60_pad_type_0, strides = var_1699, weight = nlp_net_default_encoder_transformer_layers_5_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_85_cast_fp16)[name = string("x_60_cast_fp16")];
            fp16 var_1704_to_fp16 = const()[name = string("op_1704_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1705_cast_fp16 = mul(x = x_60_cast_fp16, y = var_1704_to_fp16)[name = string("op_1705_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1706_cast_fp16 = sigmoid(x = var_1705_cast_fp16)[name = string("op_1706_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_87_cast_fp16 = mul(x = x_60_cast_fp16, y = var_1706_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<int32, [2]> var_1710 = const()[name = string("op_1710"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1712 = const()[name = string("op_1712"), val = tensor<int32, [2]>([1, 1])];
            string input0_23_pad_type_0 = const()[name = string("input0_23_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_23_pad_0 = const()[name = string("input0_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43867072))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(43865984))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44915712)))];
            tensor<fp16, [1, 512, 1, 256]> input0_23_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_bias_to_fp16, dilations = var_1712, groups = var_46, pad = input0_23_pad_0, pad_type = input0_23_pad_type_0, strides = var_1710, weight = nlp_net_default_encoder_transformer_layers_5_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_87_cast_fp16)[name = string("input0_23_cast_fp16")];
            tensor<int32, [1]> input_89_axes_0 = const()[name = string("input_89_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_89_gamma_0_to_fp16 = const()[name = string("input_89_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52760832)))];
            tensor<fp16, [512]> input_89_beta_0_to_fp16 = const()[name = string("input_89_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52761920)))];
            fp16 var_1727_to_fp16 = const()[name = string("op_1727_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = input_89_beta_0_to_fp16, epsilon = var_1727_to_fp16, gamma = input_89_gamma_0_to_fp16, x = input0_23_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<int32, [2]> var_1739 = const()[name = string("op_1739"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1741 = const()[name = string("op_1741"), val = tensor<int32, [2]>([1, 1])];
            string x_62_pad_type_0 = const()[name = string("x_62_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_62_pad_0 = const()[name = string("x_62_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52763328))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52763008))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52828928)))];
            tensor<fp16, [1, 128, 1, 256]> x_62_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_1741, groups = var_46, pad = x_62_pad_0, pad_type = x_62_pad_type_0, strides = var_1739, weight = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_89_cast_fp16)[name = string("x_62_cast_fp16")];
            fp16 var_1744_to_fp16 = const()[name = string("op_1744_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1745_cast_fp16 = mul(x = x_62_cast_fp16, y = var_1744_to_fp16)[name = string("op_1745_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1746_cast_fp16 = sigmoid(x = var_1745_cast_fp16)[name = string("op_1746_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_91_cast_fp16 = mul(x = x_62_cast_fp16, y = var_1746_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<int32, [2]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1752 = const()[name = string("op_1752"), val = tensor<int32, [2]>([1, 1])];
            string x_64_pad_type_0 = const()[name = string("x_64_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_64_pad_0 = const()[name = string("x_64_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52830336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52829248))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52895936)))];
            tensor<fp16, [1, 512, 1, 256]> x_64_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_1752, groups = var_46, pad = x_64_pad_0, pad_type = x_64_pad_type_0, strides = var_1750, weight = nlp_net_default_encoder_transformer_layers_5_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_91_cast_fp16)[name = string("x_64_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_12_cast_fp16 = add(x = x_64_cast_fp16, y = input0_23_cast_fp16)[name = string("f_12_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_12_cast_fp16 = add(x = f_12_cast_fp16, y = inputs0_12_cast_fp16)[name = string("x1_12_cast_fp16")];
            fp16 var_1757_to_fp16 = const()[name = string("op_1757_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_1758_cast_fp16 = mul(x = inputs3_1_cast_fp16, y = var_1757_to_fp16)[name = string("op_1758_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs4_1_cast_fp16 = add(x = var_1758_cast_fp16, y = x1_12_cast_fp16)[name = string("inputs4_1_cast_fp16")];
            tensor<int32, [1]> k_27_axes_0 = const()[name = string("k_27_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_27_gamma_0_to_fp16 = const()[name = string("k_27_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44916800)))];
            tensor<fp16, [512]> k_27_beta_0_to_fp16 = const()[name = string("k_27_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44917888)))];
            fp16 var_1776_to_fp16 = const()[name = string("op_1776_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_27_cast_fp16 = layer_norm(axes = k_27_axes_0, beta = k_27_beta_0_to_fp16, epsilon = var_1776_to_fp16, gamma = k_27_gamma_0_to_fp16, x = inputs4_1_cast_fp16)[name = string("k_27_cast_fp16")];
            tensor<int32, [2]> var_1795 = const()[name = string("op_1795"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1797 = const()[name = string("op_1797"), val = tensor<int32, [2]>([1, 1])];
            string var_1799_pad_type_0 = const()[name = string("op_1799_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1799_pad_0 = const()[name = string("op_1799_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44920064))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(44918976))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45182272)))];
            tensor<fp16, [1, 512, 1, 256]> var_1799_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_bias_to_fp16, dilations = var_1797, groups = var_46, pad = var_1799_pad_0, pad_type = var_1799_pad_type_0, strides = var_1795, weight = nlp_net_default_encoder_transformer_layers_6_attn_q_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1799_cast_fp16")];
            tensor<int32, [2]> var_1802 = const()[name = string("op_1802"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1804 = const()[name = string("op_1804"), val = tensor<int32, [2]>([1, 1])];
            string k_29_pad_type_0 = const()[name = string("k_29_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_29_pad_0 = const()[name = string("k_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45184448))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45183360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45446656)))];
            tensor<fp16, [1, 512, 1, 256]> k_29_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_bias_to_fp16, dilations = var_1804, groups = var_46, pad = k_29_pad_0, pad_type = k_29_pad_type_0, strides = var_1802, weight = nlp_net_default_encoder_transformer_layers_6_attn_k_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<int32, [2]> var_1809 = const()[name = string("op_1809"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1811 = const()[name = string("op_1811"), val = tensor<int32, [2]>([1, 1])];
            string var_1813_pad_type_0 = const()[name = string("op_1813_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_1813_pad_0 = const()[name = string("op_1813_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45448832))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45447744))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45711040)))];
            tensor<fp16, [1, 512, 1, 256]> var_1813_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_bias_to_fp16, dilations = var_1811, groups = var_46, pad = var_1813_pad_0, pad_type = var_1813_pad_type_0, strides = var_1809, weight = nlp_net_default_encoder_transformer_layers_6_attn_v_proj_weight_to_fp16_affine_quantized, x = k_27_cast_fp16)[name = string("op_1813_cast_fp16")];
            tensor<int32, [8]> tile_24 = const()[name = string("tile_24"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1814_axis_0 = const()[name = string("op_1814_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1814_cast_fp16_7 = split(axis = var_1814_axis_0, split_sizes = tile_24, x = var_1799_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<int32, [4]> var_1823_perm_0 = const()[name = string("op_1823_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_25 = const()[name = string("tile_25"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1824_axis_0 = const()[name = string("op_1824_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_2 = transpose(perm = var_1823_perm_0, x = k_29_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_1824_cast_fp16_7 = split(axis = var_1824_axis_0, split_sizes = tile_25, x = transpose_2)[name = string("op_1824_cast_fp16")];
            tensor<int32, [8]> tile_26 = const()[name = string("tile_26"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_1833_axis_0 = const()[name = string("op_1833_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_1833_cast_fp16_7 = split(axis = var_1833_axis_0, split_sizes = tile_26, x = var_1813_cast_fp16)[name = string("op_1833_cast_fp16")];
            string var_1843_equation_0 = const()[name = string("op_1843_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1843_cast_fp16 = einsum(equation = var_1843_equation_0, values = (var_1824_cast_fp16_0, var_1814_cast_fp16_0))[name = string("op_1843_cast_fp16")];
            fp16 var_1844_to_fp16 = const()[name = string("op_1844_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1845_cast_fp16 = mul(x = var_1843_cast_fp16, y = var_1844_to_fp16)[name = string("op_1845_cast_fp16")];
            string var_1847_equation_0 = const()[name = string("op_1847_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1847_cast_fp16 = einsum(equation = var_1847_equation_0, values = (var_1824_cast_fp16_1, var_1814_cast_fp16_1))[name = string("op_1847_cast_fp16")];
            fp16 var_1848_to_fp16 = const()[name = string("op_1848_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1849_cast_fp16 = mul(x = var_1847_cast_fp16, y = var_1848_to_fp16)[name = string("op_1849_cast_fp16")];
            string var_1851_equation_0 = const()[name = string("op_1851_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1851_cast_fp16 = einsum(equation = var_1851_equation_0, values = (var_1824_cast_fp16_2, var_1814_cast_fp16_2))[name = string("op_1851_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1853_cast_fp16 = mul(x = var_1851_cast_fp16, y = var_1852_to_fp16)[name = string("op_1853_cast_fp16")];
            string var_1855_equation_0 = const()[name = string("op_1855_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1855_cast_fp16 = einsum(equation = var_1855_equation_0, values = (var_1824_cast_fp16_3, var_1814_cast_fp16_3))[name = string("op_1855_cast_fp16")];
            fp16 var_1856_to_fp16 = const()[name = string("op_1856_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1857_cast_fp16 = mul(x = var_1855_cast_fp16, y = var_1856_to_fp16)[name = string("op_1857_cast_fp16")];
            string var_1859_equation_0 = const()[name = string("op_1859_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1859_cast_fp16 = einsum(equation = var_1859_equation_0, values = (var_1824_cast_fp16_4, var_1814_cast_fp16_4))[name = string("op_1859_cast_fp16")];
            fp16 var_1860_to_fp16 = const()[name = string("op_1860_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1861_cast_fp16 = mul(x = var_1859_cast_fp16, y = var_1860_to_fp16)[name = string("op_1861_cast_fp16")];
            string var_1863_equation_0 = const()[name = string("op_1863_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1863_cast_fp16 = einsum(equation = var_1863_equation_0, values = (var_1824_cast_fp16_5, var_1814_cast_fp16_5))[name = string("op_1863_cast_fp16")];
            fp16 var_1864_to_fp16 = const()[name = string("op_1864_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1865_cast_fp16 = mul(x = var_1863_cast_fp16, y = var_1864_to_fp16)[name = string("op_1865_cast_fp16")];
            string var_1867_equation_0 = const()[name = string("op_1867_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1867_cast_fp16 = einsum(equation = var_1867_equation_0, values = (var_1824_cast_fp16_6, var_1814_cast_fp16_6))[name = string("op_1867_cast_fp16")];
            fp16 var_1868_to_fp16 = const()[name = string("op_1868_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1869_cast_fp16 = mul(x = var_1867_cast_fp16, y = var_1868_to_fp16)[name = string("op_1869_cast_fp16")];
            string var_1871_equation_0 = const()[name = string("op_1871_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_1871_cast_fp16 = einsum(equation = var_1871_equation_0, values = (var_1824_cast_fp16_7, var_1814_cast_fp16_7))[name = string("op_1871_cast_fp16")];
            fp16 var_1872_to_fp16 = const()[name = string("op_1872_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_1873_cast_fp16 = mul(x = var_1871_cast_fp16, y = var_1872_to_fp16)[name = string("op_1873_cast_fp16")];
            bool attn_weights_14_interleave_0 = const()[name = string("attn_weights_14_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_14_cast_fp16 = concat(axis = var_47, interleave = attn_weights_14_interleave_0, values = (var_1845_cast_fp16, var_1849_cast_fp16, var_1853_cast_fp16, var_1857_cast_fp16, var_1861_cast_fp16, var_1865_cast_fp16, var_1869_cast_fp16, var_1873_cast_fp16))[name = string("attn_weights_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_14_cast_fp16 = add(x = attn_weights_14_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_14_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_93_cast_fp16 = softmax(axis = var_46, x = attn_weights0_14_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<int32, [8]> tile_27 = const()[name = string("tile_27"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_1879_axis_0 = const()[name = string("op_1879_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_1879_cast_fp16_7 = split(axis = var_1879_axis_0, split_sizes = tile_27, x = input_93_cast_fp16)[name = string("op_1879_cast_fp16")];
            string var_1889_equation_0 = const()[name = string("op_1889_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1889_cast_fp16 = einsum(equation = var_1889_equation_0, values = (var_1833_cast_fp16_0, var_1879_cast_fp16_0))[name = string("op_1889_cast_fp16")];
            string var_1891_equation_0 = const()[name = string("op_1891_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1891_cast_fp16 = einsum(equation = var_1891_equation_0, values = (var_1833_cast_fp16_1, var_1879_cast_fp16_1))[name = string("op_1891_cast_fp16")];
            string var_1893_equation_0 = const()[name = string("op_1893_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1893_cast_fp16 = einsum(equation = var_1893_equation_0, values = (var_1833_cast_fp16_2, var_1879_cast_fp16_2))[name = string("op_1893_cast_fp16")];
            string var_1895_equation_0 = const()[name = string("op_1895_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1895_cast_fp16 = einsum(equation = var_1895_equation_0, values = (var_1833_cast_fp16_3, var_1879_cast_fp16_3))[name = string("op_1895_cast_fp16")];
            string var_1897_equation_0 = const()[name = string("op_1897_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1897_cast_fp16 = einsum(equation = var_1897_equation_0, values = (var_1833_cast_fp16_4, var_1879_cast_fp16_4))[name = string("op_1897_cast_fp16")];
            string var_1899_equation_0 = const()[name = string("op_1899_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1899_cast_fp16 = einsum(equation = var_1899_equation_0, values = (var_1833_cast_fp16_5, var_1879_cast_fp16_5))[name = string("op_1899_cast_fp16")];
            string var_1901_equation_0 = const()[name = string("op_1901_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1901_cast_fp16 = einsum(equation = var_1901_equation_0, values = (var_1833_cast_fp16_6, var_1879_cast_fp16_6))[name = string("op_1901_cast_fp16")];
            string var_1903_equation_0 = const()[name = string("op_1903_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_1903_cast_fp16 = einsum(equation = var_1903_equation_0, values = (var_1833_cast_fp16_7, var_1879_cast_fp16_7))[name = string("op_1903_cast_fp16")];
            bool attn_41_interleave_0 = const()[name = string("attn_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_41_cast_fp16 = concat(axis = var_46, interleave = attn_41_interleave_0, values = (var_1889_cast_fp16, var_1891_cast_fp16, var_1893_cast_fp16, var_1895_cast_fp16, var_1897_cast_fp16, var_1899_cast_fp16, var_1901_cast_fp16, var_1903_cast_fp16))[name = string("attn_41_cast_fp16")];
            tensor<int32, [2]> var_1911 = const()[name = string("op_1911"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1913 = const()[name = string("op_1913"), val = tensor<int32, [2]>([1, 1])];
            string inputs_29_pad_type_0 = const()[name = string("inputs_29_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_29_pad_0 = const()[name = string("inputs_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45713216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45712128))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45975424)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_29_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_bias_to_fp16, dilations = var_1913, groups = var_46, pad = inputs_29_pad_0, pad_type = inputs_29_pad_type_0, strides = var_1911, weight = nlp_net_default_encoder_transformer_layers_6_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_41_cast_fp16)[name = string("inputs_29_cast_fp16")];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_95_gamma_0_to_fp16 = const()[name = string("input_95_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52897024)))];
            tensor<fp16, [512]> input_95_beta_0_to_fp16 = const()[name = string("input_95_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52898112)))];
            fp16 var_1927_to_fp16 = const()[name = string("op_1927_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = input_95_beta_0_to_fp16, epsilon = var_1927_to_fp16, gamma = input_95_gamma_0_to_fp16, x = inputs_29_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<int32, [2]> var_1939 = const()[name = string("op_1939"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1941 = const()[name = string("op_1941"), val = tensor<int32, [2]>([1, 1])];
            string x_66_pad_type_0 = const()[name = string("x_66_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_66_pad_0 = const()[name = string("x_66_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52899520))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52899200))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52965120)))];
            tensor<fp16, [1, 128, 1, 256]> x_66_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_bias_to_fp16, dilations = var_1941, groups = var_46, pad = x_66_pad_0, pad_type = x_66_pad_type_0, strides = var_1939, weight = nlp_net_default_encoder_transformer_layers_6_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_95_cast_fp16)[name = string("x_66_cast_fp16")];
            fp16 var_1944_to_fp16 = const()[name = string("op_1944_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_1945_cast_fp16 = mul(x = x_66_cast_fp16, y = var_1944_to_fp16)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_1946_cast_fp16 = sigmoid(x = var_1945_cast_fp16)[name = string("op_1946_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_97_cast_fp16 = mul(x = x_66_cast_fp16, y = var_1946_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<int32, [2]> var_1950 = const()[name = string("op_1950"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1952 = const()[name = string("op_1952"), val = tensor<int32, [2]>([1, 1])];
            string x_68_pad_type_0 = const()[name = string("x_68_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_68_pad_0 = const()[name = string("x_68_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52966528))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(52965440))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53032128)))];
            tensor<fp16, [1, 512, 1, 256]> x_68_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_bias_to_fp16, dilations = var_1952, groups = var_46, pad = x_68_pad_0, pad_type = x_68_pad_type_0, strides = var_1950, weight = nlp_net_default_encoder_transformer_layers_6_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_97_cast_fp16)[name = string("x_68_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_43_cast_fp16 = add(x = x_68_cast_fp16, y = inputs_29_cast_fp16)[name = string("attn_43_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_14_cast_fp16 = add(x = inputs4_1_cast_fp16, y = attn_43_cast_fp16)[name = string("inputs0_14_cast_fp16")];
            tensor<int32, [1]> input_99_axes_0 = const()[name = string("input_99_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_99_gamma_0_to_fp16 = const()[name = string("input_99_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45976512)))];
            tensor<fp16, [512]> input_99_beta_0_to_fp16 = const()[name = string("input_99_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45977600)))];
            fp16 var_1965_to_fp16 = const()[name = string("op_1965_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = input_99_beta_0_to_fp16, epsilon = var_1965_to_fp16, gamma = input_99_gamma_0_to_fp16, x = inputs0_14_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<int32, [2]> var_1979 = const()[name = string("op_1979"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1981 = const()[name = string("op_1981"), val = tensor<int32, [2]>([1, 1])];
            string x_70_pad_type_0 = const()[name = string("x_70_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_70_pad_0 = const()[name = string("x_70_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45982848))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(45978688))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47035648))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47031488))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_70_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_1981, groups = var_46, pad = x_70_pad_0, pad_type = x_70_pad_type_0, strides = var_1979, weight = nlp_net_default_encoder_transformer_layers_6_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_99_cast_fp16)[name = string("x_70_cast_fp16")];
            fp16 var_1984_to_fp16 = const()[name = string("op_1984_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_1985_cast_fp16 = mul(x = x_70_cast_fp16, y = var_1984_to_fp16)[name = string("op_1985_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_1986_cast_fp16 = sigmoid(x = var_1985_cast_fp16)[name = string("op_1986_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_101_cast_fp16 = mul(x = x_70_cast_fp16, y = var_1986_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<int32, [2]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_1992 = const()[name = string("op_1992"), val = tensor<int32, [2]>([1, 1])];
            string input0_27_pad_type_0 = const()[name = string("input0_27_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_27_pad_0 = const()[name = string("input0_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47038848))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(47037760))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48087488)))];
            tensor<fp16, [1, 512, 1, 256]> input0_27_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_bias_to_fp16, dilations = var_1992, groups = var_46, pad = input0_27_pad_0, pad_type = input0_27_pad_type_0, strides = var_1990, weight = nlp_net_default_encoder_transformer_layers_6_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_101_cast_fp16)[name = string("input0_27_cast_fp16")];
            tensor<int32, [1]> input_103_axes_0 = const()[name = string("input_103_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_103_gamma_0_to_fp16 = const()[name = string("input_103_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53033216)))];
            tensor<fp16, [512]> input_103_beta_0_to_fp16 = const()[name = string("input_103_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53034304)))];
            fp16 var_2007_to_fp16 = const()[name = string("op_2007_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_103_cast_fp16 = layer_norm(axes = input_103_axes_0, beta = input_103_beta_0_to_fp16, epsilon = var_2007_to_fp16, gamma = input_103_gamma_0_to_fp16, x = input0_27_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<int32, [2]> var_2019 = const()[name = string("op_2019"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [2]>([1, 1])];
            string x_72_pad_type_0 = const()[name = string("x_72_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_72_pad_0 = const()[name = string("x_72_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53035712))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53035392))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53101312)))];
            tensor<fp16, [1, 128, 1, 256]> x_72_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_2021, groups = var_46, pad = x_72_pad_0, pad_type = x_72_pad_type_0, strides = var_2019, weight = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_103_cast_fp16)[name = string("x_72_cast_fp16")];
            fp16 var_2024_to_fp16 = const()[name = string("op_2024_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_2025_cast_fp16 = mul(x = x_72_cast_fp16, y = var_2024_to_fp16)[name = string("op_2025_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_2026_cast_fp16 = sigmoid(x = var_2025_cast_fp16)[name = string("op_2026_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_105_cast_fp16 = mul(x = x_72_cast_fp16, y = var_2026_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<int32, [2]> var_2030 = const()[name = string("op_2030"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2032 = const()[name = string("op_2032"), val = tensor<int32, [2]>([1, 1])];
            string x_74_pad_type_0 = const()[name = string("x_74_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_74_pad_0 = const()[name = string("x_74_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53102720))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53101632))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53168320)))];
            tensor<fp16, [1, 512, 1, 256]> x_74_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_2032, groups = var_46, pad = x_74_pad_0, pad_type = x_74_pad_type_0, strides = var_2030, weight = nlp_net_default_encoder_transformer_layers_6_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_105_cast_fp16)[name = string("x_74_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_14_cast_fp16 = add(x = x_74_cast_fp16, y = input0_27_cast_fp16)[name = string("f_14_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_14_cast_fp16 = add(x = f_14_cast_fp16, y = inputs0_14_cast_fp16)[name = string("x1_14_cast_fp16")];
            fp16 var_2037_to_fp16 = const()[name = string("op_2037_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_2038_cast_fp16 = mul(x = inputs4_1_cast_fp16, y = var_2037_to_fp16)[name = string("op_2038_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs5_1_cast_fp16 = add(x = var_2038_cast_fp16, y = x1_14_cast_fp16)[name = string("inputs5_1_cast_fp16")];
            tensor<int32, [1]> k_2_axes_0 = const()[name = string("k_2_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> k_2_gamma_0_to_fp16 = const()[name = string("k_2_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48088576)))];
            tensor<fp16, [512]> k_2_beta_0_to_fp16 = const()[name = string("k_2_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48089664)))];
            fp16 var_2056_to_fp16 = const()[name = string("op_2056_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> k_2_cast_fp16 = layer_norm(axes = k_2_axes_0, beta = k_2_beta_0_to_fp16, epsilon = var_2056_to_fp16, gamma = k_2_gamma_0_to_fp16, x = inputs5_1_cast_fp16)[name = string("k_2_cast_fp16")];
            tensor<int32, [2]> var_2075 = const()[name = string("op_2075"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2077 = const()[name = string("op_2077"), val = tensor<int32, [2]>([1, 1])];
            string var_2079_pad_type_0 = const()[name = string("op_2079_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_2079_pad_0 = const()[name = string("op_2079_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48091840))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48090752))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48354048)))];
            tensor<fp16, [1, 512, 1, 256]> var_2079_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_bias_to_fp16, dilations = var_2077, groups = var_46, pad = var_2079_pad_0, pad_type = var_2079_pad_type_0, strides = var_2075, weight = nlp_net_default_encoder_transformer_layers_7_attn_q_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_2079_cast_fp16")];
            tensor<int32, [2]> var_2082 = const()[name = string("op_2082"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2084 = const()[name = string("op_2084"), val = tensor<int32, [2]>([1, 1])];
            string k_1_pad_type_0 = const()[name = string("k_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> k_1_pad_0 = const()[name = string("k_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48356224))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48355136))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48618432)))];
            tensor<fp16, [1, 512, 1, 256]> k_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_bias_to_fp16, dilations = var_2084, groups = var_46, pad = k_1_pad_0, pad_type = k_1_pad_type_0, strides = var_2082, weight = nlp_net_default_encoder_transformer_layers_7_attn_k_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<int32, [2]> var_2089 = const()[name = string("op_2089"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2091 = const()[name = string("op_2091"), val = tensor<int32, [2]>([1, 1])];
            string var_2093_pad_type_0 = const()[name = string("op_2093_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_2093_pad_0 = const()[name = string("op_2093_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48620608))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48619520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48882816)))];
            tensor<fp16, [1, 512, 1, 256]> var_2093_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_bias_to_fp16, dilations = var_2091, groups = var_46, pad = var_2093_pad_0, pad_type = var_2093_pad_type_0, strides = var_2089, weight = nlp_net_default_encoder_transformer_layers_7_attn_v_proj_weight_to_fp16_affine_quantized, x = k_2_cast_fp16)[name = string("op_2093_cast_fp16")];
            tensor<int32, [8]> tile_28 = const()[name = string("tile_28"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_2094_axis_0 = const()[name = string("op_2094_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_2094_cast_fp16_7 = split(axis = var_2094_axis_0, split_sizes = tile_28, x = var_2079_cast_fp16)[name = string("op_2094_cast_fp16")];
            tensor<int32, [4]> var_2103_perm_0 = const()[name = string("op_2103_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            tensor<int32, [8]> tile_29 = const()[name = string("tile_29"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_2104_axis_0 = const()[name = string("op_2104_axis_0"), val = int32(3)];
            tensor<fp16, [1, 256, 1, 512]> transpose_1 = transpose(perm = var_2103_perm_0, x = k_1_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_0, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_1, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_2, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_3, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_4, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_5, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_6, tensor<fp16, [1, 256, 1, 64]> var_2104_cast_fp16_7 = split(axis = var_2104_axis_0, split_sizes = tile_29, x = transpose_1)[name = string("op_2104_cast_fp16")];
            tensor<int32, [8]> tile_30 = const()[name = string("tile_30"), val = tensor<int32, [8]>([64, 64, 64, 64, 64, 64, 64, 64])];
            int32 var_2113_axis_0 = const()[name = string("op_2113_axis_0"), val = int32(1)];
            tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_0, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_1, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_2, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_3, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_4, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_5, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_6, tensor<fp16, [1, 64, 1, 256]> var_2113_cast_fp16_7 = split(axis = var_2113_axis_0, split_sizes = tile_30, x = var_2093_cast_fp16)[name = string("op_2113_cast_fp16")];
            string var_2123_equation_0 = const()[name = string("op_2123_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2123_cast_fp16 = einsum(equation = var_2123_equation_0, values = (var_2104_cast_fp16_0, var_2094_cast_fp16_0))[name = string("op_2123_cast_fp16")];
            fp16 var_2124_to_fp16 = const()[name = string("op_2124_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2125_cast_fp16 = mul(x = var_2123_cast_fp16, y = var_2124_to_fp16)[name = string("op_2125_cast_fp16")];
            string var_2127_equation_0 = const()[name = string("op_2127_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2127_cast_fp16 = einsum(equation = var_2127_equation_0, values = (var_2104_cast_fp16_1, var_2094_cast_fp16_1))[name = string("op_2127_cast_fp16")];
            fp16 var_2128_to_fp16 = const()[name = string("op_2128_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2129_cast_fp16 = mul(x = var_2127_cast_fp16, y = var_2128_to_fp16)[name = string("op_2129_cast_fp16")];
            string var_2131_equation_0 = const()[name = string("op_2131_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2131_cast_fp16 = einsum(equation = var_2131_equation_0, values = (var_2104_cast_fp16_2, var_2094_cast_fp16_2))[name = string("op_2131_cast_fp16")];
            fp16 var_2132_to_fp16 = const()[name = string("op_2132_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2133_cast_fp16 = mul(x = var_2131_cast_fp16, y = var_2132_to_fp16)[name = string("op_2133_cast_fp16")];
            string var_2135_equation_0 = const()[name = string("op_2135_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2135_cast_fp16 = einsum(equation = var_2135_equation_0, values = (var_2104_cast_fp16_3, var_2094_cast_fp16_3))[name = string("op_2135_cast_fp16")];
            fp16 var_2136_to_fp16 = const()[name = string("op_2136_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2137_cast_fp16 = mul(x = var_2135_cast_fp16, y = var_2136_to_fp16)[name = string("op_2137_cast_fp16")];
            string var_2139_equation_0 = const()[name = string("op_2139_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2139_cast_fp16 = einsum(equation = var_2139_equation_0, values = (var_2104_cast_fp16_4, var_2094_cast_fp16_4))[name = string("op_2139_cast_fp16")];
            fp16 var_2140_to_fp16 = const()[name = string("op_2140_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2141_cast_fp16 = mul(x = var_2139_cast_fp16, y = var_2140_to_fp16)[name = string("op_2141_cast_fp16")];
            string var_2143_equation_0 = const()[name = string("op_2143_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2143_cast_fp16 = einsum(equation = var_2143_equation_0, values = (var_2104_cast_fp16_5, var_2094_cast_fp16_5))[name = string("op_2143_cast_fp16")];
            fp16 var_2144_to_fp16 = const()[name = string("op_2144_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2145_cast_fp16 = mul(x = var_2143_cast_fp16, y = var_2144_to_fp16)[name = string("op_2145_cast_fp16")];
            string var_2147_equation_0 = const()[name = string("op_2147_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2147_cast_fp16 = einsum(equation = var_2147_equation_0, values = (var_2104_cast_fp16_6, var_2094_cast_fp16_6))[name = string("op_2147_cast_fp16")];
            fp16 var_2148_to_fp16 = const()[name = string("op_2148_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2149_cast_fp16 = mul(x = var_2147_cast_fp16, y = var_2148_to_fp16)[name = string("op_2149_cast_fp16")];
            string var_2151_equation_0 = const()[name = string("op_2151_equation_0"), val = string("bkhc,bchq->bkhq")];
            tensor<fp16, [1, 256, 1, 256]> var_2151_cast_fp16 = einsum(equation = var_2151_equation_0, values = (var_2104_cast_fp16_7, var_2094_cast_fp16_7))[name = string("op_2151_cast_fp16")];
            fp16 var_2152_to_fp16 = const()[name = string("op_2152_to_fp16"), val = fp16(0.125)];
            tensor<fp16, [1, 256, 1, 256]> var_2153_cast_fp16 = mul(x = var_2151_cast_fp16, y = var_2152_to_fp16)[name = string("op_2153_cast_fp16")];
            bool attn_weights_1_interleave_0 = const()[name = string("attn_weights_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 256, 8, 256]> attn_weights_1_cast_fp16 = concat(axis = var_47, interleave = attn_weights_1_interleave_0, values = (var_2125_cast_fp16, var_2129_cast_fp16, var_2133_cast_fp16, var_2137_cast_fp16, var_2141_cast_fp16, var_2145_cast_fp16, var_2149_cast_fp16, var_2153_cast_fp16))[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> attn_weights0_1_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = var_33_cast_fp16)[name = string("attn_weights0_1_cast_fp16")];
            tensor<fp16, [1, 256, 8, 256]> input_4_cast_fp16 = softmax(axis = var_46, x = attn_weights0_1_cast_fp16)[name = string("input_4_cast_fp16")];
            tensor<int32, [8]> tile_31 = const()[name = string("tile_31"), val = tensor<int32, [8]>([1, 1, 1, 1, 1, 1, 1, 1])];
            int32 var_2159_axis_0 = const()[name = string("op_2159_axis_0"), val = int32(2)];
            tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_0, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_1, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_2, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_3, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_4, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_5, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_6, tensor<fp16, [1, 256, 1, 256]> var_2159_cast_fp16_7 = split(axis = var_2159_axis_0, split_sizes = tile_31, x = input_4_cast_fp16)[name = string("op_2159_cast_fp16")];
            string var_2169_equation_0 = const()[name = string("op_2169_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2169_cast_fp16 = einsum(equation = var_2169_equation_0, values = (var_2113_cast_fp16_0, var_2159_cast_fp16_0))[name = string("op_2169_cast_fp16")];
            string var_2171_equation_0 = const()[name = string("op_2171_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2171_cast_fp16 = einsum(equation = var_2171_equation_0, values = (var_2113_cast_fp16_1, var_2159_cast_fp16_1))[name = string("op_2171_cast_fp16")];
            string var_2173_equation_0 = const()[name = string("op_2173_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2173_cast_fp16 = einsum(equation = var_2173_equation_0, values = (var_2113_cast_fp16_2, var_2159_cast_fp16_2))[name = string("op_2173_cast_fp16")];
            string var_2175_equation_0 = const()[name = string("op_2175_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2175_cast_fp16 = einsum(equation = var_2175_equation_0, values = (var_2113_cast_fp16_3, var_2159_cast_fp16_3))[name = string("op_2175_cast_fp16")];
            string var_2177_equation_0 = const()[name = string("op_2177_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2177_cast_fp16 = einsum(equation = var_2177_equation_0, values = (var_2113_cast_fp16_4, var_2159_cast_fp16_4))[name = string("op_2177_cast_fp16")];
            string var_2179_equation_0 = const()[name = string("op_2179_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2179_cast_fp16 = einsum(equation = var_2179_equation_0, values = (var_2113_cast_fp16_5, var_2159_cast_fp16_5))[name = string("op_2179_cast_fp16")];
            string var_2181_equation_0 = const()[name = string("op_2181_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2181_cast_fp16 = einsum(equation = var_2181_equation_0, values = (var_2113_cast_fp16_6, var_2159_cast_fp16_6))[name = string("op_2181_cast_fp16")];
            string var_2183_equation_0 = const()[name = string("op_2183_equation_0"), val = string("bchk,bkhq->bchq")];
            tensor<fp16, [1, 64, 1, 256]> var_2183_cast_fp16 = einsum(equation = var_2183_equation_0, values = (var_2113_cast_fp16_7, var_2159_cast_fp16_7))[name = string("op_2183_cast_fp16")];
            bool attn_4_interleave_0 = const()[name = string("attn_4_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 512, 1, 256]> attn_4_cast_fp16 = concat(axis = var_46, interleave = attn_4_interleave_0, values = (var_2169_cast_fp16, var_2171_cast_fp16, var_2173_cast_fp16, var_2175_cast_fp16, var_2177_cast_fp16, var_2179_cast_fp16, var_2181_cast_fp16, var_2183_cast_fp16))[name = string("attn_4_cast_fp16")];
            tensor<int32, [2]> var_2191 = const()[name = string("op_2191"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2193 = const()[name = string("op_2193"), val = tensor<int32, [2]>([1, 1])];
            string inputs_4_pad_type_0 = const()[name = string("inputs_4_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> inputs_4_pad_0 = const()[name = string("inputs_4_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48884992))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(48883904))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49147200)))];
            tensor<fp16, [1, 512, 1, 256]> inputs_4_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_bias_to_fp16, dilations = var_2193, groups = var_46, pad = inputs_4_pad_0, pad_type = inputs_4_pad_type_0, strides = var_2191, weight = nlp_net_default_encoder_transformer_layers_7_attn_out_proj_weight_to_fp16_affine_quantized, x = attn_4_cast_fp16)[name = string("inputs_4_cast_fp16")];
            tensor<int32, [1]> input_6_axes_0 = const()[name = string("input_6_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_6_gamma_0_to_fp16 = const()[name = string("input_6_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53169408)))];
            tensor<fp16, [512]> input_6_beta_0_to_fp16 = const()[name = string("input_6_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53170496)))];
            fp16 var_2207_to_fp16 = const()[name = string("op_2207_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_6_cast_fp16 = layer_norm(axes = input_6_axes_0, beta = input_6_beta_0_to_fp16, epsilon = var_2207_to_fp16, gamma = input_6_gamma_0_to_fp16, x = inputs_4_cast_fp16)[name = string("input_6_cast_fp16")];
            tensor<int32, [2]> var_2219 = const()[name = string("op_2219"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2221 = const()[name = string("op_2221"), val = tensor<int32, [2]>([1, 1])];
            string x_5_pad_type_0 = const()[name = string("x_5_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_5_pad_0 = const()[name = string("x_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53171904))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53171584))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53237504)))];
            tensor<fp16, [1, 128, 1, 256]> x_5_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_bias_to_fp16, dilations = var_2221, groups = var_46, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = var_2219, weight = nlp_net_default_encoder_transformer_layers_7_attn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_6_cast_fp16)[name = string("x_5_cast_fp16")];
            fp16 var_2224_to_fp16 = const()[name = string("op_2224_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_2225_cast_fp16 = mul(x = x_5_cast_fp16, y = var_2224_to_fp16)[name = string("op_2225_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_2226_cast_fp16 = sigmoid(x = var_2225_cast_fp16)[name = string("op_2226_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_12_cast_fp16 = mul(x = x_5_cast_fp16, y = var_2226_cast_fp16)[name = string("input_12_cast_fp16")];
            tensor<int32, [2]> var_2230 = const()[name = string("op_2230"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2232 = const()[name = string("op_2232"), val = tensor<int32, [2]>([1, 1])];
            string x_3_pad_type_0 = const()[name = string("x_3_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_3_pad_0 = const()[name = string("x_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53238912))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53237824))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53304512)))];
            tensor<fp16, [1, 512, 1, 256]> x_3_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_bias_to_fp16, dilations = var_2232, groups = var_46, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = var_2230, weight = nlp_net_default_encoder_transformer_layers_7_attn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_12_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> attn_1_cast_fp16 = add(x = x_3_cast_fp16, y = inputs_4_cast_fp16)[name = string("attn_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs0_1_cast_fp16 = add(x = inputs5_1_cast_fp16, y = attn_1_cast_fp16)[name = string("inputs0_1_cast_fp16")];
            tensor<int32, [1]> input_8_axes_0 = const()[name = string("input_8_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_8_gamma_0_to_fp16 = const()[name = string("input_8_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49148288)))];
            tensor<fp16, [512]> input_8_beta_0_to_fp16 = const()[name = string("input_8_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49149376)))];
            fp16 var_2245_to_fp16 = const()[name = string("op_2245_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_8_cast_fp16 = layer_norm(axes = input_8_axes_0, beta = input_8_beta_0_to_fp16, epsilon = var_2245_to_fp16, gamma = input_8_gamma_0_to_fp16, x = inputs0_1_cast_fp16)[name = string("input_8_cast_fp16")];
            tensor<int32, [2]> var_2259 = const()[name = string("op_2259"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2261 = const()[name = string("op_2261"), val = tensor<int32, [2]>([1, 1])];
            string x_7_pad_type_0 = const()[name = string("x_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_7_pad_0 = const()[name = string("x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [2048, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49154624))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(49150464))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [2048]> nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized"), quantized_data = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50207424))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50203264))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(26945920)))];
            tensor<fp16, [1, 2048, 1, 256]> x_7_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_bias_to_fp16_affine_quantized, dilations = var_2261, groups = var_46, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = var_2259, weight = nlp_net_default_encoder_transformer_layers_7_ffn_expand_proj_weight_to_fp16_affine_quantized, x = input_8_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_2264_to_fp16 = const()[name = string("op_2264_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 2048, 1, 256]> var_2265_cast_fp16 = mul(x = x_7_cast_fp16, y = var_2264_to_fp16)[name = string("op_2265_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> var_2266_cast_fp16 = sigmoid(x = var_2265_cast_fp16)[name = string("op_2266_cast_fp16")];
            tensor<fp16, [1, 2048, 1, 256]> input_10_cast_fp16 = mul(x = x_7_cast_fp16, y = var_2266_cast_fp16)[name = string("input_10_cast_fp16")];
            tensor<int32, [2]> var_2270 = const()[name = string("op_2270"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2272 = const()[name = string("op_2272"), val = tensor<int32, [2]>([1, 1])];
            string input0_1_pad_type_0 = const()[name = string("input0_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input0_1_pad_0 = const()[name = string("input0_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 2048, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50210624))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(50209536))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51259264)))];
            tensor<fp16, [1, 512, 1, 256]> input0_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_bias_to_fp16, dilations = var_2272, groups = var_46, pad = input0_1_pad_0, pad_type = input0_1_pad_type_0, strides = var_2270, weight = nlp_net_default_encoder_transformer_layers_7_ffn_contract_proj_weight_to_fp16_affine_quantized, x = input_10_cast_fp16)[name = string("input0_1_cast_fp16")];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> input_1_gamma_0_to_fp16 = const()[name = string("input_1_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53305600)))];
            tensor<fp16, [512]> input_1_beta_0_to_fp16 = const()[name = string("input_1_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53306688)))];
            fp16 var_2287_to_fp16 = const()[name = string("op_2287_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> input_1_cast_fp16 = layer_norm(axes = input_1_axes_0, beta = input_1_beta_0_to_fp16, epsilon = var_2287_to_fp16, gamma = input_1_gamma_0_to_fp16, x = input0_1_cast_fp16)[name = string("input_1_cast_fp16")];
            tensor<int32, [2]> var_2299 = const()[name = string("op_2299"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2301 = const()[name = string("op_2301"), val = tensor<int32, [2]>([1, 1])];
            string x_2_pad_type_0 = const()[name = string("x_2_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_2_pad_0 = const()[name = string("x_2_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [128, 512, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [128, 512, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53308096))), scale = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53307776))), zero_point = tensor<int8, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51264704)))];
            tensor<fp16, [128]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53373696)))];
            tensor<fp16, [1, 128, 1, 256]> x_2_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_bias_to_fp16, dilations = var_2301, groups = var_46, pad = x_2_pad_0, pad_type = x_2_pad_type_0, strides = var_2299, weight = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_contract_proj_weight_to_fp16_affine_quantized, x = input_1_cast_fp16)[name = string("x_2_cast_fp16")];
            fp16 var_2304_to_fp16 = const()[name = string("op_2304_to_fp16"), val = fp16(1.70214844)];
            tensor<fp16, [1, 128, 1, 256]> var_2305_cast_fp16 = mul(x = x_2_cast_fp16, y = var_2304_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> var_2306_cast_fp16 = sigmoid(x = var_2305_cast_fp16)[name = string("op_2306_cast_fp16")];
            tensor<fp16, [1, 128, 1, 256]> input_2_cast_fp16 = mul(x = x_2_cast_fp16, y = var_2306_cast_fp16)[name = string("input_2_cast_fp16")];
            tensor<int32, [2]> var_2310 = const()[name = string("op_2310"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> var_2312 = const()[name = string("op_2312"), val = tensor<int32, [2]>([1, 1])];
            string x_1_pad_type_0 = const()[name = string("x_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> x_1_pad_0 = const()[name = string("x_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<fp16, [512, 128, 1, 1]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized = constexpr_affine_dequantize()[axis = int32(0), name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized"), quantized_data = tensor<int8, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53375104))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53374016))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(25885632)))];
            tensor<fp16, [512]> nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16 = const()[name = string("nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(53440704)))];
            tensor<fp16, [1, 512, 1, 256]> x_1_cast_fp16 = conv(bias = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_bias_to_fp16, dilations = var_2312, groups = var_46, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = var_2310, weight = nlp_net_default_encoder_transformer_layers_7_ffn_adapter_expand_proj_weight_to_fp16_affine_quantized, x = input_2_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> f_1_cast_fp16 = add(x = x_1_cast_fp16, y = input0_1_cast_fp16)[name = string("f_1_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> x1_1_cast_fp16 = add(x = f_1_cast_fp16, y = inputs0_1_cast_fp16)[name = string("x1_1_cast_fp16")];
            fp16 var_2317_to_fp16 = const()[name = string("op_2317_to_fp16"), val = fp16(0)];
            tensor<fp16, [1, 512, 1, 256]> var_2318_cast_fp16 = mul(x = inputs5_1_cast_fp16, y = var_2317_to_fp16)[name = string("op_2318_cast_fp16")];
            tensor<fp16, [1, 512, 1, 256]> inputs6_1_cast_fp16 = add(x = var_2318_cast_fp16, y = x1_1_cast_fp16)[name = string("inputs6_1_cast_fp16")];
            tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [512]> embeddings_1_gamma_0_to_fp16 = const()[name = string("embeddings_1_gamma_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51260352)))];
            tensor<fp16, [512]> embeddings_1_beta_0_to_fp16 = const()[name = string("embeddings_1_beta_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights.bin"), offset = uint64(51261440)))];
            fp16 var_2330_to_fp16 = const()[name = string("op_2330_to_fp16"), val = fp16(1.00135803e-05)];
            tensor<fp16, [1, 512, 1, 256]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = embeddings_1_beta_0_to_fp16, epsilon = var_2330_to_fp16, gamma = embeddings_1_gamma_0_to_fp16, x = inputs6_1_cast_fp16)[name = string("embeddings_1_cast_fp16")];
            tensor<int32, [4]> mlm_embeddings_1_perm_0 = const()[name = string("mlm_embeddings_1_perm_0"), val = tensor<int32, [4]>([0, 3, 2, 1])];
            string mlm_embeddings_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("mlm_embeddings_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
            tensor<fp16, [1, 256, 1, 512]> transpose_0 = transpose(perm = mlm_embeddings_1_perm_0, x = embeddings_1_cast_fp16)[name = string("transpose_0")];
            tensor<fp32, [1, 256, 1, 512]> mlm_embeddings = cast(dtype = mlm_embeddings_1_cast_fp16_to_fp32_dtype_0, x = transpose_0)[name = string("cast_0")];
        } -> (mlm_embeddings);
}