program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3400.23.1"}, {"coremlc-version", "3400.24.1"}}), mldb_token = string("mldb-5qbg63zgxe")]
{
    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(0x1p+0)];
            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(-0x1.388p+13)];
            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, [35000, 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, [35000, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [35000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17955200))), zero_point = tensor<int8, [35000]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17920128)))];
            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/weight.bin"), offset = uint64(18025280))), scale = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18156736))), zero_point = tensor<int8, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18156416)))];
            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/weight.bin"), offset = uint64(18157312)))];
            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/weight.bin"), offset = uint64(18158400)))];
            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/weight.bin"), offset = uint64(18159488)))];
            fp16 var_94_to_fp16 = const()[name = string("op_94_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(18160576))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18423360))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18422784)))];
            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/weight.bin"), offset = uint64(18424448)))];
            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/weight.bin"), offset = uint64(18425536))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18688320))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18687744)))];
            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/weight.bin"), offset = uint64(18689408)))];
            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/weight.bin"), offset = uint64(18690496))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18953280))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18952704)))];
            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/weight.bin"), offset = uint64(18954368)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(18955456))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19218240))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19217664)))];
            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/weight.bin"), offset = uint64(19219328)))];
            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/weight.bin"), offset = uint64(19220416)))];
            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/weight.bin"), offset = uint64(19221504)))];
            fp16 var_243_to_fp16 = const()[name = string("op_243_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(19222592))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20273344))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20271232)))];
            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/weight.bin"), offset = uint64(20277504))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20281728))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20279616)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(20285888))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21335104))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21334528)))];
            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/weight.bin"), offset = uint64(21336192)))];
            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(0x0p+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/weight.bin"), offset = uint64(21337280)))];
            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/weight.bin"), offset = uint64(21338368)))];
            fp16 var_292_to_fp16 = const()[name = string("op_292_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(21339456))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21602240))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21601664)))];
            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/weight.bin"), offset = uint64(21603328)))];
            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/weight.bin"), offset = uint64(21604416))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21867200))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21866624)))];
            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/weight.bin"), offset = uint64(21868288)))];
            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/weight.bin"), offset = uint64(21869376))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22132160))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22131584)))];
            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/weight.bin"), offset = uint64(22133248)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(22134336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22397120))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22396544)))];
            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/weight.bin"), offset = uint64(22398208)))];
            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/weight.bin"), offset = uint64(22399296)))];
            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/weight.bin"), offset = uint64(22400384)))];
            fp16 var_441_to_fp16 = const()[name = string("op_441_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(22401472))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23452224))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23450112)))];
            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/weight.bin"), offset = uint64(23456384))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23460608))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23458496)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(23464768))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24513984))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24513408)))];
            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/weight.bin"), offset = uint64(24515072)))];
            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(0x0p+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/weight.bin"), offset = uint64(24516160)))];
            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/weight.bin"), offset = uint64(24517248)))];
            fp16 var_490_to_fp16 = const()[name = string("op_490_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(24518336))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24781120))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24780544)))];
            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/weight.bin"), offset = uint64(24782208)))];
            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/weight.bin"), offset = uint64(24783296))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25046080))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25045504)))];
            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/weight.bin"), offset = uint64(25047168)))];
            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/weight.bin"), offset = uint64(25048256))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25311040))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25310464)))];
            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/weight.bin"), offset = uint64(25312128)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(25313216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25576000))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25575424)))];
            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/weight.bin"), offset = uint64(25577088)))];
            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/weight.bin"), offset = uint64(25578176)))];
            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/weight.bin"), offset = uint64(25579264)))];
            fp16 var_639_to_fp16 = const()[name = string("op_639_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(25580352))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26631104))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26628992)))];
            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/weight.bin"), offset = uint64(26635264))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26639488))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26637376)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(26643648))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27692864))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27692288)))];
            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/weight.bin"), offset = uint64(27693952)))];
            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(0x0p+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/weight.bin"), offset = uint64(27695040)))];
            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/weight.bin"), offset = uint64(27696128)))];
            fp16 var_688_to_fp16 = const()[name = string("op_688_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(27697216))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27960000))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27959424)))];
            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/weight.bin"), offset = uint64(27961088)))];
            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/weight.bin"), offset = uint64(27962176))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28224960))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28224384)))];
            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/weight.bin"), offset = uint64(28226048)))];
            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/weight.bin"), offset = uint64(28227136))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28489920))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28489344)))];
            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/weight.bin"), offset = uint64(28491008)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(28492096))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28754880))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28754304)))];
            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/weight.bin"), offset = uint64(28755968)))];
            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/weight.bin"), offset = uint64(28757056)))];
            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/weight.bin"), offset = uint64(28758144)))];
            fp16 var_837_to_fp16 = const()[name = string("op_837_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(28759232))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29809984))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29807872)))];
            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/weight.bin"), offset = uint64(29814144))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29818368))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29816256)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(29822528))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30871744))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30871168)))];
            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/weight.bin"), offset = uint64(30872832)))];
            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(0x0p+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/weight.bin"), offset = uint64(30873920)))];
            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/weight.bin"), offset = uint64(30875008)))];
            fp16 var_886_to_fp16 = const()[name = string("op_886_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(30876096))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31138880))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31138304)))];
            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/weight.bin"), offset = uint64(31139968)))];
            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/weight.bin"), offset = uint64(31141056))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31403840))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31403264)))];
            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/weight.bin"), offset = uint64(31404928)))];
            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/weight.bin"), offset = uint64(31406016))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31668800))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31668224)))];
            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/weight.bin"), offset = uint64(31669888)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(31670976))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31933760))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31933184)))];
            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/weight.bin"), offset = uint64(31934848)))];
            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/weight.bin"), offset = uint64(31935936)))];
            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/weight.bin"), offset = uint64(31937024)))];
            fp16 var_1035_to_fp16 = const()[name = string("op_1035_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(31938112))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32988864))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32986752)))];
            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/weight.bin"), offset = uint64(32993024))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32997248))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32995136)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(33001408))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34050624))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34050048)))];
            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/weight.bin"), offset = uint64(34051712)))];
            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(0x0p+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/weight.bin"), offset = uint64(34052800)))];
            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/weight.bin"), offset = uint64(34053888)))];
            fp16 var_1084_to_fp16 = const()[name = string("op_1084_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(34054976))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34317760))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34317184)))];
            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/weight.bin"), offset = uint64(34318848)))];
            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/weight.bin"), offset = uint64(34319936))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34582720))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34582144)))];
            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/weight.bin"), offset = uint64(34583808)))];
            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/weight.bin"), offset = uint64(34584896))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34847680))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34847104)))];
            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/weight.bin"), offset = uint64(34848768)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(34849856))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35112640))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35112064)))];
            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/weight.bin"), offset = uint64(35113728)))];
            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/weight.bin"), offset = uint64(35114816)))];
            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/weight.bin"), offset = uint64(35115904)))];
            fp16 var_1233_to_fp16 = const()[name = string("op_1233_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(35116992))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36167744))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36165632)))];
            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/weight.bin"), offset = uint64(36171904))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36176128))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36174016)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(36180288))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37229504))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37228928)))];
            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/weight.bin"), offset = uint64(37230592)))];
            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(0x0p+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/weight.bin"), offset = uint64(37231680)))];
            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/weight.bin"), offset = uint64(37232768)))];
            fp16 var_1282_to_fp16 = const()[name = string("op_1282_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(37233856))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37496640))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37496064)))];
            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/weight.bin"), offset = uint64(37497728)))];
            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/weight.bin"), offset = uint64(37498816))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37761600))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37761024)))];
            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/weight.bin"), offset = uint64(37762688)))];
            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/weight.bin"), offset = uint64(37763776))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38026560))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38025984)))];
            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/weight.bin"), offset = uint64(38027648)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(38028736))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38291520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38290944)))];
            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/weight.bin"), offset = uint64(38292608)))];
            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/weight.bin"), offset = uint64(38293696)))];
            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/weight.bin"), offset = uint64(38294784)))];
            fp16 var_1431_to_fp16 = const()[name = string("op_1431_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(38295872))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39346624))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39344512)))];
            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/weight.bin"), offset = uint64(39350784))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39355008))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39352896)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(39359168))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40408384))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40407808)))];
            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/weight.bin"), offset = uint64(40409472)))];
            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(0x0p+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/weight.bin"), offset = uint64(40410560)))];
            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/weight.bin"), offset = uint64(40411648)))];
            fp16 var_1480_to_fp16 = const()[name = string("op_1480_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(40412736))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40675520))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40674944)))];
            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/weight.bin"), offset = uint64(40676608)))];
            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/weight.bin"), offset = uint64(40677696))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40940480))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40939904)))];
            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/weight.bin"), offset = uint64(40941568)))];
            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/weight.bin"), offset = uint64(40942656))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41205440))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41204864)))];
            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/weight.bin"), offset = uint64(41206528)))];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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(0x1p-3)];
            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/weight.bin"), offset = uint64(41207616))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41470400))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41469824)))];
            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/weight.bin"), offset = uint64(41471488)))];
            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/weight.bin"), offset = uint64(41472576)))];
            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/weight.bin"), offset = uint64(41473664)))];
            fp16 var_1629_to_fp16 = const()[name = string("op_1629_to_fp16"), val = fp16(0x1.5p-17)];
            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/weight.bin"), offset = uint64(41474752))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42525504))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42523392)))];
            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/weight.bin"), offset = uint64(42529664))), scale = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42533888))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42531776)))];
            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(0x1.b3cp+0)];
            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/weight.bin"), offset = uint64(42538048))), scale = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43587264))), zero_point = tensor<int8, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43586688)))];
            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/weight.bin"), offset = uint64(43588352)))];
            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(0x0p+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/weight.bin"), offset = uint64(43589440)))];
            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/weight.bin"), offset = uint64(43590528)))];
            fp16 var_1674_to_fp16 = const()[name = string("op_1674_to_fp16"), val = fp16(0x1.5p-17)];
            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);
}