program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.11.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios19>(tensor<fp16, [1, 512]> attention_mask, tensor<int32, [1, 512]> input_ids, tensor<fp16, [32, 512]> pool_matrix) {
            tensor<fp16, [32768, 128]> encoder_sin_cached = const()[name = string("encoder_sin_cached"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [32768, 128]> encoder_cos_cached = const()[name = string("encoder_cos_cached"), val = tensor<fp16, [32768, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8388736)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_0_self_attn_q_proj_weight = const()[name = string("encoder_layers_0_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16777408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_0_self_attn_k_proj_weight = const()[name = string("encoder_layers_0_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20971776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_0_self_attn_v_proj_weight = const()[name = string("encoder_layers_0_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23068992)))];
            tensor<fp16, [128]> encoder_layers_0_self_attn_q_norm_weight = const()[name = string("encoder_layers_0_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166208)))];
            tensor<fp16, [128]> encoder_layers_0_self_attn_k_norm_weight = const()[name = string("encoder_layers_0_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_0_mlp_gate_proj_weight = const()[name = string("encoder_layers_0_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25166848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_0_mlp_up_proj_weight = const()[name = string("encoder_layers_0_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31458368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_0_mlp_down_proj_weight = const()[name = string("encoder_layers_0_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37749888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_1_self_attn_q_proj_weight = const()[name = string("encoder_layers_1_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44041408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_1_self_attn_k_proj_weight = const()[name = string("encoder_layers_1_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48235776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_1_self_attn_v_proj_weight = const()[name = string("encoder_layers_1_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50332992)))];
            tensor<fp16, [128]> encoder_layers_1_self_attn_q_norm_weight = const()[name = string("encoder_layers_1_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430208)))];
            tensor<fp16, [128]> encoder_layers_1_self_attn_k_norm_weight = const()[name = string("encoder_layers_1_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_1_mlp_gate_proj_weight = const()[name = string("encoder_layers_1_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52430848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_1_mlp_up_proj_weight = const()[name = string("encoder_layers_1_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58722368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_1_mlp_down_proj_weight = const()[name = string("encoder_layers_1_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65013888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_2_self_attn_q_proj_weight = const()[name = string("encoder_layers_2_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71305408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_2_self_attn_k_proj_weight = const()[name = string("encoder_layers_2_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75499776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_2_self_attn_v_proj_weight = const()[name = string("encoder_layers_2_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77596992)))];
            tensor<fp16, [128]> encoder_layers_2_self_attn_q_norm_weight = const()[name = string("encoder_layers_2_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694208)))];
            tensor<fp16, [128]> encoder_layers_2_self_attn_k_norm_weight = const()[name = string("encoder_layers_2_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_2_mlp_gate_proj_weight = const()[name = string("encoder_layers_2_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79694848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_2_mlp_up_proj_weight = const()[name = string("encoder_layers_2_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85986368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_2_mlp_down_proj_weight = const()[name = string("encoder_layers_2_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92277888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_3_self_attn_q_proj_weight = const()[name = string("encoder_layers_3_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98569408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_3_self_attn_k_proj_weight = const()[name = string("encoder_layers_3_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102763776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_3_self_attn_v_proj_weight = const()[name = string("encoder_layers_3_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104860992)))];
            tensor<fp16, [128]> encoder_layers_3_self_attn_q_norm_weight = const()[name = string("encoder_layers_3_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958208)))];
            tensor<fp16, [128]> encoder_layers_3_self_attn_k_norm_weight = const()[name = string("encoder_layers_3_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_3_mlp_gate_proj_weight = const()[name = string("encoder_layers_3_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106958848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_3_mlp_up_proj_weight = const()[name = string("encoder_layers_3_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113250368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_3_mlp_down_proj_weight = const()[name = string("encoder_layers_3_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119541888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_4_self_attn_q_proj_weight = const()[name = string("encoder_layers_4_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125833408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_4_self_attn_k_proj_weight = const()[name = string("encoder_layers_4_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130027776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_4_self_attn_v_proj_weight = const()[name = string("encoder_layers_4_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132124992)))];
            tensor<fp16, [128]> encoder_layers_4_self_attn_q_norm_weight = const()[name = string("encoder_layers_4_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222208)))];
            tensor<fp16, [128]> encoder_layers_4_self_attn_k_norm_weight = const()[name = string("encoder_layers_4_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_4_mlp_gate_proj_weight = const()[name = string("encoder_layers_4_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134222848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_4_mlp_up_proj_weight = const()[name = string("encoder_layers_4_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140514368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_4_mlp_down_proj_weight = const()[name = string("encoder_layers_4_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146805888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_5_self_attn_q_proj_weight = const()[name = string("encoder_layers_5_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153097408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_5_self_attn_k_proj_weight = const()[name = string("encoder_layers_5_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157291776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_5_self_attn_v_proj_weight = const()[name = string("encoder_layers_5_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159388992)))];
            tensor<fp16, [128]> encoder_layers_5_self_attn_q_norm_weight = const()[name = string("encoder_layers_5_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486208)))];
            tensor<fp16, [128]> encoder_layers_5_self_attn_k_norm_weight = const()[name = string("encoder_layers_5_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_5_mlp_gate_proj_weight = const()[name = string("encoder_layers_5_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161486848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_5_mlp_up_proj_weight = const()[name = string("encoder_layers_5_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167778368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_5_mlp_down_proj_weight = const()[name = string("encoder_layers_5_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174069888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_6_self_attn_q_proj_weight = const()[name = string("encoder_layers_6_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180361408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_6_self_attn_k_proj_weight = const()[name = string("encoder_layers_6_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184555776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_6_self_attn_v_proj_weight = const()[name = string("encoder_layers_6_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186652992)))];
            tensor<fp16, [128]> encoder_layers_6_self_attn_q_norm_weight = const()[name = string("encoder_layers_6_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750208)))];
            tensor<fp16, [128]> encoder_layers_6_self_attn_k_norm_weight = const()[name = string("encoder_layers_6_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_6_mlp_gate_proj_weight = const()[name = string("encoder_layers_6_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(188750848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_6_mlp_up_proj_weight = const()[name = string("encoder_layers_6_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195042368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_6_mlp_down_proj_weight = const()[name = string("encoder_layers_6_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201333888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_7_self_attn_q_proj_weight = const()[name = string("encoder_layers_7_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207625408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_7_self_attn_k_proj_weight = const()[name = string("encoder_layers_7_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211819776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_7_self_attn_v_proj_weight = const()[name = string("encoder_layers_7_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(213916992)))];
            tensor<fp16, [128]> encoder_layers_7_self_attn_q_norm_weight = const()[name = string("encoder_layers_7_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014208)))];
            tensor<fp16, [128]> encoder_layers_7_self_attn_k_norm_weight = const()[name = string("encoder_layers_7_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_7_mlp_gate_proj_weight = const()[name = string("encoder_layers_7_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216014848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_7_mlp_up_proj_weight = const()[name = string("encoder_layers_7_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222306368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_7_mlp_down_proj_weight = const()[name = string("encoder_layers_7_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228597888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_8_self_attn_q_proj_weight = const()[name = string("encoder_layers_8_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234889408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_8_self_attn_k_proj_weight = const()[name = string("encoder_layers_8_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239083776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_8_self_attn_v_proj_weight = const()[name = string("encoder_layers_8_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241180992)))];
            tensor<fp16, [128]> encoder_layers_8_self_attn_q_norm_weight = const()[name = string("encoder_layers_8_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278208)))];
            tensor<fp16, [128]> encoder_layers_8_self_attn_k_norm_weight = const()[name = string("encoder_layers_8_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_8_mlp_gate_proj_weight = const()[name = string("encoder_layers_8_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243278848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_8_mlp_up_proj_weight = const()[name = string("encoder_layers_8_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249570368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_8_mlp_down_proj_weight = const()[name = string("encoder_layers_8_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255861888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_9_self_attn_q_proj_weight = const()[name = string("encoder_layers_9_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262153408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_9_self_attn_k_proj_weight = const()[name = string("encoder_layers_9_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266347776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_9_self_attn_v_proj_weight = const()[name = string("encoder_layers_9_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268444992)))];
            tensor<fp16, [128]> encoder_layers_9_self_attn_q_norm_weight = const()[name = string("encoder_layers_9_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542208)))];
            tensor<fp16, [128]> encoder_layers_9_self_attn_k_norm_weight = const()[name = string("encoder_layers_9_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_9_mlp_gate_proj_weight = const()[name = string("encoder_layers_9_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270542848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_9_mlp_up_proj_weight = const()[name = string("encoder_layers_9_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276834368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_9_mlp_down_proj_weight = const()[name = string("encoder_layers_9_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283125888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_10_self_attn_q_proj_weight = const()[name = string("encoder_layers_10_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289417408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_10_self_attn_k_proj_weight = const()[name = string("encoder_layers_10_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293611776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_10_self_attn_v_proj_weight = const()[name = string("encoder_layers_10_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295708992)))];
            tensor<fp16, [128]> encoder_layers_10_self_attn_q_norm_weight = const()[name = string("encoder_layers_10_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806208)))];
            tensor<fp16, [128]> encoder_layers_10_self_attn_k_norm_weight = const()[name = string("encoder_layers_10_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_10_mlp_gate_proj_weight = const()[name = string("encoder_layers_10_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297806848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_10_mlp_up_proj_weight = const()[name = string("encoder_layers_10_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304098368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_10_mlp_down_proj_weight = const()[name = string("encoder_layers_10_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310389888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_11_self_attn_q_proj_weight = const()[name = string("encoder_layers_11_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316681408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_11_self_attn_k_proj_weight = const()[name = string("encoder_layers_11_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320875776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_11_self_attn_v_proj_weight = const()[name = string("encoder_layers_11_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322972992)))];
            tensor<fp16, [128]> encoder_layers_11_self_attn_q_norm_weight = const()[name = string("encoder_layers_11_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070208)))];
            tensor<fp16, [128]> encoder_layers_11_self_attn_k_norm_weight = const()[name = string("encoder_layers_11_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_11_mlp_gate_proj_weight = const()[name = string("encoder_layers_11_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325070848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_11_mlp_up_proj_weight = const()[name = string("encoder_layers_11_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331362368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_11_mlp_down_proj_weight = const()[name = string("encoder_layers_11_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337653888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_12_self_attn_q_proj_weight = const()[name = string("encoder_layers_12_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343945408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_12_self_attn_k_proj_weight = const()[name = string("encoder_layers_12_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348139776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_12_self_attn_v_proj_weight = const()[name = string("encoder_layers_12_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350236992)))];
            tensor<fp16, [128]> encoder_layers_12_self_attn_q_norm_weight = const()[name = string("encoder_layers_12_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334208)))];
            tensor<fp16, [128]> encoder_layers_12_self_attn_k_norm_weight = const()[name = string("encoder_layers_12_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_12_mlp_gate_proj_weight = const()[name = string("encoder_layers_12_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352334848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_12_mlp_up_proj_weight = const()[name = string("encoder_layers_12_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358626368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_12_mlp_down_proj_weight = const()[name = string("encoder_layers_12_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364917888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_13_self_attn_q_proj_weight = const()[name = string("encoder_layers_13_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371209408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_13_self_attn_k_proj_weight = const()[name = string("encoder_layers_13_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375403776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_13_self_attn_v_proj_weight = const()[name = string("encoder_layers_13_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377500992)))];
            tensor<fp16, [128]> encoder_layers_13_self_attn_q_norm_weight = const()[name = string("encoder_layers_13_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598208)))];
            tensor<fp16, [128]> encoder_layers_13_self_attn_k_norm_weight = const()[name = string("encoder_layers_13_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_13_mlp_gate_proj_weight = const()[name = string("encoder_layers_13_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379598848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_13_mlp_up_proj_weight = const()[name = string("encoder_layers_13_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385890368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_13_mlp_down_proj_weight = const()[name = string("encoder_layers_13_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392181888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_14_self_attn_q_proj_weight = const()[name = string("encoder_layers_14_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398473408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_14_self_attn_k_proj_weight = const()[name = string("encoder_layers_14_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402667776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_14_self_attn_v_proj_weight = const()[name = string("encoder_layers_14_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404764992)))];
            tensor<fp16, [128]> encoder_layers_14_self_attn_q_norm_weight = const()[name = string("encoder_layers_14_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862208)))];
            tensor<fp16, [128]> encoder_layers_14_self_attn_k_norm_weight = const()[name = string("encoder_layers_14_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_14_mlp_gate_proj_weight = const()[name = string("encoder_layers_14_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406862848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_14_mlp_up_proj_weight = const()[name = string("encoder_layers_14_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(413154368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_14_mlp_down_proj_weight = const()[name = string("encoder_layers_14_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419445888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_15_self_attn_q_proj_weight = const()[name = string("encoder_layers_15_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425737408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_15_self_attn_k_proj_weight = const()[name = string("encoder_layers_15_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429931776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_15_self_attn_v_proj_weight = const()[name = string("encoder_layers_15_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432028992)))];
            tensor<fp16, [128]> encoder_layers_15_self_attn_q_norm_weight = const()[name = string("encoder_layers_15_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126208)))];
            tensor<fp16, [128]> encoder_layers_15_self_attn_k_norm_weight = const()[name = string("encoder_layers_15_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_15_mlp_gate_proj_weight = const()[name = string("encoder_layers_15_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434126848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_15_mlp_up_proj_weight = const()[name = string("encoder_layers_15_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440418368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_15_mlp_down_proj_weight = const()[name = string("encoder_layers_15_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446709888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_16_self_attn_q_proj_weight = const()[name = string("encoder_layers_16_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453001408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_16_self_attn_k_proj_weight = const()[name = string("encoder_layers_16_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457195776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_16_self_attn_v_proj_weight = const()[name = string("encoder_layers_16_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459292992)))];
            tensor<fp16, [128]> encoder_layers_16_self_attn_q_norm_weight = const()[name = string("encoder_layers_16_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390208)))];
            tensor<fp16, [128]> encoder_layers_16_self_attn_k_norm_weight = const()[name = string("encoder_layers_16_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_16_mlp_gate_proj_weight = const()[name = string("encoder_layers_16_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461390848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_16_mlp_up_proj_weight = const()[name = string("encoder_layers_16_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467682368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_16_mlp_down_proj_weight = const()[name = string("encoder_layers_16_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473973888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_17_self_attn_q_proj_weight = const()[name = string("encoder_layers_17_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480265408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_17_self_attn_k_proj_weight = const()[name = string("encoder_layers_17_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484459776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_17_self_attn_v_proj_weight = const()[name = string("encoder_layers_17_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486556992)))];
            tensor<fp16, [128]> encoder_layers_17_self_attn_q_norm_weight = const()[name = string("encoder_layers_17_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654208)))];
            tensor<fp16, [128]> encoder_layers_17_self_attn_k_norm_weight = const()[name = string("encoder_layers_17_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_17_mlp_gate_proj_weight = const()[name = string("encoder_layers_17_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488654848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_17_mlp_up_proj_weight = const()[name = string("encoder_layers_17_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494946368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_17_mlp_down_proj_weight = const()[name = string("encoder_layers_17_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501237888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_18_self_attn_q_proj_weight = const()[name = string("encoder_layers_18_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(507529408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_18_self_attn_k_proj_weight = const()[name = string("encoder_layers_18_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511723776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_18_self_attn_v_proj_weight = const()[name = string("encoder_layers_18_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513820992)))];
            tensor<fp16, [128]> encoder_layers_18_self_attn_q_norm_weight = const()[name = string("encoder_layers_18_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918208)))];
            tensor<fp16, [128]> encoder_layers_18_self_attn_k_norm_weight = const()[name = string("encoder_layers_18_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_18_mlp_gate_proj_weight = const()[name = string("encoder_layers_18_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515918848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_18_mlp_up_proj_weight = const()[name = string("encoder_layers_18_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522210368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_18_mlp_down_proj_weight = const()[name = string("encoder_layers_18_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528501888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_19_self_attn_q_proj_weight = const()[name = string("encoder_layers_19_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534793408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_19_self_attn_k_proj_weight = const()[name = string("encoder_layers_19_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538987776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_19_self_attn_v_proj_weight = const()[name = string("encoder_layers_19_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541084992)))];
            tensor<fp16, [128]> encoder_layers_19_self_attn_q_norm_weight = const()[name = string("encoder_layers_19_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182208)))];
            tensor<fp16, [128]> encoder_layers_19_self_attn_k_norm_weight = const()[name = string("encoder_layers_19_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_19_mlp_gate_proj_weight = const()[name = string("encoder_layers_19_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543182848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_19_mlp_up_proj_weight = const()[name = string("encoder_layers_19_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549474368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_19_mlp_down_proj_weight = const()[name = string("encoder_layers_19_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555765888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_20_self_attn_q_proj_weight = const()[name = string("encoder_layers_20_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562057408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_20_self_attn_k_proj_weight = const()[name = string("encoder_layers_20_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566251776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_20_self_attn_v_proj_weight = const()[name = string("encoder_layers_20_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568348992)))];
            tensor<fp16, [128]> encoder_layers_20_self_attn_q_norm_weight = const()[name = string("encoder_layers_20_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446208)))];
            tensor<fp16, [128]> encoder_layers_20_self_attn_k_norm_weight = const()[name = string("encoder_layers_20_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_20_mlp_gate_proj_weight = const()[name = string("encoder_layers_20_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570446848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_20_mlp_up_proj_weight = const()[name = string("encoder_layers_20_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576738368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_20_mlp_down_proj_weight = const()[name = string("encoder_layers_20_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583029888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_21_self_attn_q_proj_weight = const()[name = string("encoder_layers_21_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589321408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_21_self_attn_k_proj_weight = const()[name = string("encoder_layers_21_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593515776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_21_self_attn_v_proj_weight = const()[name = string("encoder_layers_21_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595612992)))];
            tensor<fp16, [128]> encoder_layers_21_self_attn_q_norm_weight = const()[name = string("encoder_layers_21_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710208)))];
            tensor<fp16, [128]> encoder_layers_21_self_attn_k_norm_weight = const()[name = string("encoder_layers_21_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_21_mlp_gate_proj_weight = const()[name = string("encoder_layers_21_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597710848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_21_mlp_up_proj_weight = const()[name = string("encoder_layers_21_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604002368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_21_mlp_down_proj_weight = const()[name = string("encoder_layers_21_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610293888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_22_self_attn_q_proj_weight = const()[name = string("encoder_layers_22_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616585408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_22_self_attn_k_proj_weight = const()[name = string("encoder_layers_22_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620779776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_22_self_attn_v_proj_weight = const()[name = string("encoder_layers_22_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622876992)))];
            tensor<fp16, [128]> encoder_layers_22_self_attn_q_norm_weight = const()[name = string("encoder_layers_22_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974208)))];
            tensor<fp16, [128]> encoder_layers_22_self_attn_k_norm_weight = const()[name = string("encoder_layers_22_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_22_mlp_gate_proj_weight = const()[name = string("encoder_layers_22_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624974848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_22_mlp_up_proj_weight = const()[name = string("encoder_layers_22_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631266368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_22_mlp_down_proj_weight = const()[name = string("encoder_layers_22_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637557888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_23_self_attn_q_proj_weight = const()[name = string("encoder_layers_23_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(643849408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_23_self_attn_k_proj_weight = const()[name = string("encoder_layers_23_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648043776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_23_self_attn_v_proj_weight = const()[name = string("encoder_layers_23_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(650140992)))];
            tensor<fp16, [128]> encoder_layers_23_self_attn_q_norm_weight = const()[name = string("encoder_layers_23_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238208)))];
            tensor<fp16, [128]> encoder_layers_23_self_attn_k_norm_weight = const()[name = string("encoder_layers_23_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_23_mlp_gate_proj_weight = const()[name = string("encoder_layers_23_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652238848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_23_mlp_up_proj_weight = const()[name = string("encoder_layers_23_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(658530368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_23_mlp_down_proj_weight = const()[name = string("encoder_layers_23_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664821888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_24_self_attn_q_proj_weight = const()[name = string("encoder_layers_24_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671113408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_24_self_attn_k_proj_weight = const()[name = string("encoder_layers_24_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675307776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_24_self_attn_v_proj_weight = const()[name = string("encoder_layers_24_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677404992)))];
            tensor<fp16, [128]> encoder_layers_24_self_attn_q_norm_weight = const()[name = string("encoder_layers_24_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502208)))];
            tensor<fp16, [128]> encoder_layers_24_self_attn_k_norm_weight = const()[name = string("encoder_layers_24_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_24_mlp_gate_proj_weight = const()[name = string("encoder_layers_24_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679502848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_24_mlp_up_proj_weight = const()[name = string("encoder_layers_24_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(685794368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_24_mlp_down_proj_weight = const()[name = string("encoder_layers_24_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692085888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_25_self_attn_q_proj_weight = const()[name = string("encoder_layers_25_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698377408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_25_self_attn_k_proj_weight = const()[name = string("encoder_layers_25_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702571776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_25_self_attn_v_proj_weight = const()[name = string("encoder_layers_25_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704668992)))];
            tensor<fp16, [128]> encoder_layers_25_self_attn_q_norm_weight = const()[name = string("encoder_layers_25_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766208)))];
            tensor<fp16, [128]> encoder_layers_25_self_attn_k_norm_weight = const()[name = string("encoder_layers_25_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_25_mlp_gate_proj_weight = const()[name = string("encoder_layers_25_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706766848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_25_mlp_up_proj_weight = const()[name = string("encoder_layers_25_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713058368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_25_mlp_down_proj_weight = const()[name = string("encoder_layers_25_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(719349888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_26_self_attn_q_proj_weight = const()[name = string("encoder_layers_26_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725641408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_26_self_attn_k_proj_weight = const()[name = string("encoder_layers_26_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729835776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_26_self_attn_v_proj_weight = const()[name = string("encoder_layers_26_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731932992)))];
            tensor<fp16, [128]> encoder_layers_26_self_attn_q_norm_weight = const()[name = string("encoder_layers_26_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030208)))];
            tensor<fp16, [128]> encoder_layers_26_self_attn_k_norm_weight = const()[name = string("encoder_layers_26_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_26_mlp_gate_proj_weight = const()[name = string("encoder_layers_26_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734030848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_26_mlp_up_proj_weight = const()[name = string("encoder_layers_26_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740322368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_26_mlp_down_proj_weight = const()[name = string("encoder_layers_26_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746613888)))];
            tensor<fp16, [2048, 1024, 1, 1]> encoder_layers_27_self_attn_q_proj_weight = const()[name = string("encoder_layers_27_self_attn_q_proj_weight"), val = tensor<fp16, [2048, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752905408)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_27_self_attn_k_proj_weight = const()[name = string("encoder_layers_27_self_attn_k_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(757099776)))];
            tensor<fp16, [1024, 1024, 1, 1]> encoder_layers_27_self_attn_v_proj_weight = const()[name = string("encoder_layers_27_self_attn_v_proj_weight"), val = tensor<fp16, [1024, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(759196992)))];
            tensor<fp16, [128]> encoder_layers_27_self_attn_q_norm_weight = const()[name = string("encoder_layers_27_self_attn_q_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294208)))];
            tensor<fp16, [128]> encoder_layers_27_self_attn_k_norm_weight = const()[name = string("encoder_layers_27_self_attn_k_norm_weight"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294528)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_27_mlp_gate_proj_weight = const()[name = string("encoder_layers_27_mlp_gate_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761294848)))];
            tensor<fp16, [3072, 1024, 1, 1]> encoder_layers_27_mlp_up_proj_weight = const()[name = string("encoder_layers_27_mlp_up_proj_weight"), val = tensor<fp16, [3072, 1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767586368)))];
            tensor<fp16, [1024, 3072, 1, 1]> encoder_layers_27_mlp_down_proj_weight = const()[name = string("encoder_layers_27_mlp_down_proj_weight"), val = tensor<fp16, [1024, 3072, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773877888)))];
            int32 var_18 = const()[name = string("op_18"), val = int32(-1)];
            int32 var_20 = const()[name = string("op_20"), val = int32(1)];
            int32 var_83_batch_dims_0 = const()[name = string("op_83_batch_dims_0"), val = int32(0)];
            bool var_83_validate_indices_0 = const()[name = string("op_83_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> encoder_embed_tokens_weight_to_fp16 = const()[name = string("encoder_embed_tokens_weight_to_fp16"), val = tensor<fp16, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(780169408)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 512]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 512]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 var_83_cast_fp16_axis_0 = const()[name = string("op_83_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 512, 1024]> var_83_cast_fp16 = gather(axis = var_83_cast_fp16_axis_0, batch_dims = var_83_batch_dims_0, indices = select_0, validate_indices = var_83_validate_indices_0, x = encoder_embed_tokens_weight_to_fp16)[name = string("op_83_cast_fp16")];
            fp16 fill_like_0_value_0_to_fp16 = const()[name = string("fill_like_0_value_0_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 512]> fill_like_0_cast_fp16 = fill_like(ref_tensor = attention_mask, value = fill_like_0_value_0_to_fp16)[name = string("fill_like_0_cast_fp16")];
            bool var_86_exclusive_0 = const()[name = string("op_86_exclusive_0"), val = bool(false)];
            bool var_86_reverse_0 = const()[name = string("op_86_reverse_0"), val = bool(false)];
            tensor<fp16, [1, 512]> var_86_cast_fp16 = cumsum(axis = var_20, exclusive = var_86_exclusive_0, reverse = var_86_reverse_0, x = fill_like_0_cast_fp16)[name = string("op_86_cast_fp16")];
            fp16 var_87_to_fp16 = const()[name = string("op_87_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 512]> var_88_cast_fp16 = sub(x = var_86_cast_fp16, y = var_87_to_fp16)[name = string("op_88_cast_fp16")];
            string position_ids_dtype_0 = const()[name = string("position_ids_dtype_0"), val = string("int32")];
            tensor<int32, [2]> pos_begin_0 = const()[name = string("pos_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> pos_end_0 = const()[name = string("pos_end_0"), val = tensor<int32, [2]>([1, 512])];
            tensor<bool, [2]> pos_end_mask_0 = const()[name = string("pos_end_mask_0"), val = tensor<bool, [2]>([false, true])];
            tensor<bool, [2]> pos_squeeze_mask_0 = const()[name = string("pos_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
            tensor<int32, [1, 512]> var_88_cast_fp16_to_int32 = cast(dtype = position_ids_dtype_0, x = var_88_cast_fp16)[name = string("cast_231")];
            tensor<int32, [512]> pos = slice_by_index(begin = pos_begin_0, end = pos_end_0, end_mask = pos_end_mask_0, squeeze_mask = pos_squeeze_mask_0, x = var_88_cast_fp16_to_int32)[name = string("pos")];
            int32 var_91_batch_dims_0 = const()[name = string("op_91_batch_dims_0"), val = int32(0)];
            bool var_91_validate_indices_0 = const()[name = string("op_91_validate_indices_0"), val = bool(false)];
            int32 greater_equal_1_y_0 = const()[name = string("greater_equal_1_y_0"), val = int32(0)];
            tensor<bool, [512]> greater_equal_1 = greater_equal(x = pos, y = greater_equal_1_y_0)[name = string("greater_equal_1")];
            int32 slice_by_index_1 = const()[name = string("slice_by_index_1"), val = int32(32768)];
            tensor<int32, [512]> add_1 = add(x = pos, y = slice_by_index_1)[name = string("add_1")];
            tensor<int32, [512]> select_1 = select(a = pos, b = add_1, cond = greater_equal_1)[name = string("select_1")];
            int32 var_91_axis_0 = const()[name = string("op_91_axis_0"), val = int32(0)];
            tensor<fp16, [512, 128]> var_91 = gather(axis = var_91_axis_0, batch_dims = var_91_batch_dims_0, indices = select_1, validate_indices = var_91_validate_indices_0, x = encoder_cos_cached)[name = string("op_91")];
            tensor<int32, [4]> var_92 = const()[name = string("op_92"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<fp16, [1, 1, 512, 128]> cos = reshape(shape = var_92, x = var_91)[name = string("cos")];
            int32 var_94_batch_dims_0 = const()[name = string("op_94_batch_dims_0"), val = int32(0)];
            bool var_94_validate_indices_0 = const()[name = string("op_94_validate_indices_0"), val = bool(false)];
            int32 var_94_axis_0 = const()[name = string("op_94_axis_0"), val = int32(0)];
            tensor<fp16, [512, 128]> var_94 = gather(axis = var_94_axis_0, batch_dims = var_94_batch_dims_0, indices = select_1, validate_indices = var_94_validate_indices_0, x = encoder_sin_cached)[name = string("op_94")];
            tensor<int32, [4]> var_95 = const()[name = string("op_95"), val = tensor<int32, [4]>([1, 1, 512, 128])];
            tensor<fp16, [1, 1, 512, 128]> sin = reshape(shape = var_95, x = var_94)[name = string("sin")];
            fp16 var_12_to_fp16 = const()[name = string("op_12_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 512]> var_98_cast_fp16 = sub(x = var_12_to_fp16, y = attention_mask)[name = string("op_98_cast_fp16")];
            fp16 var_100_to_fp16 = const()[name = string("op_100_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 512]> key_pad_cast_fp16 = mul(x = var_98_cast_fp16, y = var_100_to_fp16)[name = string("key_pad_cast_fp16")];
            tensor<int32, [4]> var_102 = const()[name = string("op_102"), val = tensor<int32, [4]>([1, 1, 1, 512])];
            tensor<fp16, [1, 1, 1, 512]> var_103_cast_fp16 = reshape(shape = var_102, x = key_pad_cast_fp16)[name = string("op_103_cast_fp16")];
            tensor<int32, [4]> causal_mask_reps_0 = const()[name = string("causal_mask_reps_0"), val = tensor<int32, [4]>([1, 1, 512, 1])];
            tensor<fp16, [1, 1, 512, 512]> causal_mask_cast_fp16 = tile(reps = causal_mask_reps_0, x = var_103_cast_fp16)[name = string("causal_mask_cast_fp16")];
            fp16 var_6_promoted_to_fp16 = const()[name = string("op_6_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_129_cast_fp16 = pow(x = var_83_cast_fp16, y = var_6_promoted_to_fp16)[name = string("op_129_cast_fp16")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = string("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1_keep_dims_0 = const()[name = string("var_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_1_cast_fp16 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_129_cast_fp16)[name = string("var_1_cast_fp16")];
            fp16 var_132_to_fp16 = const()[name = string("op_132_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_133_cast_fp16 = add(x = var_1_cast_fp16, y = var_132_to_fp16)[name = string("op_133_cast_fp16")];
            fp32 var_134_epsilon_0 = const()[name = string("op_134_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_134_cast_fp16 = rsqrt(epsilon = var_134_epsilon_0, x = var_133_cast_fp16)[name = string("op_134_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_3_cast_fp16 = mul(x = var_83_cast_fp16, y = var_134_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_0_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_0_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1091334400)))];
            tensor<fp16, [1, 512, 1024]> var_137_cast_fp16 = mul(x = x_3_cast_fp16, y = encoder_layers_0_input_layernorm_weight_promoted_to_fp16)[name = string("op_137_cast_fp16")];
            tensor<int32, [3]> var_142 = const()[name = string("op_142"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_143 = transpose(perm = var_142, x = var_137_cast_fp16)[name = string("transpose_251")];
            tensor<fp16, [1, 1024, 1, 512]> input_1 = expand_dims(axes = input_1_axes_0, x = var_143)[name = string("input_1")];
            string var_150_pad_type_0 = const()[name = string("op_150_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_150_strides_0 = const()[name = string("op_150_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_150_pad_0 = const()[name = string("op_150_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_150_dilations_0 = const()[name = string("op_150_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_150_groups_0 = const()[name = string("op_150_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_150 = conv(dilations = var_150_dilations_0, groups = var_150_groups_0, pad = var_150_pad_0, pad_type = var_150_pad_type_0, strides = var_150_strides_0, weight = encoder_layers_0_self_attn_q_proj_weight, x = input_1)[name = string("op_150")];
            tensor<int32, [4]> var_151 = const()[name = string("op_151"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_152 = reshape(shape = var_151, x = var_150)[name = string("op_152")];
            tensor<int32, [4]> var_153 = const()[name = string("op_153"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_160_pad_type_0 = const()[name = string("op_160_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_160_strides_0 = const()[name = string("op_160_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_160_pad_0 = const()[name = string("op_160_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_160_dilations_0 = const()[name = string("op_160_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_160_groups_0 = const()[name = string("op_160_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_160 = conv(dilations = var_160_dilations_0, groups = var_160_groups_0, pad = var_160_pad_0, pad_type = var_160_pad_type_0, strides = var_160_strides_0, weight = encoder_layers_0_self_attn_k_proj_weight, x = input_1)[name = string("op_160")];
            tensor<int32, [4]> var_161 = const()[name = string("op_161"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_162 = reshape(shape = var_161, x = var_160)[name = string("op_162")];
            tensor<int32, [4]> var_163 = const()[name = string("op_163"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_170_pad_type_0 = const()[name = string("op_170_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_170_strides_0 = const()[name = string("op_170_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_170_pad_0 = const()[name = string("op_170_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_170_dilations_0 = const()[name = string("op_170_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_170_groups_0 = const()[name = string("op_170_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_170 = conv(dilations = var_170_dilations_0, groups = var_170_groups_0, pad = var_170_pad_0, pad_type = var_170_pad_type_0, strides = var_170_strides_0, weight = encoder_layers_0_self_attn_v_proj_weight, x = input_1)[name = string("op_170")];
            tensor<int32, [4]> var_171 = const()[name = string("op_171"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_172 = reshape(shape = var_171, x = var_170)[name = string("op_172")];
            tensor<int32, [4]> var_173 = const()[name = string("op_173"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_1_to_fp16 = const()[name = string("op_6_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_1 = transpose(perm = var_153, x = var_152)[name = string("transpose_250")];
            tensor<fp16, [1, 16, 512, 128]> var_179_cast_fp16 = pow(x = q_1, y = var_6_promoted_1_to_fp16)[name = string("op_179_cast_fp16")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = string("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3_keep_dims_0 = const()[name = string("var_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_3_cast_fp16 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_179_cast_fp16)[name = string("var_3_cast_fp16")];
            fp16 var_182_to_fp16 = const()[name = string("op_182_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_183_cast_fp16 = add(x = var_3_cast_fp16, y = var_182_to_fp16)[name = string("op_183_cast_fp16")];
            fp32 var_184_epsilon_0 = const()[name = string("op_184_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_184_cast_fp16 = rsqrt(epsilon = var_184_epsilon_0, x = var_183_cast_fp16)[name = string("op_184_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_11_cast_fp16 = mul(x = q_1, y = var_184_cast_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_3 = mul(x = x_11_cast_fp16, y = encoder_layers_0_self_attn_q_norm_weight)[name = string("q_3")];
            fp16 var_6_promoted_2_to_fp16 = const()[name = string("op_6_promoted_2_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_1 = transpose(perm = var_163, x = var_162)[name = string("transpose_249")];
            tensor<fp16, [1, 8, 512, 128]> var_192_cast_fp16 = pow(x = k_1, y = var_6_promoted_2_to_fp16)[name = string("op_192_cast_fp16")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = string("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5_keep_dims_0 = const()[name = string("var_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_5_cast_fp16 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_192_cast_fp16)[name = string("var_5_cast_fp16")];
            fp16 var_195_to_fp16 = const()[name = string("op_195_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_196_cast_fp16 = add(x = var_5_cast_fp16, y = var_195_to_fp16)[name = string("op_196_cast_fp16")];
            fp32 var_197_epsilon_0 = const()[name = string("op_197_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_197_cast_fp16 = rsqrt(epsilon = var_197_epsilon_0, x = var_196_cast_fp16)[name = string("op_197_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_17_cast_fp16 = mul(x = k_1, y = var_197_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_3 = mul(x = x_17_cast_fp16, y = encoder_layers_0_self_attn_k_norm_weight)[name = string("k_3")];
            tensor<fp16, [1, 16, 512, 128]> var_201 = mul(x = q_3, y = cos)[name = string("op_201")];
            tensor<int32, [2]> var_202_split_sizes_0 = const()[name = string("op_202_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_202_axis_0 = const()[name = string("op_202_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_202_0, tensor<fp16, [1, 16, 512, 64]> var_202_1 = split(axis = var_202_axis_0, split_sizes = var_202_split_sizes_0, x = q_3)[name = string("op_202")];
            fp16 const_3_promoted = const()[name = string("const_3_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_204 = mul(x = var_202_1, y = const_3_promoted)[name = string("op_204")];
            bool var_206_interleave_0 = const()[name = string("op_206_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_206 = concat(axis = var_18, interleave = var_206_interleave_0, values = (var_204, var_202_0))[name = string("op_206")];
            tensor<fp16, [1, 16, 512, 128]> var_207 = mul(x = var_206, y = sin)[name = string("op_207")];
            tensor<fp16, [1, 16, 512, 128]> query_1 = add(x = var_201, y = var_207)[name = string("query_1")];
            tensor<fp16, [1, 8, 512, 128]> var_209 = mul(x = k_3, y = cos)[name = string("op_209")];
            tensor<int32, [2]> var_210_split_sizes_0 = const()[name = string("op_210_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_210_axis_0 = const()[name = string("op_210_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_210_0, tensor<fp16, [1, 8, 512, 64]> var_210_1 = split(axis = var_210_axis_0, split_sizes = var_210_split_sizes_0, x = k_3)[name = string("op_210")];
            fp16 const_4_promoted = const()[name = string("const_4_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_212 = mul(x = var_210_1, y = const_4_promoted)[name = string("op_212")];
            bool var_214_interleave_0 = const()[name = string("op_214_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_214 = concat(axis = var_18, interleave = var_214_interleave_0, values = (var_212, var_210_0))[name = string("op_214")];
            tensor<fp16, [1, 8, 512, 128]> var_215 = mul(x = var_214, y = sin)[name = string("op_215")];
            tensor<fp16, [1, 8, 512, 128]> x_19 = add(x = var_209, y = var_215)[name = string("x_19")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = string("op_217_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_217 = expand_dims(axes = var_217_axes_0, x = x_19)[name = string("op_217")];
            tensor<int32, [5]> x_21_reps_0 = const()[name = string("x_21_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_21 = tile(reps = x_21_reps_0, x = var_217)[name = string("x_21")];
            tensor<int32, [4]> var_220 = const()[name = string("op_220"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_1 = reshape(shape = var_220, x = x_21)[name = string("key_1")];
            tensor<int32, [1]> var_222_axes_0 = const()[name = string("op_222_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_23 = transpose(perm = var_173, x = var_172)[name = string("transpose_248")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_222 = expand_dims(axes = var_222_axes_0, x = x_23)[name = string("op_222")];
            tensor<int32, [5]> x_25_reps_0 = const()[name = string("x_25_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_25 = tile(reps = x_25_reps_0, x = var_222)[name = string("x_25")];
            tensor<int32, [4]> var_225 = const()[name = string("op_225"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_1 = reshape(shape = var_225, x = x_25)[name = string("value_1")];
            bool var_230_transpose_x_1 = const()[name = string("op_230_transpose_x_1"), val = bool(false)];
            bool var_230_transpose_y_1 = const()[name = string("op_230_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_230_cast_fp16 = matmul(transpose_x = var_230_transpose_x_1, transpose_y = var_230_transpose_y_1, x = query_1, y = key_1)[name = string("op_230_cast_fp16")];
            fp16 var_231_to_fp16 = const()[name = string("op_231_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_1_cast_fp16 = mul(x = var_230_cast_fp16, y = var_231_to_fp16)[name = string("attn_weights_1_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_3_cast_fp16 = add(x = attn_weights_1_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_3_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_235_cast_fp16 = softmax(axis = var_18, x = attn_weights_3_cast_fp16)[name = string("op_235_cast_fp16")];
            bool var_239_transpose_x_0 = const()[name = string("op_239_transpose_x_0"), val = bool(false)];
            bool var_239_transpose_y_0 = const()[name = string("op_239_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_239_cast_fp16 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = var_235_cast_fp16, y = value_1)[name = string("op_239_cast_fp16")];
            tensor<int32, [4]> var_241 = const()[name = string("op_241"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_244 = const()[name = string("op_244"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_242 = transpose(perm = var_241, x = var_239_cast_fp16)[name = string("transpose_247")];
            tensor<fp16, [1, 512, 2048]> attn_out_3 = reshape(shape = var_244, x = var_242)[name = string("attn_out_3")];
            tensor<int32, [3]> var_246 = const()[name = string("op_246"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_0 = const()[name = string("squeeze_0"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1091336512)))];
            string var_255_pad_type_0 = const()[name = string("op_255_pad_type_0"), val = string("valid")];
            int32 var_255_groups_0 = const()[name = string("op_255_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_255_strides_0 = const()[name = string("op_255_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_255_pad_0 = const()[name = string("op_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_255_dilations_0 = const()[name = string("op_255_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_247 = transpose(perm = var_246, x = attn_out_3)[name = string("transpose_246")];
            tensor<fp16, [1, 1024, 512]> var_255 = conv(dilations = var_255_dilations_0, groups = var_255_groups_0, pad = var_255_pad_0, pad_type = var_255_pad_type_0, strides = var_255_strides_0, weight = squeeze_0, x = var_247)[name = string("op_255")];
            tensor<int32, [3]> var_256 = const()[name = string("op_256"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_5 = transpose(perm = var_256, x = var_255)[name = string("transpose_245")];
            tensor<fp16, [1, 512, 1024]> x_27_cast_fp16 = add(x = var_83_cast_fp16, y = attn_out_5)[name = string("x_27_cast_fp16")];
            fp16 var_6_promoted_3_to_fp16 = const()[name = string("op_6_promoted_3_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_262_cast_fp16 = pow(x = x_27_cast_fp16, y = var_6_promoted_3_to_fp16)[name = string("op_262_cast_fp16")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = string("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7_keep_dims_0 = const()[name = string("var_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_7_cast_fp16 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_262_cast_fp16)[name = string("var_7_cast_fp16")];
            fp16 var_265_to_fp16 = const()[name = string("op_265_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_266_cast_fp16 = add(x = var_7_cast_fp16, y = var_265_to_fp16)[name = string("op_266_cast_fp16")];
            fp32 var_267_epsilon_0 = const()[name = string("op_267_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_267_cast_fp16 = rsqrt(epsilon = var_267_epsilon_0, x = var_266_cast_fp16)[name = string("op_267_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_31_cast_fp16 = mul(x = x_27_cast_fp16, y = var_267_cast_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095530880)))];
            tensor<fp16, [1, 512, 1024]> var_270_cast_fp16 = mul(x = x_31_cast_fp16, y = encoder_layers_0_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_270_cast_fp16")];
            tensor<int32, [3]> var_275 = const()[name = string("op_275"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_276 = transpose(perm = var_275, x = var_270_cast_fp16)[name = string("transpose_244")];
            tensor<fp16, [1, 1024, 1, 512]> input_5 = expand_dims(axes = input_5_axes_0, x = var_276)[name = string("input_5")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_7 = conv(dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = encoder_layers_0_mlp_gate_proj_weight, x = input_5)[name = string("input_7")];
            string up_1_pad_type_0 = const()[name = string("up_1_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_1_strides_0 = const()[name = string("up_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_1_pad_0 = const()[name = string("up_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_1_dilations_0 = const()[name = string("up_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_1_groups_0 = const()[name = string("up_1_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_1 = conv(dilations = up_1_dilations_0, groups = up_1_groups_0, pad = up_1_pad_0, pad_type = up_1_pad_type_0, strides = up_1_strides_0, weight = encoder_layers_0_mlp_up_proj_weight, x = input_5)[name = string("up_1")];
            tensor<fp16, [1, 3072, 1, 512]> var_290 = silu(x = input_7)[name = string("op_290")];
            tensor<fp16, [1, 3072, 1, 512]> input_9 = mul(x = var_290, y = up_1)[name = string("input_9")];
            string var_297_pad_type_0 = const()[name = string("op_297_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_297_strides_0 = const()[name = string("op_297_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_297_pad_0 = const()[name = string("op_297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_297_dilations_0 = const()[name = string("op_297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_297_groups_0 = const()[name = string("op_297_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_297 = conv(dilations = var_297_dilations_0, groups = var_297_groups_0, pad = var_297_pad_0, pad_type = var_297_pad_type_0, strides = var_297_strides_0, weight = encoder_layers_0_mlp_down_proj_weight, x = input_9)[name = string("op_297")];
            tensor<int32, [1]> var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_298 = squeeze(axes = var_298_axes_0, x = var_297)[name = string("op_298")];
            tensor<int32, [3]> var_299 = const()[name = string("op_299"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_1 = transpose(perm = var_299, x = var_298)[name = string("transpose_243")];
            tensor<fp16, [1, 512, 1024]> hidden_states_3_cast_fp16 = add(x = x_27_cast_fp16, y = mlp_out_1)[name = string("hidden_states_3_cast_fp16")];
            fp16 var_6_promoted_4_to_fp16 = const()[name = string("op_6_promoted_4_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_326_cast_fp16 = pow(x = hidden_states_3_cast_fp16, y = var_6_promoted_4_to_fp16)[name = string("op_326_cast_fp16")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = string("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9_keep_dims_0 = const()[name = string("var_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_9_cast_fp16 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_326_cast_fp16)[name = string("var_9_cast_fp16")];
            fp16 var_329_to_fp16 = const()[name = string("op_329_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_330_cast_fp16 = add(x = var_9_cast_fp16, y = var_329_to_fp16)[name = string("op_330_cast_fp16")];
            fp32 var_331_epsilon_0 = const()[name = string("op_331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_331_cast_fp16 = rsqrt(epsilon = var_331_epsilon_0, x = var_330_cast_fp16)[name = string("op_331_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_37_cast_fp16 = mul(x = hidden_states_3_cast_fp16, y = var_331_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_1_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_1_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095532992)))];
            tensor<fp16, [1, 512, 1024]> var_334_cast_fp16 = mul(x = x_37_cast_fp16, y = encoder_layers_1_input_layernorm_weight_promoted_to_fp16)[name = string("op_334_cast_fp16")];
            tensor<int32, [3]> var_339 = const()[name = string("op_339"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_340 = transpose(perm = var_339, x = var_334_cast_fp16)[name = string("transpose_242")];
            tensor<fp16, [1, 1024, 1, 512]> input_11 = expand_dims(axes = input_11_axes_0, x = var_340)[name = string("input_11")];
            string var_347_pad_type_0 = const()[name = string("op_347_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_347_strides_0 = const()[name = string("op_347_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_347_pad_0 = const()[name = string("op_347_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_347_dilations_0 = const()[name = string("op_347_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_347_groups_0 = const()[name = string("op_347_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_347 = conv(dilations = var_347_dilations_0, groups = var_347_groups_0, pad = var_347_pad_0, pad_type = var_347_pad_type_0, strides = var_347_strides_0, weight = encoder_layers_1_self_attn_q_proj_weight, x = input_11)[name = string("op_347")];
            tensor<int32, [4]> var_348 = const()[name = string("op_348"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_349 = reshape(shape = var_348, x = var_347)[name = string("op_349")];
            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_357_pad_type_0 = const()[name = string("op_357_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_357_strides_0 = const()[name = string("op_357_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_357_pad_0 = const()[name = string("op_357_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_357_dilations_0 = const()[name = string("op_357_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_357_groups_0 = const()[name = string("op_357_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_357 = conv(dilations = var_357_dilations_0, groups = var_357_groups_0, pad = var_357_pad_0, pad_type = var_357_pad_type_0, strides = var_357_strides_0, weight = encoder_layers_1_self_attn_k_proj_weight, x = input_11)[name = string("op_357")];
            tensor<int32, [4]> var_358 = const()[name = string("op_358"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_359 = reshape(shape = var_358, x = var_357)[name = string("op_359")];
            tensor<int32, [4]> var_360 = const()[name = string("op_360"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_367_pad_type_0 = const()[name = string("op_367_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_367_strides_0 = const()[name = string("op_367_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_367_pad_0 = const()[name = string("op_367_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_367_dilations_0 = const()[name = string("op_367_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_367_groups_0 = const()[name = string("op_367_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_367 = conv(dilations = var_367_dilations_0, groups = var_367_groups_0, pad = var_367_pad_0, pad_type = var_367_pad_type_0, strides = var_367_strides_0, weight = encoder_layers_1_self_attn_v_proj_weight, x = input_11)[name = string("op_367")];
            tensor<int32, [4]> var_368 = const()[name = string("op_368"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_369 = reshape(shape = var_368, x = var_367)[name = string("op_369")];
            tensor<int32, [4]> var_370 = const()[name = string("op_370"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_5_to_fp16 = const()[name = string("op_6_promoted_5_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_7 = transpose(perm = var_350, x = var_349)[name = string("transpose_241")];
            tensor<fp16, [1, 16, 512, 128]> var_376_cast_fp16 = pow(x = q_7, y = var_6_promoted_5_to_fp16)[name = string("op_376_cast_fp16")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = string("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_11_keep_dims_0 = const()[name = string("var_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_11_cast_fp16 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_376_cast_fp16)[name = string("var_11_cast_fp16")];
            fp16 var_379_to_fp16 = const()[name = string("op_379_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_380_cast_fp16 = add(x = var_11_cast_fp16, y = var_379_to_fp16)[name = string("op_380_cast_fp16")];
            fp32 var_381_epsilon_0 = const()[name = string("op_381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_381_cast_fp16 = rsqrt(epsilon = var_381_epsilon_0, x = var_380_cast_fp16)[name = string("op_381_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_45_cast_fp16 = mul(x = q_7, y = var_381_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_9 = mul(x = x_45_cast_fp16, y = encoder_layers_1_self_attn_q_norm_weight)[name = string("q_9")];
            fp16 var_6_promoted_6_to_fp16 = const()[name = string("op_6_promoted_6_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_7 = transpose(perm = var_360, x = var_359)[name = string("transpose_240")];
            tensor<fp16, [1, 8, 512, 128]> var_389_cast_fp16 = pow(x = k_7, y = var_6_promoted_6_to_fp16)[name = string("op_389_cast_fp16")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = string("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_13_keep_dims_0 = const()[name = string("var_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_13_cast_fp16 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_389_cast_fp16)[name = string("var_13_cast_fp16")];
            fp16 var_392_to_fp16 = const()[name = string("op_392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_393_cast_fp16 = add(x = var_13_cast_fp16, y = var_392_to_fp16)[name = string("op_393_cast_fp16")];
            fp32 var_394_epsilon_0 = const()[name = string("op_394_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_394_cast_fp16 = rsqrt(epsilon = var_394_epsilon_0, x = var_393_cast_fp16)[name = string("op_394_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_51_cast_fp16 = mul(x = k_7, y = var_394_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_9 = mul(x = x_51_cast_fp16, y = encoder_layers_1_self_attn_k_norm_weight)[name = string("k_9")];
            tensor<fp16, [1, 16, 512, 128]> var_398 = mul(x = q_9, y = cos)[name = string("op_398")];
            tensor<int32, [2]> var_399_split_sizes_0 = const()[name = string("op_399_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_399_axis_0 = const()[name = string("op_399_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_399_0, tensor<fp16, [1, 16, 512, 64]> var_399_1 = split(axis = var_399_axis_0, split_sizes = var_399_split_sizes_0, x = q_9)[name = string("op_399")];
            fp16 const_6_promoted = const()[name = string("const_6_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_401 = mul(x = var_399_1, y = const_6_promoted)[name = string("op_401")];
            bool var_403_interleave_0 = const()[name = string("op_403_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_403 = concat(axis = var_18, interleave = var_403_interleave_0, values = (var_401, var_399_0))[name = string("op_403")];
            tensor<fp16, [1, 16, 512, 128]> var_404 = mul(x = var_403, y = sin)[name = string("op_404")];
            tensor<fp16, [1, 16, 512, 128]> query_3 = add(x = var_398, y = var_404)[name = string("query_3")];
            tensor<fp16, [1, 8, 512, 128]> var_406 = mul(x = k_9, y = cos)[name = string("op_406")];
            tensor<int32, [2]> var_407_split_sizes_0 = const()[name = string("op_407_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_407_axis_0 = const()[name = string("op_407_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_407_0, tensor<fp16, [1, 8, 512, 64]> var_407_1 = split(axis = var_407_axis_0, split_sizes = var_407_split_sizes_0, x = k_9)[name = string("op_407")];
            fp16 const_7_promoted = const()[name = string("const_7_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_409 = mul(x = var_407_1, y = const_7_promoted)[name = string("op_409")];
            bool var_411_interleave_0 = const()[name = string("op_411_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_411 = concat(axis = var_18, interleave = var_411_interleave_0, values = (var_409, var_407_0))[name = string("op_411")];
            tensor<fp16, [1, 8, 512, 128]> var_412 = mul(x = var_411, y = sin)[name = string("op_412")];
            tensor<fp16, [1, 8, 512, 128]> x_53 = add(x = var_406, y = var_412)[name = string("x_53")];
            tensor<int32, [1]> var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_414 = expand_dims(axes = var_414_axes_0, x = x_53)[name = string("op_414")];
            tensor<int32, [5]> x_55_reps_0 = const()[name = string("x_55_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_55 = tile(reps = x_55_reps_0, x = var_414)[name = string("x_55")];
            tensor<int32, [4]> var_417 = const()[name = string("op_417"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_3 = reshape(shape = var_417, x = x_55)[name = string("key_3")];
            tensor<int32, [1]> var_419_axes_0 = const()[name = string("op_419_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_57 = transpose(perm = var_370, x = var_369)[name = string("transpose_239")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_419 = expand_dims(axes = var_419_axes_0, x = x_57)[name = string("op_419")];
            tensor<int32, [5]> x_59_reps_0 = const()[name = string("x_59_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_59 = tile(reps = x_59_reps_0, x = var_419)[name = string("x_59")];
            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_3 = reshape(shape = var_422, x = x_59)[name = string("value_3")];
            bool var_427_transpose_x_1 = const()[name = string("op_427_transpose_x_1"), val = bool(false)];
            bool var_427_transpose_y_1 = const()[name = string("op_427_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_427_cast_fp16 = matmul(transpose_x = var_427_transpose_x_1, transpose_y = var_427_transpose_y_1, x = query_3, y = key_3)[name = string("op_427_cast_fp16")];
            fp16 var_428_to_fp16 = const()[name = string("op_428_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_7_cast_fp16 = mul(x = var_427_cast_fp16, y = var_428_to_fp16)[name = string("attn_weights_7_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_9_cast_fp16 = add(x = attn_weights_7_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_9_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_432_cast_fp16 = softmax(axis = var_18, x = attn_weights_9_cast_fp16)[name = string("op_432_cast_fp16")];
            bool var_436_transpose_x_0 = const()[name = string("op_436_transpose_x_0"), val = bool(false)];
            bool var_436_transpose_y_0 = const()[name = string("op_436_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_436_cast_fp16 = matmul(transpose_x = var_436_transpose_x_0, transpose_y = var_436_transpose_y_0, x = var_432_cast_fp16, y = value_3)[name = string("op_436_cast_fp16")];
            tensor<int32, [4]> var_438 = const()[name = string("op_438"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_441 = const()[name = string("op_441"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_439 = transpose(perm = var_438, x = var_436_cast_fp16)[name = string("transpose_238")];
            tensor<fp16, [1, 512, 2048]> attn_out_9 = reshape(shape = var_441, x = var_439)[name = string("attn_out_9")];
            tensor<int32, [3]> var_443 = const()[name = string("op_443"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_1 = const()[name = string("squeeze_1"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1095535104)))];
            string var_452_pad_type_0 = const()[name = string("op_452_pad_type_0"), val = string("valid")];
            int32 var_452_groups_0 = const()[name = string("op_452_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_452_strides_0 = const()[name = string("op_452_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_452_pad_0 = const()[name = string("op_452_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_452_dilations_0 = const()[name = string("op_452_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_444 = transpose(perm = var_443, x = attn_out_9)[name = string("transpose_237")];
            tensor<fp16, [1, 1024, 512]> var_452 = conv(dilations = var_452_dilations_0, groups = var_452_groups_0, pad = var_452_pad_0, pad_type = var_452_pad_type_0, strides = var_452_strides_0, weight = squeeze_1, x = var_444)[name = string("op_452")];
            tensor<int32, [3]> var_453 = const()[name = string("op_453"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_11 = transpose(perm = var_453, x = var_452)[name = string("transpose_236")];
            tensor<fp16, [1, 512, 1024]> x_61_cast_fp16 = add(x = hidden_states_3_cast_fp16, y = attn_out_11)[name = string("x_61_cast_fp16")];
            fp16 var_6_promoted_7_to_fp16 = const()[name = string("op_6_promoted_7_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_459_cast_fp16 = pow(x = x_61_cast_fp16, y = var_6_promoted_7_to_fp16)[name = string("op_459_cast_fp16")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = string("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_15_keep_dims_0 = const()[name = string("var_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_15_cast_fp16 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_459_cast_fp16)[name = string("var_15_cast_fp16")];
            fp16 var_462_to_fp16 = const()[name = string("op_462_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_463_cast_fp16 = add(x = var_15_cast_fp16, y = var_462_to_fp16)[name = string("op_463_cast_fp16")];
            fp32 var_464_epsilon_0 = const()[name = string("op_464_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_464_cast_fp16 = rsqrt(epsilon = var_464_epsilon_0, x = var_463_cast_fp16)[name = string("op_464_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_65_cast_fp16 = mul(x = x_61_cast_fp16, y = var_464_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099729472)))];
            tensor<fp16, [1, 512, 1024]> var_467_cast_fp16 = mul(x = x_65_cast_fp16, y = encoder_layers_1_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_467_cast_fp16")];
            tensor<int32, [3]> var_472 = const()[name = string("op_472"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_473 = transpose(perm = var_472, x = var_467_cast_fp16)[name = string("transpose_235")];
            tensor<fp16, [1, 1024, 1, 512]> input_15 = expand_dims(axes = input_15_axes_0, x = var_473)[name = string("input_15")];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_17 = conv(dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = encoder_layers_1_mlp_gate_proj_weight, x = input_15)[name = string("input_17")];
            string up_3_pad_type_0 = const()[name = string("up_3_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_3_strides_0 = const()[name = string("up_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_3_pad_0 = const()[name = string("up_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_3_dilations_0 = const()[name = string("up_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_3_groups_0 = const()[name = string("up_3_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_3 = conv(dilations = up_3_dilations_0, groups = up_3_groups_0, pad = up_3_pad_0, pad_type = up_3_pad_type_0, strides = up_3_strides_0, weight = encoder_layers_1_mlp_up_proj_weight, x = input_15)[name = string("up_3")];
            tensor<fp16, [1, 3072, 1, 512]> var_487 = silu(x = input_17)[name = string("op_487")];
            tensor<fp16, [1, 3072, 1, 512]> input_19 = mul(x = var_487, y = up_3)[name = string("input_19")];
            string var_494_pad_type_0 = const()[name = string("op_494_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_494_strides_0 = const()[name = string("op_494_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_494_pad_0 = const()[name = string("op_494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_494_dilations_0 = const()[name = string("op_494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_494_groups_0 = const()[name = string("op_494_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_494 = conv(dilations = var_494_dilations_0, groups = var_494_groups_0, pad = var_494_pad_0, pad_type = var_494_pad_type_0, strides = var_494_strides_0, weight = encoder_layers_1_mlp_down_proj_weight, x = input_19)[name = string("op_494")];
            tensor<int32, [1]> var_495_axes_0 = const()[name = string("op_495_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_495 = squeeze(axes = var_495_axes_0, x = var_494)[name = string("op_495")];
            tensor<int32, [3]> var_496 = const()[name = string("op_496"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_3 = transpose(perm = var_496, x = var_495)[name = string("transpose_234")];
            tensor<fp16, [1, 512, 1024]> hidden_states_5_cast_fp16 = add(x = x_61_cast_fp16, y = mlp_out_3)[name = string("hidden_states_5_cast_fp16")];
            fp16 var_6_promoted_8_to_fp16 = const()[name = string("op_6_promoted_8_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_523_cast_fp16 = pow(x = hidden_states_5_cast_fp16, y = var_6_promoted_8_to_fp16)[name = string("op_523_cast_fp16")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = string("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_17_keep_dims_0 = const()[name = string("var_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_17_cast_fp16 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_523_cast_fp16)[name = string("var_17_cast_fp16")];
            fp16 var_526_to_fp16 = const()[name = string("op_526_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_527_cast_fp16 = add(x = var_17_cast_fp16, y = var_526_to_fp16)[name = string("op_527_cast_fp16")];
            fp32 var_528_epsilon_0 = const()[name = string("op_528_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_528_cast_fp16 = rsqrt(epsilon = var_528_epsilon_0, x = var_527_cast_fp16)[name = string("op_528_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_71_cast_fp16 = mul(x = hidden_states_5_cast_fp16, y = var_528_cast_fp16)[name = string("x_71_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_2_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_2_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099731584)))];
            tensor<fp16, [1, 512, 1024]> var_531_cast_fp16 = mul(x = x_71_cast_fp16, y = encoder_layers_2_input_layernorm_weight_promoted_to_fp16)[name = string("op_531_cast_fp16")];
            tensor<int32, [3]> var_536 = const()[name = string("op_536"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_537 = transpose(perm = var_536, x = var_531_cast_fp16)[name = string("transpose_233")];
            tensor<fp16, [1, 1024, 1, 512]> input_21 = expand_dims(axes = input_21_axes_0, x = var_537)[name = string("input_21")];
            string var_544_pad_type_0 = const()[name = string("op_544_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_544_strides_0 = const()[name = string("op_544_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_544_pad_0 = const()[name = string("op_544_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_544_dilations_0 = const()[name = string("op_544_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_544_groups_0 = const()[name = string("op_544_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_544 = conv(dilations = var_544_dilations_0, groups = var_544_groups_0, pad = var_544_pad_0, pad_type = var_544_pad_type_0, strides = var_544_strides_0, weight = encoder_layers_2_self_attn_q_proj_weight, x = input_21)[name = string("op_544")];
            tensor<int32, [4]> var_545 = const()[name = string("op_545"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_546 = reshape(shape = var_545, x = var_544)[name = string("op_546")];
            tensor<int32, [4]> var_547 = const()[name = string("op_547"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_554_pad_type_0 = const()[name = string("op_554_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_554_strides_0 = const()[name = string("op_554_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_554_pad_0 = const()[name = string("op_554_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_554_dilations_0 = const()[name = string("op_554_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_554_groups_0 = const()[name = string("op_554_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_554 = conv(dilations = var_554_dilations_0, groups = var_554_groups_0, pad = var_554_pad_0, pad_type = var_554_pad_type_0, strides = var_554_strides_0, weight = encoder_layers_2_self_attn_k_proj_weight, x = input_21)[name = string("op_554")];
            tensor<int32, [4]> var_555 = const()[name = string("op_555"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_556 = reshape(shape = var_555, x = var_554)[name = string("op_556")];
            tensor<int32, [4]> var_557 = const()[name = string("op_557"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_564_pad_type_0 = const()[name = string("op_564_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_564_strides_0 = const()[name = string("op_564_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_564_pad_0 = const()[name = string("op_564_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_564_dilations_0 = const()[name = string("op_564_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_564_groups_0 = const()[name = string("op_564_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_564 = conv(dilations = var_564_dilations_0, groups = var_564_groups_0, pad = var_564_pad_0, pad_type = var_564_pad_type_0, strides = var_564_strides_0, weight = encoder_layers_2_self_attn_v_proj_weight, x = input_21)[name = string("op_564")];
            tensor<int32, [4]> var_565 = const()[name = string("op_565"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_566 = reshape(shape = var_565, x = var_564)[name = string("op_566")];
            tensor<int32, [4]> var_567 = const()[name = string("op_567"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_9_to_fp16 = const()[name = string("op_6_promoted_9_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_13 = transpose(perm = var_547, x = var_546)[name = string("transpose_232")];
            tensor<fp16, [1, 16, 512, 128]> var_573_cast_fp16 = pow(x = q_13, y = var_6_promoted_9_to_fp16)[name = string("op_573_cast_fp16")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = string("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_19_keep_dims_0 = const()[name = string("var_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_19_cast_fp16 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_573_cast_fp16)[name = string("var_19_cast_fp16")];
            fp16 var_576_to_fp16 = const()[name = string("op_576_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_577_cast_fp16 = add(x = var_19_cast_fp16, y = var_576_to_fp16)[name = string("op_577_cast_fp16")];
            fp32 var_578_epsilon_0 = const()[name = string("op_578_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_578_cast_fp16 = rsqrt(epsilon = var_578_epsilon_0, x = var_577_cast_fp16)[name = string("op_578_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_79_cast_fp16 = mul(x = q_13, y = var_578_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_15 = mul(x = x_79_cast_fp16, y = encoder_layers_2_self_attn_q_norm_weight)[name = string("q_15")];
            fp16 var_6_promoted_10_to_fp16 = const()[name = string("op_6_promoted_10_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_13 = transpose(perm = var_557, x = var_556)[name = string("transpose_231")];
            tensor<fp16, [1, 8, 512, 128]> var_586_cast_fp16 = pow(x = k_13, y = var_6_promoted_10_to_fp16)[name = string("op_586_cast_fp16")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = string("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_21_keep_dims_0 = const()[name = string("var_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_21_cast_fp16 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_586_cast_fp16)[name = string("var_21_cast_fp16")];
            fp16 var_589_to_fp16 = const()[name = string("op_589_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_590_cast_fp16 = add(x = var_21_cast_fp16, y = var_589_to_fp16)[name = string("op_590_cast_fp16")];
            fp32 var_591_epsilon_0 = const()[name = string("op_591_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_591_cast_fp16 = rsqrt(epsilon = var_591_epsilon_0, x = var_590_cast_fp16)[name = string("op_591_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_85_cast_fp16 = mul(x = k_13, y = var_591_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_15 = mul(x = x_85_cast_fp16, y = encoder_layers_2_self_attn_k_norm_weight)[name = string("k_15")];
            tensor<fp16, [1, 16, 512, 128]> var_595 = mul(x = q_15, y = cos)[name = string("op_595")];
            tensor<int32, [2]> var_596_split_sizes_0 = const()[name = string("op_596_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_596_axis_0 = const()[name = string("op_596_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_596_0, tensor<fp16, [1, 16, 512, 64]> var_596_1 = split(axis = var_596_axis_0, split_sizes = var_596_split_sizes_0, x = q_15)[name = string("op_596")];
            fp16 const_9_promoted = const()[name = string("const_9_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_598 = mul(x = var_596_1, y = const_9_promoted)[name = string("op_598")];
            bool var_600_interleave_0 = const()[name = string("op_600_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_600 = concat(axis = var_18, interleave = var_600_interleave_0, values = (var_598, var_596_0))[name = string("op_600")];
            tensor<fp16, [1, 16, 512, 128]> var_601 = mul(x = var_600, y = sin)[name = string("op_601")];
            tensor<fp16, [1, 16, 512, 128]> query_5 = add(x = var_595, y = var_601)[name = string("query_5")];
            tensor<fp16, [1, 8, 512, 128]> var_603 = mul(x = k_15, y = cos)[name = string("op_603")];
            tensor<int32, [2]> var_604_split_sizes_0 = const()[name = string("op_604_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_604_axis_0 = const()[name = string("op_604_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_604_0, tensor<fp16, [1, 8, 512, 64]> var_604_1 = split(axis = var_604_axis_0, split_sizes = var_604_split_sizes_0, x = k_15)[name = string("op_604")];
            fp16 const_10_promoted = const()[name = string("const_10_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_606 = mul(x = var_604_1, y = const_10_promoted)[name = string("op_606")];
            bool var_608_interleave_0 = const()[name = string("op_608_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_608 = concat(axis = var_18, interleave = var_608_interleave_0, values = (var_606, var_604_0))[name = string("op_608")];
            tensor<fp16, [1, 8, 512, 128]> var_609 = mul(x = var_608, y = sin)[name = string("op_609")];
            tensor<fp16, [1, 8, 512, 128]> x_87 = add(x = var_603, y = var_609)[name = string("x_87")];
            tensor<int32, [1]> var_611_axes_0 = const()[name = string("op_611_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_611 = expand_dims(axes = var_611_axes_0, x = x_87)[name = string("op_611")];
            tensor<int32, [5]> x_89_reps_0 = const()[name = string("x_89_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_89 = tile(reps = x_89_reps_0, x = var_611)[name = string("x_89")];
            tensor<int32, [4]> var_614 = const()[name = string("op_614"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_5 = reshape(shape = var_614, x = x_89)[name = string("key_5")];
            tensor<int32, [1]> var_616_axes_0 = const()[name = string("op_616_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_91 = transpose(perm = var_567, x = var_566)[name = string("transpose_230")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_616 = expand_dims(axes = var_616_axes_0, x = x_91)[name = string("op_616")];
            tensor<int32, [5]> x_93_reps_0 = const()[name = string("x_93_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_93 = tile(reps = x_93_reps_0, x = var_616)[name = string("x_93")];
            tensor<int32, [4]> var_619 = const()[name = string("op_619"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_5 = reshape(shape = var_619, x = x_93)[name = string("value_5")];
            bool var_624_transpose_x_1 = const()[name = string("op_624_transpose_x_1"), val = bool(false)];
            bool var_624_transpose_y_1 = const()[name = string("op_624_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_624_cast_fp16 = matmul(transpose_x = var_624_transpose_x_1, transpose_y = var_624_transpose_y_1, x = query_5, y = key_5)[name = string("op_624_cast_fp16")];
            fp16 var_625_to_fp16 = const()[name = string("op_625_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_13_cast_fp16 = mul(x = var_624_cast_fp16, y = var_625_to_fp16)[name = string("attn_weights_13_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_15_cast_fp16 = add(x = attn_weights_13_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_15_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_629_cast_fp16 = softmax(axis = var_18, x = attn_weights_15_cast_fp16)[name = string("op_629_cast_fp16")];
            bool var_633_transpose_x_0 = const()[name = string("op_633_transpose_x_0"), val = bool(false)];
            bool var_633_transpose_y_0 = const()[name = string("op_633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_633_cast_fp16 = matmul(transpose_x = var_633_transpose_x_0, transpose_y = var_633_transpose_y_0, x = var_629_cast_fp16, y = value_5)[name = string("op_633_cast_fp16")];
            tensor<int32, [4]> var_635 = const()[name = string("op_635"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_638 = const()[name = string("op_638"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_636 = transpose(perm = var_635, x = var_633_cast_fp16)[name = string("transpose_229")];
            tensor<fp16, [1, 512, 2048]> attn_out_15 = reshape(shape = var_638, x = var_636)[name = string("attn_out_15")];
            tensor<int32, [3]> var_640 = const()[name = string("op_640"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_2 = const()[name = string("squeeze_2"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1099733696)))];
            string var_649_pad_type_0 = const()[name = string("op_649_pad_type_0"), val = string("valid")];
            int32 var_649_groups_0 = const()[name = string("op_649_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_649_strides_0 = const()[name = string("op_649_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_649_pad_0 = const()[name = string("op_649_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_649_dilations_0 = const()[name = string("op_649_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_641 = transpose(perm = var_640, x = attn_out_15)[name = string("transpose_228")];
            tensor<fp16, [1, 1024, 512]> var_649 = conv(dilations = var_649_dilations_0, groups = var_649_groups_0, pad = var_649_pad_0, pad_type = var_649_pad_type_0, strides = var_649_strides_0, weight = squeeze_2, x = var_641)[name = string("op_649")];
            tensor<int32, [3]> var_650 = const()[name = string("op_650"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_17 = transpose(perm = var_650, x = var_649)[name = string("transpose_227")];
            tensor<fp16, [1, 512, 1024]> x_95_cast_fp16 = add(x = hidden_states_5_cast_fp16, y = attn_out_17)[name = string("x_95_cast_fp16")];
            fp16 var_6_promoted_11_to_fp16 = const()[name = string("op_6_promoted_11_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_656_cast_fp16 = pow(x = x_95_cast_fp16, y = var_6_promoted_11_to_fp16)[name = string("op_656_cast_fp16")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = string("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_23_keep_dims_0 = const()[name = string("var_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_23_cast_fp16 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_656_cast_fp16)[name = string("var_23_cast_fp16")];
            fp16 var_659_to_fp16 = const()[name = string("op_659_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_660_cast_fp16 = add(x = var_23_cast_fp16, y = var_659_to_fp16)[name = string("op_660_cast_fp16")];
            fp32 var_661_epsilon_0 = const()[name = string("op_661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_661_cast_fp16 = rsqrt(epsilon = var_661_epsilon_0, x = var_660_cast_fp16)[name = string("op_661_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_99_cast_fp16 = mul(x = x_95_cast_fp16, y = var_661_cast_fp16)[name = string("x_99_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103928064)))];
            tensor<fp16, [1, 512, 1024]> var_664_cast_fp16 = mul(x = x_99_cast_fp16, y = encoder_layers_2_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_664_cast_fp16")];
            tensor<int32, [3]> var_669 = const()[name = string("op_669"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_670 = transpose(perm = var_669, x = var_664_cast_fp16)[name = string("transpose_226")];
            tensor<fp16, [1, 1024, 1, 512]> input_25 = expand_dims(axes = input_25_axes_0, x = var_670)[name = string("input_25")];
            string input_27_pad_type_0 = const()[name = string("input_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_27_strides_0 = const()[name = string("input_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_27_pad_0 = const()[name = string("input_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_27_dilations_0 = const()[name = string("input_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_27_groups_0 = const()[name = string("input_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_27 = conv(dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = encoder_layers_2_mlp_gate_proj_weight, x = input_25)[name = string("input_27")];
            string up_5_pad_type_0 = const()[name = string("up_5_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_5_strides_0 = const()[name = string("up_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_5_pad_0 = const()[name = string("up_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_5_dilations_0 = const()[name = string("up_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_5_groups_0 = const()[name = string("up_5_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_5 = conv(dilations = up_5_dilations_0, groups = up_5_groups_0, pad = up_5_pad_0, pad_type = up_5_pad_type_0, strides = up_5_strides_0, weight = encoder_layers_2_mlp_up_proj_weight, x = input_25)[name = string("up_5")];
            tensor<fp16, [1, 3072, 1, 512]> var_684 = silu(x = input_27)[name = string("op_684")];
            tensor<fp16, [1, 3072, 1, 512]> input_29 = mul(x = var_684, y = up_5)[name = string("input_29")];
            string var_691_pad_type_0 = const()[name = string("op_691_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_691_strides_0 = const()[name = string("op_691_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_691_pad_0 = const()[name = string("op_691_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_691_dilations_0 = const()[name = string("op_691_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_691_groups_0 = const()[name = string("op_691_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_691 = conv(dilations = var_691_dilations_0, groups = var_691_groups_0, pad = var_691_pad_0, pad_type = var_691_pad_type_0, strides = var_691_strides_0, weight = encoder_layers_2_mlp_down_proj_weight, x = input_29)[name = string("op_691")];
            tensor<int32, [1]> var_692_axes_0 = const()[name = string("op_692_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_692 = squeeze(axes = var_692_axes_0, x = var_691)[name = string("op_692")];
            tensor<int32, [3]> var_693 = const()[name = string("op_693"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_5 = transpose(perm = var_693, x = var_692)[name = string("transpose_225")];
            tensor<fp16, [1, 512, 1024]> hidden_states_7_cast_fp16 = add(x = x_95_cast_fp16, y = mlp_out_5)[name = string("hidden_states_7_cast_fp16")];
            fp16 var_6_promoted_12_to_fp16 = const()[name = string("op_6_promoted_12_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_720_cast_fp16 = pow(x = hidden_states_7_cast_fp16, y = var_6_promoted_12_to_fp16)[name = string("op_720_cast_fp16")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = string("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_25_keep_dims_0 = const()[name = string("var_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_25_cast_fp16 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_720_cast_fp16)[name = string("var_25_cast_fp16")];
            fp16 var_723_to_fp16 = const()[name = string("op_723_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_724_cast_fp16 = add(x = var_25_cast_fp16, y = var_723_to_fp16)[name = string("op_724_cast_fp16")];
            fp32 var_725_epsilon_0 = const()[name = string("op_725_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_725_cast_fp16 = rsqrt(epsilon = var_725_epsilon_0, x = var_724_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_105_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = var_725_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_3_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_3_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103930176)))];
            tensor<fp16, [1, 512, 1024]> var_728_cast_fp16 = mul(x = x_105_cast_fp16, y = encoder_layers_3_input_layernorm_weight_promoted_to_fp16)[name = string("op_728_cast_fp16")];
            tensor<int32, [3]> var_733 = const()[name = string("op_733"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_734 = transpose(perm = var_733, x = var_728_cast_fp16)[name = string("transpose_224")];
            tensor<fp16, [1, 1024, 1, 512]> input_31 = expand_dims(axes = input_31_axes_0, x = var_734)[name = string("input_31")];
            string var_741_pad_type_0 = const()[name = string("op_741_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_741_strides_0 = const()[name = string("op_741_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_741_pad_0 = const()[name = string("op_741_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_741_dilations_0 = const()[name = string("op_741_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_741_groups_0 = const()[name = string("op_741_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_741 = conv(dilations = var_741_dilations_0, groups = var_741_groups_0, pad = var_741_pad_0, pad_type = var_741_pad_type_0, strides = var_741_strides_0, weight = encoder_layers_3_self_attn_q_proj_weight, x = input_31)[name = string("op_741")];
            tensor<int32, [4]> var_742 = const()[name = string("op_742"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_743 = reshape(shape = var_742, x = var_741)[name = string("op_743")];
            tensor<int32, [4]> var_744 = const()[name = string("op_744"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_751_pad_type_0 = const()[name = string("op_751_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_751_strides_0 = const()[name = string("op_751_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_751_pad_0 = const()[name = string("op_751_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_751_dilations_0 = const()[name = string("op_751_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_751_groups_0 = const()[name = string("op_751_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_751 = conv(dilations = var_751_dilations_0, groups = var_751_groups_0, pad = var_751_pad_0, pad_type = var_751_pad_type_0, strides = var_751_strides_0, weight = encoder_layers_3_self_attn_k_proj_weight, x = input_31)[name = string("op_751")];
            tensor<int32, [4]> var_752 = const()[name = string("op_752"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_753 = reshape(shape = var_752, x = var_751)[name = string("op_753")];
            tensor<int32, [4]> var_754 = const()[name = string("op_754"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_761_pad_type_0 = const()[name = string("op_761_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_761_strides_0 = const()[name = string("op_761_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_761_pad_0 = const()[name = string("op_761_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_761_dilations_0 = const()[name = string("op_761_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_761_groups_0 = const()[name = string("op_761_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_761 = conv(dilations = var_761_dilations_0, groups = var_761_groups_0, pad = var_761_pad_0, pad_type = var_761_pad_type_0, strides = var_761_strides_0, weight = encoder_layers_3_self_attn_v_proj_weight, x = input_31)[name = string("op_761")];
            tensor<int32, [4]> var_762 = const()[name = string("op_762"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_763 = reshape(shape = var_762, x = var_761)[name = string("op_763")];
            tensor<int32, [4]> var_764 = const()[name = string("op_764"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_13_to_fp16 = const()[name = string("op_6_promoted_13_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_19 = transpose(perm = var_744, x = var_743)[name = string("transpose_223")];
            tensor<fp16, [1, 16, 512, 128]> var_770_cast_fp16 = pow(x = q_19, y = var_6_promoted_13_to_fp16)[name = string("op_770_cast_fp16")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = string("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_27_keep_dims_0 = const()[name = string("var_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_27_cast_fp16 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_770_cast_fp16)[name = string("var_27_cast_fp16")];
            fp16 var_773_to_fp16 = const()[name = string("op_773_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_774_cast_fp16 = add(x = var_27_cast_fp16, y = var_773_to_fp16)[name = string("op_774_cast_fp16")];
            fp32 var_775_epsilon_0 = const()[name = string("op_775_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_775_cast_fp16 = rsqrt(epsilon = var_775_epsilon_0, x = var_774_cast_fp16)[name = string("op_775_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_113_cast_fp16 = mul(x = q_19, y = var_775_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_21 = mul(x = x_113_cast_fp16, y = encoder_layers_3_self_attn_q_norm_weight)[name = string("q_21")];
            fp16 var_6_promoted_14_to_fp16 = const()[name = string("op_6_promoted_14_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_19 = transpose(perm = var_754, x = var_753)[name = string("transpose_222")];
            tensor<fp16, [1, 8, 512, 128]> var_783_cast_fp16 = pow(x = k_19, y = var_6_promoted_14_to_fp16)[name = string("op_783_cast_fp16")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = string("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_29_keep_dims_0 = const()[name = string("var_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_29_cast_fp16 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_783_cast_fp16)[name = string("var_29_cast_fp16")];
            fp16 var_786_to_fp16 = const()[name = string("op_786_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_787_cast_fp16 = add(x = var_29_cast_fp16, y = var_786_to_fp16)[name = string("op_787_cast_fp16")];
            fp32 var_788_epsilon_0 = const()[name = string("op_788_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_788_cast_fp16 = rsqrt(epsilon = var_788_epsilon_0, x = var_787_cast_fp16)[name = string("op_788_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_119_cast_fp16 = mul(x = k_19, y = var_788_cast_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_21 = mul(x = x_119_cast_fp16, y = encoder_layers_3_self_attn_k_norm_weight)[name = string("k_21")];
            tensor<fp16, [1, 16, 512, 128]> var_792 = mul(x = q_21, y = cos)[name = string("op_792")];
            tensor<int32, [2]> var_793_split_sizes_0 = const()[name = string("op_793_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_793_axis_0 = const()[name = string("op_793_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_793_0, tensor<fp16, [1, 16, 512, 64]> var_793_1 = split(axis = var_793_axis_0, split_sizes = var_793_split_sizes_0, x = q_21)[name = string("op_793")];
            fp16 const_12_promoted = const()[name = string("const_12_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_795 = mul(x = var_793_1, y = const_12_promoted)[name = string("op_795")];
            bool var_797_interleave_0 = const()[name = string("op_797_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_797 = concat(axis = var_18, interleave = var_797_interleave_0, values = (var_795, var_793_0))[name = string("op_797")];
            tensor<fp16, [1, 16, 512, 128]> var_798 = mul(x = var_797, y = sin)[name = string("op_798")];
            tensor<fp16, [1, 16, 512, 128]> query_7 = add(x = var_792, y = var_798)[name = string("query_7")];
            tensor<fp16, [1, 8, 512, 128]> var_800 = mul(x = k_21, y = cos)[name = string("op_800")];
            tensor<int32, [2]> var_801_split_sizes_0 = const()[name = string("op_801_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_801_axis_0 = const()[name = string("op_801_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_801_0, tensor<fp16, [1, 8, 512, 64]> var_801_1 = split(axis = var_801_axis_0, split_sizes = var_801_split_sizes_0, x = k_21)[name = string("op_801")];
            fp16 const_13_promoted = const()[name = string("const_13_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_803 = mul(x = var_801_1, y = const_13_promoted)[name = string("op_803")];
            bool var_805_interleave_0 = const()[name = string("op_805_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_805 = concat(axis = var_18, interleave = var_805_interleave_0, values = (var_803, var_801_0))[name = string("op_805")];
            tensor<fp16, [1, 8, 512, 128]> var_806 = mul(x = var_805, y = sin)[name = string("op_806")];
            tensor<fp16, [1, 8, 512, 128]> x_121 = add(x = var_800, y = var_806)[name = string("x_121")];
            tensor<int32, [1]> var_808_axes_0 = const()[name = string("op_808_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_808 = expand_dims(axes = var_808_axes_0, x = x_121)[name = string("op_808")];
            tensor<int32, [5]> x_123_reps_0 = const()[name = string("x_123_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_123 = tile(reps = x_123_reps_0, x = var_808)[name = string("x_123")];
            tensor<int32, [4]> var_811 = const()[name = string("op_811"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_7 = reshape(shape = var_811, x = x_123)[name = string("key_7")];
            tensor<int32, [1]> var_813_axes_0 = const()[name = string("op_813_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_125 = transpose(perm = var_764, x = var_763)[name = string("transpose_221")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_813 = expand_dims(axes = var_813_axes_0, x = x_125)[name = string("op_813")];
            tensor<int32, [5]> x_127_reps_0 = const()[name = string("x_127_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_127 = tile(reps = x_127_reps_0, x = var_813)[name = string("x_127")];
            tensor<int32, [4]> var_816 = const()[name = string("op_816"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_7 = reshape(shape = var_816, x = x_127)[name = string("value_7")];
            bool var_821_transpose_x_1 = const()[name = string("op_821_transpose_x_1"), val = bool(false)];
            bool var_821_transpose_y_1 = const()[name = string("op_821_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_821_cast_fp16 = matmul(transpose_x = var_821_transpose_x_1, transpose_y = var_821_transpose_y_1, x = query_7, y = key_7)[name = string("op_821_cast_fp16")];
            fp16 var_822_to_fp16 = const()[name = string("op_822_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_19_cast_fp16 = mul(x = var_821_cast_fp16, y = var_822_to_fp16)[name = string("attn_weights_19_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_21_cast_fp16 = add(x = attn_weights_19_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_21_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_826_cast_fp16 = softmax(axis = var_18, x = attn_weights_21_cast_fp16)[name = string("op_826_cast_fp16")];
            bool var_830_transpose_x_0 = const()[name = string("op_830_transpose_x_0"), val = bool(false)];
            bool var_830_transpose_y_0 = const()[name = string("op_830_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_830_cast_fp16 = matmul(transpose_x = var_830_transpose_x_0, transpose_y = var_830_transpose_y_0, x = var_826_cast_fp16, y = value_7)[name = string("op_830_cast_fp16")];
            tensor<int32, [4]> var_832 = const()[name = string("op_832"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_835 = const()[name = string("op_835"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_833 = transpose(perm = var_832, x = var_830_cast_fp16)[name = string("transpose_220")];
            tensor<fp16, [1, 512, 2048]> attn_out_21 = reshape(shape = var_835, x = var_833)[name = string("attn_out_21")];
            tensor<int32, [3]> var_837 = const()[name = string("op_837"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_3 = const()[name = string("squeeze_3"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1103932288)))];
            string var_846_pad_type_0 = const()[name = string("op_846_pad_type_0"), val = string("valid")];
            int32 var_846_groups_0 = const()[name = string("op_846_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_846_strides_0 = const()[name = string("op_846_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_846_pad_0 = const()[name = string("op_846_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_846_dilations_0 = const()[name = string("op_846_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_838 = transpose(perm = var_837, x = attn_out_21)[name = string("transpose_219")];
            tensor<fp16, [1, 1024, 512]> var_846 = conv(dilations = var_846_dilations_0, groups = var_846_groups_0, pad = var_846_pad_0, pad_type = var_846_pad_type_0, strides = var_846_strides_0, weight = squeeze_3, x = var_838)[name = string("op_846")];
            tensor<int32, [3]> var_847 = const()[name = string("op_847"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_23 = transpose(perm = var_847, x = var_846)[name = string("transpose_218")];
            tensor<fp16, [1, 512, 1024]> x_129_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = attn_out_23)[name = string("x_129_cast_fp16")];
            fp16 var_6_promoted_15_to_fp16 = const()[name = string("op_6_promoted_15_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_853_cast_fp16 = pow(x = x_129_cast_fp16, y = var_6_promoted_15_to_fp16)[name = string("op_853_cast_fp16")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = string("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_31_keep_dims_0 = const()[name = string("var_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_31_cast_fp16 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_853_cast_fp16)[name = string("var_31_cast_fp16")];
            fp16 var_856_to_fp16 = const()[name = string("op_856_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_857_cast_fp16 = add(x = var_31_cast_fp16, y = var_856_to_fp16)[name = string("op_857_cast_fp16")];
            fp32 var_858_epsilon_0 = const()[name = string("op_858_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_858_cast_fp16 = rsqrt(epsilon = var_858_epsilon_0, x = var_857_cast_fp16)[name = string("op_858_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_133_cast_fp16 = mul(x = x_129_cast_fp16, y = var_858_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108126656)))];
            tensor<fp16, [1, 512, 1024]> var_861_cast_fp16 = mul(x = x_133_cast_fp16, y = encoder_layers_3_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_861_cast_fp16")];
            tensor<int32, [3]> var_866 = const()[name = string("op_866"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_867 = transpose(perm = var_866, x = var_861_cast_fp16)[name = string("transpose_217")];
            tensor<fp16, [1, 1024, 1, 512]> input_35 = expand_dims(axes = input_35_axes_0, x = var_867)[name = string("input_35")];
            string input_37_pad_type_0 = const()[name = string("input_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_37_strides_0 = const()[name = string("input_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_37_pad_0 = const()[name = string("input_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_37_dilations_0 = const()[name = string("input_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_37_groups_0 = const()[name = string("input_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_37 = conv(dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = encoder_layers_3_mlp_gate_proj_weight, x = input_35)[name = string("input_37")];
            string up_7_pad_type_0 = const()[name = string("up_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_7_strides_0 = const()[name = string("up_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_7_pad_0 = const()[name = string("up_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_7_dilations_0 = const()[name = string("up_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_7_groups_0 = const()[name = string("up_7_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_7 = conv(dilations = up_7_dilations_0, groups = up_7_groups_0, pad = up_7_pad_0, pad_type = up_7_pad_type_0, strides = up_7_strides_0, weight = encoder_layers_3_mlp_up_proj_weight, x = input_35)[name = string("up_7")];
            tensor<fp16, [1, 3072, 1, 512]> var_881 = silu(x = input_37)[name = string("op_881")];
            tensor<fp16, [1, 3072, 1, 512]> input_39 = mul(x = var_881, y = up_7)[name = string("input_39")];
            string var_888_pad_type_0 = const()[name = string("op_888_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_888_strides_0 = const()[name = string("op_888_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_888_pad_0 = const()[name = string("op_888_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_888_dilations_0 = const()[name = string("op_888_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_888_groups_0 = const()[name = string("op_888_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_888 = conv(dilations = var_888_dilations_0, groups = var_888_groups_0, pad = var_888_pad_0, pad_type = var_888_pad_type_0, strides = var_888_strides_0, weight = encoder_layers_3_mlp_down_proj_weight, x = input_39)[name = string("op_888")];
            tensor<int32, [1]> var_889_axes_0 = const()[name = string("op_889_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_889 = squeeze(axes = var_889_axes_0, x = var_888)[name = string("op_889")];
            tensor<int32, [3]> var_890 = const()[name = string("op_890"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_7 = transpose(perm = var_890, x = var_889)[name = string("transpose_216")];
            tensor<fp16, [1, 512, 1024]> hidden_states_9_cast_fp16 = add(x = x_129_cast_fp16, y = mlp_out_7)[name = string("hidden_states_9_cast_fp16")];
            fp16 var_6_promoted_16_to_fp16 = const()[name = string("op_6_promoted_16_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_917_cast_fp16 = pow(x = hidden_states_9_cast_fp16, y = var_6_promoted_16_to_fp16)[name = string("op_917_cast_fp16")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = string("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_33_keep_dims_0 = const()[name = string("var_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_33_cast_fp16 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_917_cast_fp16)[name = string("var_33_cast_fp16")];
            fp16 var_920_to_fp16 = const()[name = string("op_920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_921_cast_fp16 = add(x = var_33_cast_fp16, y = var_920_to_fp16)[name = string("op_921_cast_fp16")];
            fp32 var_922_epsilon_0 = const()[name = string("op_922_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_922_cast_fp16 = rsqrt(epsilon = var_922_epsilon_0, x = var_921_cast_fp16)[name = string("op_922_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_139_cast_fp16 = mul(x = hidden_states_9_cast_fp16, y = var_922_cast_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_4_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_4_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108128768)))];
            tensor<fp16, [1, 512, 1024]> var_925_cast_fp16 = mul(x = x_139_cast_fp16, y = encoder_layers_4_input_layernorm_weight_promoted_to_fp16)[name = string("op_925_cast_fp16")];
            tensor<int32, [3]> var_930 = const()[name = string("op_930"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_931 = transpose(perm = var_930, x = var_925_cast_fp16)[name = string("transpose_215")];
            tensor<fp16, [1, 1024, 1, 512]> input_41 = expand_dims(axes = input_41_axes_0, x = var_931)[name = string("input_41")];
            string var_938_pad_type_0 = const()[name = string("op_938_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_938_strides_0 = const()[name = string("op_938_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_938_pad_0 = const()[name = string("op_938_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_938_dilations_0 = const()[name = string("op_938_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_938_groups_0 = const()[name = string("op_938_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_938 = conv(dilations = var_938_dilations_0, groups = var_938_groups_0, pad = var_938_pad_0, pad_type = var_938_pad_type_0, strides = var_938_strides_0, weight = encoder_layers_4_self_attn_q_proj_weight, x = input_41)[name = string("op_938")];
            tensor<int32, [4]> var_939 = const()[name = string("op_939"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_940 = reshape(shape = var_939, x = var_938)[name = string("op_940")];
            tensor<int32, [4]> var_941 = const()[name = string("op_941"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_948_pad_type_0 = const()[name = string("op_948_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_948_strides_0 = const()[name = string("op_948_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_948_pad_0 = const()[name = string("op_948_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_948_dilations_0 = const()[name = string("op_948_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_948_groups_0 = const()[name = string("op_948_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_948 = conv(dilations = var_948_dilations_0, groups = var_948_groups_0, pad = var_948_pad_0, pad_type = var_948_pad_type_0, strides = var_948_strides_0, weight = encoder_layers_4_self_attn_k_proj_weight, x = input_41)[name = string("op_948")];
            tensor<int32, [4]> var_949 = const()[name = string("op_949"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_950 = reshape(shape = var_949, x = var_948)[name = string("op_950")];
            tensor<int32, [4]> var_951 = const()[name = string("op_951"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_958_pad_type_0 = const()[name = string("op_958_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_958_strides_0 = const()[name = string("op_958_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_958_pad_0 = const()[name = string("op_958_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_958_dilations_0 = const()[name = string("op_958_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_958_groups_0 = const()[name = string("op_958_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_958 = conv(dilations = var_958_dilations_0, groups = var_958_groups_0, pad = var_958_pad_0, pad_type = var_958_pad_type_0, strides = var_958_strides_0, weight = encoder_layers_4_self_attn_v_proj_weight, x = input_41)[name = string("op_958")];
            tensor<int32, [4]> var_959 = const()[name = string("op_959"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_960 = reshape(shape = var_959, x = var_958)[name = string("op_960")];
            tensor<int32, [4]> var_961 = const()[name = string("op_961"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_17_to_fp16 = const()[name = string("op_6_promoted_17_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_25 = transpose(perm = var_941, x = var_940)[name = string("transpose_214")];
            tensor<fp16, [1, 16, 512, 128]> var_967_cast_fp16 = pow(x = q_25, y = var_6_promoted_17_to_fp16)[name = string("op_967_cast_fp16")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = string("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_35_keep_dims_0 = const()[name = string("var_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_35_cast_fp16 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_967_cast_fp16)[name = string("var_35_cast_fp16")];
            fp16 var_970_to_fp16 = const()[name = string("op_970_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_971_cast_fp16 = add(x = var_35_cast_fp16, y = var_970_to_fp16)[name = string("op_971_cast_fp16")];
            fp32 var_972_epsilon_0 = const()[name = string("op_972_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_972_cast_fp16 = rsqrt(epsilon = var_972_epsilon_0, x = var_971_cast_fp16)[name = string("op_972_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_147_cast_fp16 = mul(x = q_25, y = var_972_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_27 = mul(x = x_147_cast_fp16, y = encoder_layers_4_self_attn_q_norm_weight)[name = string("q_27")];
            fp16 var_6_promoted_18_to_fp16 = const()[name = string("op_6_promoted_18_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_25 = transpose(perm = var_951, x = var_950)[name = string("transpose_213")];
            tensor<fp16, [1, 8, 512, 128]> var_980_cast_fp16 = pow(x = k_25, y = var_6_promoted_18_to_fp16)[name = string("op_980_cast_fp16")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = string("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_37_keep_dims_0 = const()[name = string("var_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_37_cast_fp16 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_980_cast_fp16)[name = string("var_37_cast_fp16")];
            fp16 var_983_to_fp16 = const()[name = string("op_983_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_984_cast_fp16 = add(x = var_37_cast_fp16, y = var_983_to_fp16)[name = string("op_984_cast_fp16")];
            fp32 var_985_epsilon_0 = const()[name = string("op_985_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_985_cast_fp16 = rsqrt(epsilon = var_985_epsilon_0, x = var_984_cast_fp16)[name = string("op_985_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_153_cast_fp16 = mul(x = k_25, y = var_985_cast_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_27 = mul(x = x_153_cast_fp16, y = encoder_layers_4_self_attn_k_norm_weight)[name = string("k_27")];
            tensor<fp16, [1, 16, 512, 128]> var_989 = mul(x = q_27, y = cos)[name = string("op_989")];
            tensor<int32, [2]> var_990_split_sizes_0 = const()[name = string("op_990_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_990_axis_0 = const()[name = string("op_990_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_990_0, tensor<fp16, [1, 16, 512, 64]> var_990_1 = split(axis = var_990_axis_0, split_sizes = var_990_split_sizes_0, x = q_27)[name = string("op_990")];
            fp16 const_15_promoted = const()[name = string("const_15_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_992 = mul(x = var_990_1, y = const_15_promoted)[name = string("op_992")];
            bool var_994_interleave_0 = const()[name = string("op_994_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_994 = concat(axis = var_18, interleave = var_994_interleave_0, values = (var_992, var_990_0))[name = string("op_994")];
            tensor<fp16, [1, 16, 512, 128]> var_995 = mul(x = var_994, y = sin)[name = string("op_995")];
            tensor<fp16, [1, 16, 512, 128]> query_9 = add(x = var_989, y = var_995)[name = string("query_9")];
            tensor<fp16, [1, 8, 512, 128]> var_997 = mul(x = k_27, y = cos)[name = string("op_997")];
            tensor<int32, [2]> var_998_split_sizes_0 = const()[name = string("op_998_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_998_axis_0 = const()[name = string("op_998_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_998_0, tensor<fp16, [1, 8, 512, 64]> var_998_1 = split(axis = var_998_axis_0, split_sizes = var_998_split_sizes_0, x = k_27)[name = string("op_998")];
            fp16 const_16_promoted = const()[name = string("const_16_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1000 = mul(x = var_998_1, y = const_16_promoted)[name = string("op_1000")];
            bool var_1002_interleave_0 = const()[name = string("op_1002_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1002 = concat(axis = var_18, interleave = var_1002_interleave_0, values = (var_1000, var_998_0))[name = string("op_1002")];
            tensor<fp16, [1, 8, 512, 128]> var_1003 = mul(x = var_1002, y = sin)[name = string("op_1003")];
            tensor<fp16, [1, 8, 512, 128]> x_155 = add(x = var_997, y = var_1003)[name = string("x_155")];
            tensor<int32, [1]> var_1005_axes_0 = const()[name = string("op_1005_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1005 = expand_dims(axes = var_1005_axes_0, x = x_155)[name = string("op_1005")];
            tensor<int32, [5]> x_157_reps_0 = const()[name = string("x_157_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_157 = tile(reps = x_157_reps_0, x = var_1005)[name = string("x_157")];
            tensor<int32, [4]> var_1008 = const()[name = string("op_1008"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_9 = reshape(shape = var_1008, x = x_157)[name = string("key_9")];
            tensor<int32, [1]> var_1010_axes_0 = const()[name = string("op_1010_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_159 = transpose(perm = var_961, x = var_960)[name = string("transpose_212")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1010 = expand_dims(axes = var_1010_axes_0, x = x_159)[name = string("op_1010")];
            tensor<int32, [5]> x_161_reps_0 = const()[name = string("x_161_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_161 = tile(reps = x_161_reps_0, x = var_1010)[name = string("x_161")];
            tensor<int32, [4]> var_1013 = const()[name = string("op_1013"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_9 = reshape(shape = var_1013, x = x_161)[name = string("value_9")];
            bool var_1018_transpose_x_1 = const()[name = string("op_1018_transpose_x_1"), val = bool(false)];
            bool var_1018_transpose_y_1 = const()[name = string("op_1018_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_1018_cast_fp16 = matmul(transpose_x = var_1018_transpose_x_1, transpose_y = var_1018_transpose_y_1, x = query_9, y = key_9)[name = string("op_1018_cast_fp16")];
            fp16 var_1019_to_fp16 = const()[name = string("op_1019_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_25_cast_fp16 = mul(x = var_1018_cast_fp16, y = var_1019_to_fp16)[name = string("attn_weights_25_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_27_cast_fp16 = add(x = attn_weights_25_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_27_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_1023_cast_fp16 = softmax(axis = var_18, x = attn_weights_27_cast_fp16)[name = string("op_1023_cast_fp16")];
            bool var_1027_transpose_x_0 = const()[name = string("op_1027_transpose_x_0"), val = bool(false)];
            bool var_1027_transpose_y_0 = const()[name = string("op_1027_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1027_cast_fp16 = matmul(transpose_x = var_1027_transpose_x_0, transpose_y = var_1027_transpose_y_0, x = var_1023_cast_fp16, y = value_9)[name = string("op_1027_cast_fp16")];
            tensor<int32, [4]> var_1029 = const()[name = string("op_1029"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1032 = const()[name = string("op_1032"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_1030 = transpose(perm = var_1029, x = var_1027_cast_fp16)[name = string("transpose_211")];
            tensor<fp16, [1, 512, 2048]> attn_out_27 = reshape(shape = var_1032, x = var_1030)[name = string("attn_out_27")];
            tensor<int32, [3]> var_1034 = const()[name = string("op_1034"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_4 = const()[name = string("squeeze_4"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1108130880)))];
            string var_1043_pad_type_0 = const()[name = string("op_1043_pad_type_0"), val = string("valid")];
            int32 var_1043_groups_0 = const()[name = string("op_1043_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1043_strides_0 = const()[name = string("op_1043_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1043_pad_0 = const()[name = string("op_1043_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1043_dilations_0 = const()[name = string("op_1043_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_1035 = transpose(perm = var_1034, x = attn_out_27)[name = string("transpose_210")];
            tensor<fp16, [1, 1024, 512]> var_1043 = conv(dilations = var_1043_dilations_0, groups = var_1043_groups_0, pad = var_1043_pad_0, pad_type = var_1043_pad_type_0, strides = var_1043_strides_0, weight = squeeze_4, x = var_1035)[name = string("op_1043")];
            tensor<int32, [3]> var_1044 = const()[name = string("op_1044"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_29 = transpose(perm = var_1044, x = var_1043)[name = string("transpose_209")];
            tensor<fp16, [1, 512, 1024]> x_163_cast_fp16 = add(x = hidden_states_9_cast_fp16, y = attn_out_29)[name = string("x_163_cast_fp16")];
            fp16 var_6_promoted_19_to_fp16 = const()[name = string("op_6_promoted_19_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1050_cast_fp16 = pow(x = x_163_cast_fp16, y = var_6_promoted_19_to_fp16)[name = string("op_1050_cast_fp16")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = string("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_39_keep_dims_0 = const()[name = string("var_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_39_cast_fp16 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_1050_cast_fp16)[name = string("var_39_cast_fp16")];
            fp16 var_1053_to_fp16 = const()[name = string("op_1053_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1054_cast_fp16 = add(x = var_39_cast_fp16, y = var_1053_to_fp16)[name = string("op_1054_cast_fp16")];
            fp32 var_1055_epsilon_0 = const()[name = string("op_1055_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1055_cast_fp16 = rsqrt(epsilon = var_1055_epsilon_0, x = var_1054_cast_fp16)[name = string("op_1055_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_167_cast_fp16 = mul(x = x_163_cast_fp16, y = var_1055_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112325248)))];
            tensor<fp16, [1, 512, 1024]> var_1058_cast_fp16 = mul(x = x_167_cast_fp16, y = encoder_layers_4_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1058_cast_fp16")];
            tensor<int32, [3]> var_1063 = const()[name = string("op_1063"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1064 = transpose(perm = var_1063, x = var_1058_cast_fp16)[name = string("transpose_208")];
            tensor<fp16, [1, 1024, 1, 512]> input_45 = expand_dims(axes = input_45_axes_0, x = var_1064)[name = string("input_45")];
            string input_47_pad_type_0 = const()[name = string("input_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_47_strides_0 = const()[name = string("input_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_47_pad_0 = const()[name = string("input_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_47_dilations_0 = const()[name = string("input_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_47_groups_0 = const()[name = string("input_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_47 = conv(dilations = input_47_dilations_0, groups = input_47_groups_0, pad = input_47_pad_0, pad_type = input_47_pad_type_0, strides = input_47_strides_0, weight = encoder_layers_4_mlp_gate_proj_weight, x = input_45)[name = string("input_47")];
            string up_9_pad_type_0 = const()[name = string("up_9_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_9_strides_0 = const()[name = string("up_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_9_pad_0 = const()[name = string("up_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_9_dilations_0 = const()[name = string("up_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_9_groups_0 = const()[name = string("up_9_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_9 = conv(dilations = up_9_dilations_0, groups = up_9_groups_0, pad = up_9_pad_0, pad_type = up_9_pad_type_0, strides = up_9_strides_0, weight = encoder_layers_4_mlp_up_proj_weight, x = input_45)[name = string("up_9")];
            tensor<fp16, [1, 3072, 1, 512]> var_1078 = silu(x = input_47)[name = string("op_1078")];
            tensor<fp16, [1, 3072, 1, 512]> input_49 = mul(x = var_1078, y = up_9)[name = string("input_49")];
            string var_1085_pad_type_0 = const()[name = string("op_1085_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1085_strides_0 = const()[name = string("op_1085_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1085_pad_0 = const()[name = string("op_1085_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1085_dilations_0 = const()[name = string("op_1085_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1085_groups_0 = const()[name = string("op_1085_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1085 = conv(dilations = var_1085_dilations_0, groups = var_1085_groups_0, pad = var_1085_pad_0, pad_type = var_1085_pad_type_0, strides = var_1085_strides_0, weight = encoder_layers_4_mlp_down_proj_weight, x = input_49)[name = string("op_1085")];
            tensor<int32, [1]> var_1086_axes_0 = const()[name = string("op_1086_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1086 = squeeze(axes = var_1086_axes_0, x = var_1085)[name = string("op_1086")];
            tensor<int32, [3]> var_1087 = const()[name = string("op_1087"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_9 = transpose(perm = var_1087, x = var_1086)[name = string("transpose_207")];
            tensor<fp16, [1, 512, 1024]> hidden_states_11_cast_fp16 = add(x = x_163_cast_fp16, y = mlp_out_9)[name = string("hidden_states_11_cast_fp16")];
            fp16 var_6_promoted_20_to_fp16 = const()[name = string("op_6_promoted_20_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1114_cast_fp16 = pow(x = hidden_states_11_cast_fp16, y = var_6_promoted_20_to_fp16)[name = string("op_1114_cast_fp16")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = string("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_41_keep_dims_0 = const()[name = string("var_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_41_cast_fp16 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1114_cast_fp16)[name = string("var_41_cast_fp16")];
            fp16 var_1117_to_fp16 = const()[name = string("op_1117_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1118_cast_fp16 = add(x = var_41_cast_fp16, y = var_1117_to_fp16)[name = string("op_1118_cast_fp16")];
            fp32 var_1119_epsilon_0 = const()[name = string("op_1119_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1119_cast_fp16 = rsqrt(epsilon = var_1119_epsilon_0, x = var_1118_cast_fp16)[name = string("op_1119_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_173_cast_fp16 = mul(x = hidden_states_11_cast_fp16, y = var_1119_cast_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_5_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_5_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112327360)))];
            tensor<fp16, [1, 512, 1024]> var_1122_cast_fp16 = mul(x = x_173_cast_fp16, y = encoder_layers_5_input_layernorm_weight_promoted_to_fp16)[name = string("op_1122_cast_fp16")];
            tensor<int32, [3]> var_1127 = const()[name = string("op_1127"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1128 = transpose(perm = var_1127, x = var_1122_cast_fp16)[name = string("transpose_206")];
            tensor<fp16, [1, 1024, 1, 512]> input_51 = expand_dims(axes = input_51_axes_0, x = var_1128)[name = string("input_51")];
            string var_1135_pad_type_0 = const()[name = string("op_1135_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1135_strides_0 = const()[name = string("op_1135_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1135_pad_0 = const()[name = string("op_1135_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1135_dilations_0 = const()[name = string("op_1135_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1135_groups_0 = const()[name = string("op_1135_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_1135 = conv(dilations = var_1135_dilations_0, groups = var_1135_groups_0, pad = var_1135_pad_0, pad_type = var_1135_pad_type_0, strides = var_1135_strides_0, weight = encoder_layers_5_self_attn_q_proj_weight, x = input_51)[name = string("op_1135")];
            tensor<int32, [4]> var_1136 = const()[name = string("op_1136"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_1137 = reshape(shape = var_1136, x = var_1135)[name = string("op_1137")];
            tensor<int32, [4]> var_1138 = const()[name = string("op_1138"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1145_pad_type_0 = const()[name = string("op_1145_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1145_strides_0 = const()[name = string("op_1145_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1145_pad_0 = const()[name = string("op_1145_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1145_dilations_0 = const()[name = string("op_1145_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1145_groups_0 = const()[name = string("op_1145_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1145 = conv(dilations = var_1145_dilations_0, groups = var_1145_groups_0, pad = var_1145_pad_0, pad_type = var_1145_pad_type_0, strides = var_1145_strides_0, weight = encoder_layers_5_self_attn_k_proj_weight, x = input_51)[name = string("op_1145")];
            tensor<int32, [4]> var_1146 = const()[name = string("op_1146"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1147 = reshape(shape = var_1146, x = var_1145)[name = string("op_1147")];
            tensor<int32, [4]> var_1148 = const()[name = string("op_1148"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1155_pad_type_0 = const()[name = string("op_1155_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1155_strides_0 = const()[name = string("op_1155_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1155_pad_0 = const()[name = string("op_1155_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1155_dilations_0 = const()[name = string("op_1155_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1155_groups_0 = const()[name = string("op_1155_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1155 = conv(dilations = var_1155_dilations_0, groups = var_1155_groups_0, pad = var_1155_pad_0, pad_type = var_1155_pad_type_0, strides = var_1155_strides_0, weight = encoder_layers_5_self_attn_v_proj_weight, x = input_51)[name = string("op_1155")];
            tensor<int32, [4]> var_1156 = const()[name = string("op_1156"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1157 = reshape(shape = var_1156, x = var_1155)[name = string("op_1157")];
            tensor<int32, [4]> var_1158 = const()[name = string("op_1158"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_21_to_fp16 = const()[name = string("op_6_promoted_21_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_31 = transpose(perm = var_1138, x = var_1137)[name = string("transpose_205")];
            tensor<fp16, [1, 16, 512, 128]> var_1164_cast_fp16 = pow(x = q_31, y = var_6_promoted_21_to_fp16)[name = string("op_1164_cast_fp16")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = string("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_43_keep_dims_0 = const()[name = string("var_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_43_cast_fp16 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1164_cast_fp16)[name = string("var_43_cast_fp16")];
            fp16 var_1167_to_fp16 = const()[name = string("op_1167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_1168_cast_fp16 = add(x = var_43_cast_fp16, y = var_1167_to_fp16)[name = string("op_1168_cast_fp16")];
            fp32 var_1169_epsilon_0 = const()[name = string("op_1169_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_1169_cast_fp16 = rsqrt(epsilon = var_1169_epsilon_0, x = var_1168_cast_fp16)[name = string("op_1169_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_181_cast_fp16 = mul(x = q_31, y = var_1169_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_33 = mul(x = x_181_cast_fp16, y = encoder_layers_5_self_attn_q_norm_weight)[name = string("q_33")];
            fp16 var_6_promoted_22_to_fp16 = const()[name = string("op_6_promoted_22_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_31 = transpose(perm = var_1148, x = var_1147)[name = string("transpose_204")];
            tensor<fp16, [1, 8, 512, 128]> var_1177_cast_fp16 = pow(x = k_31, y = var_6_promoted_22_to_fp16)[name = string("op_1177_cast_fp16")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = string("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_45_keep_dims_0 = const()[name = string("var_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_45_cast_fp16 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1177_cast_fp16)[name = string("var_45_cast_fp16")];
            fp16 var_1180_to_fp16 = const()[name = string("op_1180_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_1181_cast_fp16 = add(x = var_45_cast_fp16, y = var_1180_to_fp16)[name = string("op_1181_cast_fp16")];
            fp32 var_1182_epsilon_0 = const()[name = string("op_1182_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_1182_cast_fp16 = rsqrt(epsilon = var_1182_epsilon_0, x = var_1181_cast_fp16)[name = string("op_1182_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_187_cast_fp16 = mul(x = k_31, y = var_1182_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_33 = mul(x = x_187_cast_fp16, y = encoder_layers_5_self_attn_k_norm_weight)[name = string("k_33")];
            tensor<fp16, [1, 16, 512, 128]> var_1186 = mul(x = q_33, y = cos)[name = string("op_1186")];
            tensor<int32, [2]> var_1187_split_sizes_0 = const()[name = string("op_1187_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1187_axis_0 = const()[name = string("op_1187_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_1187_0, tensor<fp16, [1, 16, 512, 64]> var_1187_1 = split(axis = var_1187_axis_0, split_sizes = var_1187_split_sizes_0, x = q_33)[name = string("op_1187")];
            fp16 const_18_promoted = const()[name = string("const_18_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_1189 = mul(x = var_1187_1, y = const_18_promoted)[name = string("op_1189")];
            bool var_1191_interleave_0 = const()[name = string("op_1191_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1191 = concat(axis = var_18, interleave = var_1191_interleave_0, values = (var_1189, var_1187_0))[name = string("op_1191")];
            tensor<fp16, [1, 16, 512, 128]> var_1192 = mul(x = var_1191, y = sin)[name = string("op_1192")];
            tensor<fp16, [1, 16, 512, 128]> query_11 = add(x = var_1186, y = var_1192)[name = string("query_11")];
            tensor<fp16, [1, 8, 512, 128]> var_1194 = mul(x = k_33, y = cos)[name = string("op_1194")];
            tensor<int32, [2]> var_1195_split_sizes_0 = const()[name = string("op_1195_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1195_axis_0 = const()[name = string("op_1195_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_1195_0, tensor<fp16, [1, 8, 512, 64]> var_1195_1 = split(axis = var_1195_axis_0, split_sizes = var_1195_split_sizes_0, x = k_33)[name = string("op_1195")];
            fp16 const_19_promoted = const()[name = string("const_19_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1197 = mul(x = var_1195_1, y = const_19_promoted)[name = string("op_1197")];
            bool var_1199_interleave_0 = const()[name = string("op_1199_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1199 = concat(axis = var_18, interleave = var_1199_interleave_0, values = (var_1197, var_1195_0))[name = string("op_1199")];
            tensor<fp16, [1, 8, 512, 128]> var_1200 = mul(x = var_1199, y = sin)[name = string("op_1200")];
            tensor<fp16, [1, 8, 512, 128]> x_189 = add(x = var_1194, y = var_1200)[name = string("x_189")];
            tensor<int32, [1]> var_1202_axes_0 = const()[name = string("op_1202_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1202 = expand_dims(axes = var_1202_axes_0, x = x_189)[name = string("op_1202")];
            tensor<int32, [5]> x_191_reps_0 = const()[name = string("x_191_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_191 = tile(reps = x_191_reps_0, x = var_1202)[name = string("x_191")];
            tensor<int32, [4]> var_1205 = const()[name = string("op_1205"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_11 = reshape(shape = var_1205, x = x_191)[name = string("key_11")];
            tensor<int32, [1]> var_1207_axes_0 = const()[name = string("op_1207_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_193 = transpose(perm = var_1158, x = var_1157)[name = string("transpose_203")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1207 = expand_dims(axes = var_1207_axes_0, x = x_193)[name = string("op_1207")];
            tensor<int32, [5]> x_195_reps_0 = const()[name = string("x_195_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_195 = tile(reps = x_195_reps_0, x = var_1207)[name = string("x_195")];
            tensor<int32, [4]> var_1210 = const()[name = string("op_1210"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_11 = reshape(shape = var_1210, x = x_195)[name = string("value_11")];
            bool var_1215_transpose_x_1 = const()[name = string("op_1215_transpose_x_1"), val = bool(false)];
            bool var_1215_transpose_y_1 = const()[name = string("op_1215_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_1215_cast_fp16 = matmul(transpose_x = var_1215_transpose_x_1, transpose_y = var_1215_transpose_y_1, x = query_11, y = key_11)[name = string("op_1215_cast_fp16")];
            fp16 var_1216_to_fp16 = const()[name = string("op_1216_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_31_cast_fp16 = mul(x = var_1215_cast_fp16, y = var_1216_to_fp16)[name = string("attn_weights_31_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_33_cast_fp16 = add(x = attn_weights_31_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_33_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_1220_cast_fp16 = softmax(axis = var_18, x = attn_weights_33_cast_fp16)[name = string("op_1220_cast_fp16")];
            bool var_1224_transpose_x_0 = const()[name = string("op_1224_transpose_x_0"), val = bool(false)];
            bool var_1224_transpose_y_0 = const()[name = string("op_1224_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1224_cast_fp16 = matmul(transpose_x = var_1224_transpose_x_0, transpose_y = var_1224_transpose_y_0, x = var_1220_cast_fp16, y = value_11)[name = string("op_1224_cast_fp16")];
            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1229 = const()[name = string("op_1229"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_1227 = transpose(perm = var_1226, x = var_1224_cast_fp16)[name = string("transpose_202")];
            tensor<fp16, [1, 512, 2048]> attn_out_33 = reshape(shape = var_1229, x = var_1227)[name = string("attn_out_33")];
            tensor<int32, [3]> var_1231 = const()[name = string("op_1231"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_5 = const()[name = string("squeeze_5"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1112329472)))];
            string var_1240_pad_type_0 = const()[name = string("op_1240_pad_type_0"), val = string("valid")];
            int32 var_1240_groups_0 = const()[name = string("op_1240_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1240_strides_0 = const()[name = string("op_1240_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1240_pad_0 = const()[name = string("op_1240_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1240_dilations_0 = const()[name = string("op_1240_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_1232 = transpose(perm = var_1231, x = attn_out_33)[name = string("transpose_201")];
            tensor<fp16, [1, 1024, 512]> var_1240 = conv(dilations = var_1240_dilations_0, groups = var_1240_groups_0, pad = var_1240_pad_0, pad_type = var_1240_pad_type_0, strides = var_1240_strides_0, weight = squeeze_5, x = var_1232)[name = string("op_1240")];
            tensor<int32, [3]> var_1241 = const()[name = string("op_1241"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_35 = transpose(perm = var_1241, x = var_1240)[name = string("transpose_200")];
            tensor<fp16, [1, 512, 1024]> x_197_cast_fp16 = add(x = hidden_states_11_cast_fp16, y = attn_out_35)[name = string("x_197_cast_fp16")];
            fp16 var_6_promoted_23_to_fp16 = const()[name = string("op_6_promoted_23_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1247_cast_fp16 = pow(x = x_197_cast_fp16, y = var_6_promoted_23_to_fp16)[name = string("op_1247_cast_fp16")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = string("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_47_keep_dims_0 = const()[name = string("var_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_47_cast_fp16 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1247_cast_fp16)[name = string("var_47_cast_fp16")];
            fp16 var_1250_to_fp16 = const()[name = string("op_1250_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1251_cast_fp16 = add(x = var_47_cast_fp16, y = var_1250_to_fp16)[name = string("op_1251_cast_fp16")];
            fp32 var_1252_epsilon_0 = const()[name = string("op_1252_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1252_cast_fp16 = rsqrt(epsilon = var_1252_epsilon_0, x = var_1251_cast_fp16)[name = string("op_1252_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_201_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1252_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116523840)))];
            tensor<fp16, [1, 512, 1024]> var_1255_cast_fp16 = mul(x = x_201_cast_fp16, y = encoder_layers_5_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1255_cast_fp16")];
            tensor<int32, [3]> var_1260 = const()[name = string("op_1260"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_55_axes_0 = const()[name = string("input_55_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1261 = transpose(perm = var_1260, x = var_1255_cast_fp16)[name = string("transpose_199")];
            tensor<fp16, [1, 1024, 1, 512]> input_55 = expand_dims(axes = input_55_axes_0, x = var_1261)[name = string("input_55")];
            string input_57_pad_type_0 = const()[name = string("input_57_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_57_strides_0 = const()[name = string("input_57_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_57_pad_0 = const()[name = string("input_57_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_57_dilations_0 = const()[name = string("input_57_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_57_groups_0 = const()[name = string("input_57_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_57 = conv(dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = encoder_layers_5_mlp_gate_proj_weight, x = input_55)[name = string("input_57")];
            string up_11_pad_type_0 = const()[name = string("up_11_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_11_strides_0 = const()[name = string("up_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_11_pad_0 = const()[name = string("up_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_11_dilations_0 = const()[name = string("up_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_11_groups_0 = const()[name = string("up_11_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_11 = conv(dilations = up_11_dilations_0, groups = up_11_groups_0, pad = up_11_pad_0, pad_type = up_11_pad_type_0, strides = up_11_strides_0, weight = encoder_layers_5_mlp_up_proj_weight, x = input_55)[name = string("up_11")];
            tensor<fp16, [1, 3072, 1, 512]> var_1275 = silu(x = input_57)[name = string("op_1275")];
            tensor<fp16, [1, 3072, 1, 512]> input_59 = mul(x = var_1275, y = up_11)[name = string("input_59")];
            string var_1282_pad_type_0 = const()[name = string("op_1282_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1282_strides_0 = const()[name = string("op_1282_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1282_pad_0 = const()[name = string("op_1282_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1282_dilations_0 = const()[name = string("op_1282_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1282_groups_0 = const()[name = string("op_1282_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1282 = conv(dilations = var_1282_dilations_0, groups = var_1282_groups_0, pad = var_1282_pad_0, pad_type = var_1282_pad_type_0, strides = var_1282_strides_0, weight = encoder_layers_5_mlp_down_proj_weight, x = input_59)[name = string("op_1282")];
            tensor<int32, [1]> var_1283_axes_0 = const()[name = string("op_1283_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1283 = squeeze(axes = var_1283_axes_0, x = var_1282)[name = string("op_1283")];
            tensor<int32, [3]> var_1284 = const()[name = string("op_1284"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_11 = transpose(perm = var_1284, x = var_1283)[name = string("transpose_198")];
            tensor<fp16, [1, 512, 1024]> hidden_states_13_cast_fp16 = add(x = x_197_cast_fp16, y = mlp_out_11)[name = string("hidden_states_13_cast_fp16")];
            fp16 var_6_promoted_24_to_fp16 = const()[name = string("op_6_promoted_24_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1311_cast_fp16 = pow(x = hidden_states_13_cast_fp16, y = var_6_promoted_24_to_fp16)[name = string("op_1311_cast_fp16")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = string("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_49_keep_dims_0 = const()[name = string("var_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_49_cast_fp16 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1311_cast_fp16)[name = string("var_49_cast_fp16")];
            fp16 var_1314_to_fp16 = const()[name = string("op_1314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1315_cast_fp16 = add(x = var_49_cast_fp16, y = var_1314_to_fp16)[name = string("op_1315_cast_fp16")];
            fp32 var_1316_epsilon_0 = const()[name = string("op_1316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1316_cast_fp16 = rsqrt(epsilon = var_1316_epsilon_0, x = var_1315_cast_fp16)[name = string("op_1316_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_207_cast_fp16 = mul(x = hidden_states_13_cast_fp16, y = var_1316_cast_fp16)[name = string("x_207_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_6_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_6_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116525952)))];
            tensor<fp16, [1, 512, 1024]> var_1319_cast_fp16 = mul(x = x_207_cast_fp16, y = encoder_layers_6_input_layernorm_weight_promoted_to_fp16)[name = string("op_1319_cast_fp16")];
            tensor<int32, [3]> var_1324 = const()[name = string("op_1324"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1325 = transpose(perm = var_1324, x = var_1319_cast_fp16)[name = string("transpose_197")];
            tensor<fp16, [1, 1024, 1, 512]> input_61 = expand_dims(axes = input_61_axes_0, x = var_1325)[name = string("input_61")];
            string var_1332_pad_type_0 = const()[name = string("op_1332_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1332_strides_0 = const()[name = string("op_1332_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1332_pad_0 = const()[name = string("op_1332_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1332_dilations_0 = const()[name = string("op_1332_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1332_groups_0 = const()[name = string("op_1332_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_1332 = conv(dilations = var_1332_dilations_0, groups = var_1332_groups_0, pad = var_1332_pad_0, pad_type = var_1332_pad_type_0, strides = var_1332_strides_0, weight = encoder_layers_6_self_attn_q_proj_weight, x = input_61)[name = string("op_1332")];
            tensor<int32, [4]> var_1333 = const()[name = string("op_1333"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_1334 = reshape(shape = var_1333, x = var_1332)[name = string("op_1334")];
            tensor<int32, [4]> var_1335 = const()[name = string("op_1335"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1342_pad_type_0 = const()[name = string("op_1342_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1342_strides_0 = const()[name = string("op_1342_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1342_pad_0 = const()[name = string("op_1342_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1342_dilations_0 = const()[name = string("op_1342_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1342_groups_0 = const()[name = string("op_1342_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1342 = conv(dilations = var_1342_dilations_0, groups = var_1342_groups_0, pad = var_1342_pad_0, pad_type = var_1342_pad_type_0, strides = var_1342_strides_0, weight = encoder_layers_6_self_attn_k_proj_weight, x = input_61)[name = string("op_1342")];
            tensor<int32, [4]> var_1343 = const()[name = string("op_1343"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1344 = reshape(shape = var_1343, x = var_1342)[name = string("op_1344")];
            tensor<int32, [4]> var_1345 = const()[name = string("op_1345"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1352_pad_type_0 = const()[name = string("op_1352_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1352_strides_0 = const()[name = string("op_1352_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1352_pad_0 = const()[name = string("op_1352_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1352_dilations_0 = const()[name = string("op_1352_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1352_groups_0 = const()[name = string("op_1352_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1352 = conv(dilations = var_1352_dilations_0, groups = var_1352_groups_0, pad = var_1352_pad_0, pad_type = var_1352_pad_type_0, strides = var_1352_strides_0, weight = encoder_layers_6_self_attn_v_proj_weight, x = input_61)[name = string("op_1352")];
            tensor<int32, [4]> var_1353 = const()[name = string("op_1353"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1354 = reshape(shape = var_1353, x = var_1352)[name = string("op_1354")];
            tensor<int32, [4]> var_1355 = const()[name = string("op_1355"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_25_to_fp16 = const()[name = string("op_6_promoted_25_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_37 = transpose(perm = var_1335, x = var_1334)[name = string("transpose_196")];
            tensor<fp16, [1, 16, 512, 128]> var_1361_cast_fp16 = pow(x = q_37, y = var_6_promoted_25_to_fp16)[name = string("op_1361_cast_fp16")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = string("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_51_keep_dims_0 = const()[name = string("var_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_51_cast_fp16 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_1361_cast_fp16)[name = string("var_51_cast_fp16")];
            fp16 var_1364_to_fp16 = const()[name = string("op_1364_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_1365_cast_fp16 = add(x = var_51_cast_fp16, y = var_1364_to_fp16)[name = string("op_1365_cast_fp16")];
            fp32 var_1366_epsilon_0 = const()[name = string("op_1366_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_1366_cast_fp16 = rsqrt(epsilon = var_1366_epsilon_0, x = var_1365_cast_fp16)[name = string("op_1366_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_215_cast_fp16 = mul(x = q_37, y = var_1366_cast_fp16)[name = string("x_215_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_39 = mul(x = x_215_cast_fp16, y = encoder_layers_6_self_attn_q_norm_weight)[name = string("q_39")];
            fp16 var_6_promoted_26_to_fp16 = const()[name = string("op_6_promoted_26_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_37 = transpose(perm = var_1345, x = var_1344)[name = string("transpose_195")];
            tensor<fp16, [1, 8, 512, 128]> var_1374_cast_fp16 = pow(x = k_37, y = var_6_promoted_26_to_fp16)[name = string("op_1374_cast_fp16")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = string("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_53_keep_dims_0 = const()[name = string("var_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_53_cast_fp16 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_1374_cast_fp16)[name = string("var_53_cast_fp16")];
            fp16 var_1377_to_fp16 = const()[name = string("op_1377_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_1378_cast_fp16 = add(x = var_53_cast_fp16, y = var_1377_to_fp16)[name = string("op_1378_cast_fp16")];
            fp32 var_1379_epsilon_0 = const()[name = string("op_1379_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_1379_cast_fp16 = rsqrt(epsilon = var_1379_epsilon_0, x = var_1378_cast_fp16)[name = string("op_1379_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_221_cast_fp16 = mul(x = k_37, y = var_1379_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_39 = mul(x = x_221_cast_fp16, y = encoder_layers_6_self_attn_k_norm_weight)[name = string("k_39")];
            tensor<fp16, [1, 16, 512, 128]> var_1383 = mul(x = q_39, y = cos)[name = string("op_1383")];
            tensor<int32, [2]> var_1384_split_sizes_0 = const()[name = string("op_1384_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1384_axis_0 = const()[name = string("op_1384_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_1384_0, tensor<fp16, [1, 16, 512, 64]> var_1384_1 = split(axis = var_1384_axis_0, split_sizes = var_1384_split_sizes_0, x = q_39)[name = string("op_1384")];
            fp16 const_21_promoted = const()[name = string("const_21_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_1386 = mul(x = var_1384_1, y = const_21_promoted)[name = string("op_1386")];
            bool var_1388_interleave_0 = const()[name = string("op_1388_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1388 = concat(axis = var_18, interleave = var_1388_interleave_0, values = (var_1386, var_1384_0))[name = string("op_1388")];
            tensor<fp16, [1, 16, 512, 128]> var_1389 = mul(x = var_1388, y = sin)[name = string("op_1389")];
            tensor<fp16, [1, 16, 512, 128]> query_13 = add(x = var_1383, y = var_1389)[name = string("query_13")];
            tensor<fp16, [1, 8, 512, 128]> var_1391 = mul(x = k_39, y = cos)[name = string("op_1391")];
            tensor<int32, [2]> var_1392_split_sizes_0 = const()[name = string("op_1392_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1392_axis_0 = const()[name = string("op_1392_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_1392_0, tensor<fp16, [1, 8, 512, 64]> var_1392_1 = split(axis = var_1392_axis_0, split_sizes = var_1392_split_sizes_0, x = k_39)[name = string("op_1392")];
            fp16 const_22_promoted = const()[name = string("const_22_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1394 = mul(x = var_1392_1, y = const_22_promoted)[name = string("op_1394")];
            bool var_1396_interleave_0 = const()[name = string("op_1396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1396 = concat(axis = var_18, interleave = var_1396_interleave_0, values = (var_1394, var_1392_0))[name = string("op_1396")];
            tensor<fp16, [1, 8, 512, 128]> var_1397 = mul(x = var_1396, y = sin)[name = string("op_1397")];
            tensor<fp16, [1, 8, 512, 128]> x_223 = add(x = var_1391, y = var_1397)[name = string("x_223")];
            tensor<int32, [1]> var_1399_axes_0 = const()[name = string("op_1399_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1399 = expand_dims(axes = var_1399_axes_0, x = x_223)[name = string("op_1399")];
            tensor<int32, [5]> x_225_reps_0 = const()[name = string("x_225_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_225 = tile(reps = x_225_reps_0, x = var_1399)[name = string("x_225")];
            tensor<int32, [4]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_13 = reshape(shape = var_1402, x = x_225)[name = string("key_13")];
            tensor<int32, [1]> var_1404_axes_0 = const()[name = string("op_1404_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_227 = transpose(perm = var_1355, x = var_1354)[name = string("transpose_194")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1404 = expand_dims(axes = var_1404_axes_0, x = x_227)[name = string("op_1404")];
            tensor<int32, [5]> x_229_reps_0 = const()[name = string("x_229_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_229 = tile(reps = x_229_reps_0, x = var_1404)[name = string("x_229")];
            tensor<int32, [4]> var_1407 = const()[name = string("op_1407"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_13 = reshape(shape = var_1407, x = x_229)[name = string("value_13")];
            bool var_1412_transpose_x_1 = const()[name = string("op_1412_transpose_x_1"), val = bool(false)];
            bool var_1412_transpose_y_1 = const()[name = string("op_1412_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_1412_cast_fp16 = matmul(transpose_x = var_1412_transpose_x_1, transpose_y = var_1412_transpose_y_1, x = query_13, y = key_13)[name = string("op_1412_cast_fp16")];
            fp16 var_1413_to_fp16 = const()[name = string("op_1413_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_37_cast_fp16 = mul(x = var_1412_cast_fp16, y = var_1413_to_fp16)[name = string("attn_weights_37_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_39_cast_fp16 = add(x = attn_weights_37_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_39_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_1417_cast_fp16 = softmax(axis = var_18, x = attn_weights_39_cast_fp16)[name = string("op_1417_cast_fp16")];
            bool var_1421_transpose_x_0 = const()[name = string("op_1421_transpose_x_0"), val = bool(false)];
            bool var_1421_transpose_y_0 = const()[name = string("op_1421_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1421_cast_fp16 = matmul(transpose_x = var_1421_transpose_x_0, transpose_y = var_1421_transpose_y_0, x = var_1417_cast_fp16, y = value_13)[name = string("op_1421_cast_fp16")];
            tensor<int32, [4]> var_1423 = const()[name = string("op_1423"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1426 = const()[name = string("op_1426"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_1424 = transpose(perm = var_1423, x = var_1421_cast_fp16)[name = string("transpose_193")];
            tensor<fp16, [1, 512, 2048]> attn_out_39 = reshape(shape = var_1426, x = var_1424)[name = string("attn_out_39")];
            tensor<int32, [3]> var_1428 = const()[name = string("op_1428"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_6 = const()[name = string("squeeze_6"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1116528064)))];
            string var_1437_pad_type_0 = const()[name = string("op_1437_pad_type_0"), val = string("valid")];
            int32 var_1437_groups_0 = const()[name = string("op_1437_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1437_strides_0 = const()[name = string("op_1437_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1437_pad_0 = const()[name = string("op_1437_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1437_dilations_0 = const()[name = string("op_1437_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_1429 = transpose(perm = var_1428, x = attn_out_39)[name = string("transpose_192")];
            tensor<fp16, [1, 1024, 512]> var_1437 = conv(dilations = var_1437_dilations_0, groups = var_1437_groups_0, pad = var_1437_pad_0, pad_type = var_1437_pad_type_0, strides = var_1437_strides_0, weight = squeeze_6, x = var_1429)[name = string("op_1437")];
            tensor<int32, [3]> var_1438 = const()[name = string("op_1438"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_41 = transpose(perm = var_1438, x = var_1437)[name = string("transpose_191")];
            tensor<fp16, [1, 512, 1024]> x_231_cast_fp16 = add(x = hidden_states_13_cast_fp16, y = attn_out_41)[name = string("x_231_cast_fp16")];
            fp16 var_6_promoted_27_to_fp16 = const()[name = string("op_6_promoted_27_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1444_cast_fp16 = pow(x = x_231_cast_fp16, y = var_6_promoted_27_to_fp16)[name = string("op_1444_cast_fp16")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = string("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_55_keep_dims_0 = const()[name = string("var_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_55_cast_fp16 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_1444_cast_fp16)[name = string("var_55_cast_fp16")];
            fp16 var_1447_to_fp16 = const()[name = string("op_1447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1448_cast_fp16 = add(x = var_55_cast_fp16, y = var_1447_to_fp16)[name = string("op_1448_cast_fp16")];
            fp32 var_1449_epsilon_0 = const()[name = string("op_1449_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1449_cast_fp16 = rsqrt(epsilon = var_1449_epsilon_0, x = var_1448_cast_fp16)[name = string("op_1449_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_235_cast_fp16 = mul(x = x_231_cast_fp16, y = var_1449_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120722432)))];
            tensor<fp16, [1, 512, 1024]> var_1452_cast_fp16 = mul(x = x_235_cast_fp16, y = encoder_layers_6_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1452_cast_fp16")];
            tensor<int32, [3]> var_1457 = const()[name = string("op_1457"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1458 = transpose(perm = var_1457, x = var_1452_cast_fp16)[name = string("transpose_190")];
            tensor<fp16, [1, 1024, 1, 512]> input_65 = expand_dims(axes = input_65_axes_0, x = var_1458)[name = string("input_65")];
            string input_67_pad_type_0 = const()[name = string("input_67_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_67_strides_0 = const()[name = string("input_67_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_67_pad_0 = const()[name = string("input_67_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_67_dilations_0 = const()[name = string("input_67_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_67_groups_0 = const()[name = string("input_67_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_67 = conv(dilations = input_67_dilations_0, groups = input_67_groups_0, pad = input_67_pad_0, pad_type = input_67_pad_type_0, strides = input_67_strides_0, weight = encoder_layers_6_mlp_gate_proj_weight, x = input_65)[name = string("input_67")];
            string up_13_pad_type_0 = const()[name = string("up_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_13_strides_0 = const()[name = string("up_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_13_pad_0 = const()[name = string("up_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_13_dilations_0 = const()[name = string("up_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_13_groups_0 = const()[name = string("up_13_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_13 = conv(dilations = up_13_dilations_0, groups = up_13_groups_0, pad = up_13_pad_0, pad_type = up_13_pad_type_0, strides = up_13_strides_0, weight = encoder_layers_6_mlp_up_proj_weight, x = input_65)[name = string("up_13")];
            tensor<fp16, [1, 3072, 1, 512]> var_1472 = silu(x = input_67)[name = string("op_1472")];
            tensor<fp16, [1, 3072, 1, 512]> input_69 = mul(x = var_1472, y = up_13)[name = string("input_69")];
            string var_1479_pad_type_0 = const()[name = string("op_1479_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1479_strides_0 = const()[name = string("op_1479_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1479_pad_0 = const()[name = string("op_1479_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1479_dilations_0 = const()[name = string("op_1479_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1479_groups_0 = const()[name = string("op_1479_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1479 = conv(dilations = var_1479_dilations_0, groups = var_1479_groups_0, pad = var_1479_pad_0, pad_type = var_1479_pad_type_0, strides = var_1479_strides_0, weight = encoder_layers_6_mlp_down_proj_weight, x = input_69)[name = string("op_1479")];
            tensor<int32, [1]> var_1480_axes_0 = const()[name = string("op_1480_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1480 = squeeze(axes = var_1480_axes_0, x = var_1479)[name = string("op_1480")];
            tensor<int32, [3]> var_1481 = const()[name = string("op_1481"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_13 = transpose(perm = var_1481, x = var_1480)[name = string("transpose_189")];
            tensor<fp16, [1, 512, 1024]> hidden_states_15_cast_fp16 = add(x = x_231_cast_fp16, y = mlp_out_13)[name = string("hidden_states_15_cast_fp16")];
            fp16 var_6_promoted_28_to_fp16 = const()[name = string("op_6_promoted_28_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1508_cast_fp16 = pow(x = hidden_states_15_cast_fp16, y = var_6_promoted_28_to_fp16)[name = string("op_1508_cast_fp16")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = string("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_57_keep_dims_0 = const()[name = string("var_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_57_cast_fp16 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_1508_cast_fp16)[name = string("var_57_cast_fp16")];
            fp16 var_1511_to_fp16 = const()[name = string("op_1511_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1512_cast_fp16 = add(x = var_57_cast_fp16, y = var_1511_to_fp16)[name = string("op_1512_cast_fp16")];
            fp32 var_1513_epsilon_0 = const()[name = string("op_1513_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1513_cast_fp16 = rsqrt(epsilon = var_1513_epsilon_0, x = var_1512_cast_fp16)[name = string("op_1513_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_241_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = var_1513_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_7_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_7_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120724544)))];
            tensor<fp16, [1, 512, 1024]> var_1516_cast_fp16 = mul(x = x_241_cast_fp16, y = encoder_layers_7_input_layernorm_weight_promoted_to_fp16)[name = string("op_1516_cast_fp16")];
            tensor<int32, [3]> var_1521 = const()[name = string("op_1521"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1522 = transpose(perm = var_1521, x = var_1516_cast_fp16)[name = string("transpose_188")];
            tensor<fp16, [1, 1024, 1, 512]> input_71 = expand_dims(axes = input_71_axes_0, x = var_1522)[name = string("input_71")];
            string var_1529_pad_type_0 = const()[name = string("op_1529_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1529_strides_0 = const()[name = string("op_1529_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1529_pad_0 = const()[name = string("op_1529_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1529_dilations_0 = const()[name = string("op_1529_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1529_groups_0 = const()[name = string("op_1529_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_1529 = conv(dilations = var_1529_dilations_0, groups = var_1529_groups_0, pad = var_1529_pad_0, pad_type = var_1529_pad_type_0, strides = var_1529_strides_0, weight = encoder_layers_7_self_attn_q_proj_weight, x = input_71)[name = string("op_1529")];
            tensor<int32, [4]> var_1530 = const()[name = string("op_1530"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_1531 = reshape(shape = var_1530, x = var_1529)[name = string("op_1531")];
            tensor<int32, [4]> var_1532 = const()[name = string("op_1532"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1539_pad_type_0 = const()[name = string("op_1539_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1539_strides_0 = const()[name = string("op_1539_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1539_pad_0 = const()[name = string("op_1539_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1539_dilations_0 = const()[name = string("op_1539_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1539_groups_0 = const()[name = string("op_1539_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1539 = conv(dilations = var_1539_dilations_0, groups = var_1539_groups_0, pad = var_1539_pad_0, pad_type = var_1539_pad_type_0, strides = var_1539_strides_0, weight = encoder_layers_7_self_attn_k_proj_weight, x = input_71)[name = string("op_1539")];
            tensor<int32, [4]> var_1540 = const()[name = string("op_1540"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1541 = reshape(shape = var_1540, x = var_1539)[name = string("op_1541")];
            tensor<int32, [4]> var_1542 = const()[name = string("op_1542"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1549_pad_type_0 = const()[name = string("op_1549_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1549_strides_0 = const()[name = string("op_1549_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1549_pad_0 = const()[name = string("op_1549_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1549_dilations_0 = const()[name = string("op_1549_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1549_groups_0 = const()[name = string("op_1549_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1549 = conv(dilations = var_1549_dilations_0, groups = var_1549_groups_0, pad = var_1549_pad_0, pad_type = var_1549_pad_type_0, strides = var_1549_strides_0, weight = encoder_layers_7_self_attn_v_proj_weight, x = input_71)[name = string("op_1549")];
            tensor<int32, [4]> var_1550 = const()[name = string("op_1550"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1551 = reshape(shape = var_1550, x = var_1549)[name = string("op_1551")];
            tensor<int32, [4]> var_1552 = const()[name = string("op_1552"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_29_to_fp16 = const()[name = string("op_6_promoted_29_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_43 = transpose(perm = var_1532, x = var_1531)[name = string("transpose_187")];
            tensor<fp16, [1, 16, 512, 128]> var_1558_cast_fp16 = pow(x = q_43, y = var_6_promoted_29_to_fp16)[name = string("op_1558_cast_fp16")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = string("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_59_keep_dims_0 = const()[name = string("var_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_59_cast_fp16 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_1558_cast_fp16)[name = string("var_59_cast_fp16")];
            fp16 var_1561_to_fp16 = const()[name = string("op_1561_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_1562_cast_fp16 = add(x = var_59_cast_fp16, y = var_1561_to_fp16)[name = string("op_1562_cast_fp16")];
            fp32 var_1563_epsilon_0 = const()[name = string("op_1563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_1563_cast_fp16 = rsqrt(epsilon = var_1563_epsilon_0, x = var_1562_cast_fp16)[name = string("op_1563_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_249_cast_fp16 = mul(x = q_43, y = var_1563_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_45 = mul(x = x_249_cast_fp16, y = encoder_layers_7_self_attn_q_norm_weight)[name = string("q_45")];
            fp16 var_6_promoted_30_to_fp16 = const()[name = string("op_6_promoted_30_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_43 = transpose(perm = var_1542, x = var_1541)[name = string("transpose_186")];
            tensor<fp16, [1, 8, 512, 128]> var_1571_cast_fp16 = pow(x = k_43, y = var_6_promoted_30_to_fp16)[name = string("op_1571_cast_fp16")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = string("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_61_keep_dims_0 = const()[name = string("var_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_61_cast_fp16 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_1571_cast_fp16)[name = string("var_61_cast_fp16")];
            fp16 var_1574_to_fp16 = const()[name = string("op_1574_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_1575_cast_fp16 = add(x = var_61_cast_fp16, y = var_1574_to_fp16)[name = string("op_1575_cast_fp16")];
            fp32 var_1576_epsilon_0 = const()[name = string("op_1576_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_1576_cast_fp16 = rsqrt(epsilon = var_1576_epsilon_0, x = var_1575_cast_fp16)[name = string("op_1576_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_255_cast_fp16 = mul(x = k_43, y = var_1576_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_45 = mul(x = x_255_cast_fp16, y = encoder_layers_7_self_attn_k_norm_weight)[name = string("k_45")];
            tensor<fp16, [1, 16, 512, 128]> var_1580 = mul(x = q_45, y = cos)[name = string("op_1580")];
            tensor<int32, [2]> var_1581_split_sizes_0 = const()[name = string("op_1581_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1581_axis_0 = const()[name = string("op_1581_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_1581_0, tensor<fp16, [1, 16, 512, 64]> var_1581_1 = split(axis = var_1581_axis_0, split_sizes = var_1581_split_sizes_0, x = q_45)[name = string("op_1581")];
            fp16 const_24_promoted = const()[name = string("const_24_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_1583 = mul(x = var_1581_1, y = const_24_promoted)[name = string("op_1583")];
            bool var_1585_interleave_0 = const()[name = string("op_1585_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1585 = concat(axis = var_18, interleave = var_1585_interleave_0, values = (var_1583, var_1581_0))[name = string("op_1585")];
            tensor<fp16, [1, 16, 512, 128]> var_1586 = mul(x = var_1585, y = sin)[name = string("op_1586")];
            tensor<fp16, [1, 16, 512, 128]> query_15 = add(x = var_1580, y = var_1586)[name = string("query_15")];
            tensor<fp16, [1, 8, 512, 128]> var_1588 = mul(x = k_45, y = cos)[name = string("op_1588")];
            tensor<int32, [2]> var_1589_split_sizes_0 = const()[name = string("op_1589_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1589_axis_0 = const()[name = string("op_1589_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_1589_0, tensor<fp16, [1, 8, 512, 64]> var_1589_1 = split(axis = var_1589_axis_0, split_sizes = var_1589_split_sizes_0, x = k_45)[name = string("op_1589")];
            fp16 const_25_promoted = const()[name = string("const_25_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1591 = mul(x = var_1589_1, y = const_25_promoted)[name = string("op_1591")];
            bool var_1593_interleave_0 = const()[name = string("op_1593_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1593 = concat(axis = var_18, interleave = var_1593_interleave_0, values = (var_1591, var_1589_0))[name = string("op_1593")];
            tensor<fp16, [1, 8, 512, 128]> var_1594 = mul(x = var_1593, y = sin)[name = string("op_1594")];
            tensor<fp16, [1, 8, 512, 128]> x_257 = add(x = var_1588, y = var_1594)[name = string("x_257")];
            tensor<int32, [1]> var_1596_axes_0 = const()[name = string("op_1596_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1596 = expand_dims(axes = var_1596_axes_0, x = x_257)[name = string("op_1596")];
            tensor<int32, [5]> x_259_reps_0 = const()[name = string("x_259_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_259 = tile(reps = x_259_reps_0, x = var_1596)[name = string("x_259")];
            tensor<int32, [4]> var_1599 = const()[name = string("op_1599"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_15 = reshape(shape = var_1599, x = x_259)[name = string("key_15")];
            tensor<int32, [1]> var_1601_axes_0 = const()[name = string("op_1601_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_261 = transpose(perm = var_1552, x = var_1551)[name = string("transpose_185")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1601 = expand_dims(axes = var_1601_axes_0, x = x_261)[name = string("op_1601")];
            tensor<int32, [5]> x_263_reps_0 = const()[name = string("x_263_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_263 = tile(reps = x_263_reps_0, x = var_1601)[name = string("x_263")];
            tensor<int32, [4]> var_1604 = const()[name = string("op_1604"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_15 = reshape(shape = var_1604, x = x_263)[name = string("value_15")];
            bool var_1609_transpose_x_1 = const()[name = string("op_1609_transpose_x_1"), val = bool(false)];
            bool var_1609_transpose_y_1 = const()[name = string("op_1609_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_1609_cast_fp16 = matmul(transpose_x = var_1609_transpose_x_1, transpose_y = var_1609_transpose_y_1, x = query_15, y = key_15)[name = string("op_1609_cast_fp16")];
            fp16 var_1610_to_fp16 = const()[name = string("op_1610_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_43_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = string("attn_weights_43_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_45_cast_fp16 = add(x = attn_weights_43_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_45_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_1614_cast_fp16 = softmax(axis = var_18, x = attn_weights_45_cast_fp16)[name = string("op_1614_cast_fp16")];
            bool var_1618_transpose_x_0 = const()[name = string("op_1618_transpose_x_0"), val = bool(false)];
            bool var_1618_transpose_y_0 = const()[name = string("op_1618_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1618_cast_fp16 = matmul(transpose_x = var_1618_transpose_x_0, transpose_y = var_1618_transpose_y_0, x = var_1614_cast_fp16, y = value_15)[name = string("op_1618_cast_fp16")];
            tensor<int32, [4]> var_1620 = const()[name = string("op_1620"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1623 = const()[name = string("op_1623"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_1621 = transpose(perm = var_1620, x = var_1618_cast_fp16)[name = string("transpose_184")];
            tensor<fp16, [1, 512, 2048]> attn_out_45 = reshape(shape = var_1623, x = var_1621)[name = string("attn_out_45")];
            tensor<int32, [3]> var_1625 = const()[name = string("op_1625"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_7 = const()[name = string("squeeze_7"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1120726656)))];
            string var_1634_pad_type_0 = const()[name = string("op_1634_pad_type_0"), val = string("valid")];
            int32 var_1634_groups_0 = const()[name = string("op_1634_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1634_strides_0 = const()[name = string("op_1634_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1634_pad_0 = const()[name = string("op_1634_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1634_dilations_0 = const()[name = string("op_1634_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_1626 = transpose(perm = var_1625, x = attn_out_45)[name = string("transpose_183")];
            tensor<fp16, [1, 1024, 512]> var_1634 = conv(dilations = var_1634_dilations_0, groups = var_1634_groups_0, pad = var_1634_pad_0, pad_type = var_1634_pad_type_0, strides = var_1634_strides_0, weight = squeeze_7, x = var_1626)[name = string("op_1634")];
            tensor<int32, [3]> var_1635 = const()[name = string("op_1635"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_47 = transpose(perm = var_1635, x = var_1634)[name = string("transpose_182")];
            tensor<fp16, [1, 512, 1024]> x_265_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = attn_out_47)[name = string("x_265_cast_fp16")];
            fp16 var_6_promoted_31_to_fp16 = const()[name = string("op_6_promoted_31_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1641_cast_fp16 = pow(x = x_265_cast_fp16, y = var_6_promoted_31_to_fp16)[name = string("op_1641_cast_fp16")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = string("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_63_keep_dims_0 = const()[name = string("var_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_63_cast_fp16 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_1641_cast_fp16)[name = string("var_63_cast_fp16")];
            fp16 var_1644_to_fp16 = const()[name = string("op_1644_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1645_cast_fp16 = add(x = var_63_cast_fp16, y = var_1644_to_fp16)[name = string("op_1645_cast_fp16")];
            fp32 var_1646_epsilon_0 = const()[name = string("op_1646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1646_cast_fp16 = rsqrt(epsilon = var_1646_epsilon_0, x = var_1645_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_269_cast_fp16 = mul(x = x_265_cast_fp16, y = var_1646_cast_fp16)[name = string("x_269_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124921024)))];
            tensor<fp16, [1, 512, 1024]> var_1649_cast_fp16 = mul(x = x_269_cast_fp16, y = encoder_layers_7_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1649_cast_fp16")];
            tensor<int32, [3]> var_1654 = const()[name = string("op_1654"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_75_axes_0 = const()[name = string("input_75_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1655 = transpose(perm = var_1654, x = var_1649_cast_fp16)[name = string("transpose_181")];
            tensor<fp16, [1, 1024, 1, 512]> input_75 = expand_dims(axes = input_75_axes_0, x = var_1655)[name = string("input_75")];
            string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_77 = conv(dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = encoder_layers_7_mlp_gate_proj_weight, x = input_75)[name = string("input_77")];
            string up_15_pad_type_0 = const()[name = string("up_15_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_15_strides_0 = const()[name = string("up_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_15_pad_0 = const()[name = string("up_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_15_dilations_0 = const()[name = string("up_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_15_groups_0 = const()[name = string("up_15_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_15 = conv(dilations = up_15_dilations_0, groups = up_15_groups_0, pad = up_15_pad_0, pad_type = up_15_pad_type_0, strides = up_15_strides_0, weight = encoder_layers_7_mlp_up_proj_weight, x = input_75)[name = string("up_15")];
            tensor<fp16, [1, 3072, 1, 512]> var_1669 = silu(x = input_77)[name = string("op_1669")];
            tensor<fp16, [1, 3072, 1, 512]> input_79 = mul(x = var_1669, y = up_15)[name = string("input_79")];
            string var_1676_pad_type_0 = const()[name = string("op_1676_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1676_strides_0 = const()[name = string("op_1676_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1676_pad_0 = const()[name = string("op_1676_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1676_dilations_0 = const()[name = string("op_1676_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1676_groups_0 = const()[name = string("op_1676_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1676 = conv(dilations = var_1676_dilations_0, groups = var_1676_groups_0, pad = var_1676_pad_0, pad_type = var_1676_pad_type_0, strides = var_1676_strides_0, weight = encoder_layers_7_mlp_down_proj_weight, x = input_79)[name = string("op_1676")];
            tensor<int32, [1]> var_1677_axes_0 = const()[name = string("op_1677_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1677 = squeeze(axes = var_1677_axes_0, x = var_1676)[name = string("op_1677")];
            tensor<int32, [3]> var_1678 = const()[name = string("op_1678"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_15 = transpose(perm = var_1678, x = var_1677)[name = string("transpose_180")];
            tensor<fp16, [1, 512, 1024]> hidden_states_17_cast_fp16 = add(x = x_265_cast_fp16, y = mlp_out_15)[name = string("hidden_states_17_cast_fp16")];
            fp16 var_6_promoted_32_to_fp16 = const()[name = string("op_6_promoted_32_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1705_cast_fp16 = pow(x = hidden_states_17_cast_fp16, y = var_6_promoted_32_to_fp16)[name = string("op_1705_cast_fp16")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = string("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_65_keep_dims_0 = const()[name = string("var_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_65_cast_fp16 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_1705_cast_fp16)[name = string("var_65_cast_fp16")];
            fp16 var_1708_to_fp16 = const()[name = string("op_1708_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1709_cast_fp16 = add(x = var_65_cast_fp16, y = var_1708_to_fp16)[name = string("op_1709_cast_fp16")];
            fp32 var_1710_epsilon_0 = const()[name = string("op_1710_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1710_cast_fp16 = rsqrt(epsilon = var_1710_epsilon_0, x = var_1709_cast_fp16)[name = string("op_1710_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_275_cast_fp16 = mul(x = hidden_states_17_cast_fp16, y = var_1710_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_8_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_8_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124923136)))];
            tensor<fp16, [1, 512, 1024]> var_1713_cast_fp16 = mul(x = x_275_cast_fp16, y = encoder_layers_8_input_layernorm_weight_promoted_to_fp16)[name = string("op_1713_cast_fp16")];
            tensor<int32, [3]> var_1718 = const()[name = string("op_1718"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1719 = transpose(perm = var_1718, x = var_1713_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 1024, 1, 512]> input_81 = expand_dims(axes = input_81_axes_0, x = var_1719)[name = string("input_81")];
            string var_1726_pad_type_0 = const()[name = string("op_1726_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1726_strides_0 = const()[name = string("op_1726_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1726_pad_0 = const()[name = string("op_1726_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1726_dilations_0 = const()[name = string("op_1726_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1726_groups_0 = const()[name = string("op_1726_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_1726 = conv(dilations = var_1726_dilations_0, groups = var_1726_groups_0, pad = var_1726_pad_0, pad_type = var_1726_pad_type_0, strides = var_1726_strides_0, weight = encoder_layers_8_self_attn_q_proj_weight, x = input_81)[name = string("op_1726")];
            tensor<int32, [4]> var_1727 = const()[name = string("op_1727"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_1728 = reshape(shape = var_1727, x = var_1726)[name = string("op_1728")];
            tensor<int32, [4]> var_1729 = const()[name = string("op_1729"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1736_pad_type_0 = const()[name = string("op_1736_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1736_strides_0 = const()[name = string("op_1736_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1736_pad_0 = const()[name = string("op_1736_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1736_dilations_0 = const()[name = string("op_1736_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1736_groups_0 = const()[name = string("op_1736_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1736 = conv(dilations = var_1736_dilations_0, groups = var_1736_groups_0, pad = var_1736_pad_0, pad_type = var_1736_pad_type_0, strides = var_1736_strides_0, weight = encoder_layers_8_self_attn_k_proj_weight, x = input_81)[name = string("op_1736")];
            tensor<int32, [4]> var_1737 = const()[name = string("op_1737"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1738 = reshape(shape = var_1737, x = var_1736)[name = string("op_1738")];
            tensor<int32, [4]> var_1739 = const()[name = string("op_1739"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1746_pad_type_0 = const()[name = string("op_1746_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1746_strides_0 = const()[name = string("op_1746_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1746_pad_0 = const()[name = string("op_1746_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1746_dilations_0 = const()[name = string("op_1746_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1746_groups_0 = const()[name = string("op_1746_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1746 = conv(dilations = var_1746_dilations_0, groups = var_1746_groups_0, pad = var_1746_pad_0, pad_type = var_1746_pad_type_0, strides = var_1746_strides_0, weight = encoder_layers_8_self_attn_v_proj_weight, x = input_81)[name = string("op_1746")];
            tensor<int32, [4]> var_1747 = const()[name = string("op_1747"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1748 = reshape(shape = var_1747, x = var_1746)[name = string("op_1748")];
            tensor<int32, [4]> var_1749 = const()[name = string("op_1749"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_33_to_fp16 = const()[name = string("op_6_promoted_33_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_49 = transpose(perm = var_1729, x = var_1728)[name = string("transpose_178")];
            tensor<fp16, [1, 16, 512, 128]> var_1755_cast_fp16 = pow(x = q_49, y = var_6_promoted_33_to_fp16)[name = string("op_1755_cast_fp16")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = string("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_67_keep_dims_0 = const()[name = string("var_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_67_cast_fp16 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_1755_cast_fp16)[name = string("var_67_cast_fp16")];
            fp16 var_1758_to_fp16 = const()[name = string("op_1758_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_1759_cast_fp16 = add(x = var_67_cast_fp16, y = var_1758_to_fp16)[name = string("op_1759_cast_fp16")];
            fp32 var_1760_epsilon_0 = const()[name = string("op_1760_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_1760_cast_fp16 = rsqrt(epsilon = var_1760_epsilon_0, x = var_1759_cast_fp16)[name = string("op_1760_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_283_cast_fp16 = mul(x = q_49, y = var_1760_cast_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_51 = mul(x = x_283_cast_fp16, y = encoder_layers_8_self_attn_q_norm_weight)[name = string("q_51")];
            fp16 var_6_promoted_34_to_fp16 = const()[name = string("op_6_promoted_34_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_49 = transpose(perm = var_1739, x = var_1738)[name = string("transpose_177")];
            tensor<fp16, [1, 8, 512, 128]> var_1768_cast_fp16 = pow(x = k_49, y = var_6_promoted_34_to_fp16)[name = string("op_1768_cast_fp16")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = string("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_69_keep_dims_0 = const()[name = string("var_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_69_cast_fp16 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_1768_cast_fp16)[name = string("var_69_cast_fp16")];
            fp16 var_1771_to_fp16 = const()[name = string("op_1771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_1772_cast_fp16 = add(x = var_69_cast_fp16, y = var_1771_to_fp16)[name = string("op_1772_cast_fp16")];
            fp32 var_1773_epsilon_0 = const()[name = string("op_1773_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_1773_cast_fp16 = rsqrt(epsilon = var_1773_epsilon_0, x = var_1772_cast_fp16)[name = string("op_1773_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_289_cast_fp16 = mul(x = k_49, y = var_1773_cast_fp16)[name = string("x_289_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_51 = mul(x = x_289_cast_fp16, y = encoder_layers_8_self_attn_k_norm_weight)[name = string("k_51")];
            tensor<fp16, [1, 16, 512, 128]> var_1777 = mul(x = q_51, y = cos)[name = string("op_1777")];
            tensor<int32, [2]> var_1778_split_sizes_0 = const()[name = string("op_1778_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1778_axis_0 = const()[name = string("op_1778_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_1778_0, tensor<fp16, [1, 16, 512, 64]> var_1778_1 = split(axis = var_1778_axis_0, split_sizes = var_1778_split_sizes_0, x = q_51)[name = string("op_1778")];
            fp16 const_27_promoted = const()[name = string("const_27_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_1780 = mul(x = var_1778_1, y = const_27_promoted)[name = string("op_1780")];
            bool var_1782_interleave_0 = const()[name = string("op_1782_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1782 = concat(axis = var_18, interleave = var_1782_interleave_0, values = (var_1780, var_1778_0))[name = string("op_1782")];
            tensor<fp16, [1, 16, 512, 128]> var_1783 = mul(x = var_1782, y = sin)[name = string("op_1783")];
            tensor<fp16, [1, 16, 512, 128]> query_17 = add(x = var_1777, y = var_1783)[name = string("query_17")];
            tensor<fp16, [1, 8, 512, 128]> var_1785 = mul(x = k_51, y = cos)[name = string("op_1785")];
            tensor<int32, [2]> var_1786_split_sizes_0 = const()[name = string("op_1786_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1786_axis_0 = const()[name = string("op_1786_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_1786_0, tensor<fp16, [1, 8, 512, 64]> var_1786_1 = split(axis = var_1786_axis_0, split_sizes = var_1786_split_sizes_0, x = k_51)[name = string("op_1786")];
            fp16 const_28_promoted = const()[name = string("const_28_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1788 = mul(x = var_1786_1, y = const_28_promoted)[name = string("op_1788")];
            bool var_1790_interleave_0 = const()[name = string("op_1790_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1790 = concat(axis = var_18, interleave = var_1790_interleave_0, values = (var_1788, var_1786_0))[name = string("op_1790")];
            tensor<fp16, [1, 8, 512, 128]> var_1791 = mul(x = var_1790, y = sin)[name = string("op_1791")];
            tensor<fp16, [1, 8, 512, 128]> x_291 = add(x = var_1785, y = var_1791)[name = string("x_291")];
            tensor<int32, [1]> var_1793_axes_0 = const()[name = string("op_1793_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1793 = expand_dims(axes = var_1793_axes_0, x = x_291)[name = string("op_1793")];
            tensor<int32, [5]> x_293_reps_0 = const()[name = string("x_293_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_293 = tile(reps = x_293_reps_0, x = var_1793)[name = string("x_293")];
            tensor<int32, [4]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_17 = reshape(shape = var_1796, x = x_293)[name = string("key_17")];
            tensor<int32, [1]> var_1798_axes_0 = const()[name = string("op_1798_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_295 = transpose(perm = var_1749, x = var_1748)[name = string("transpose_176")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1798 = expand_dims(axes = var_1798_axes_0, x = x_295)[name = string("op_1798")];
            tensor<int32, [5]> x_297_reps_0 = const()[name = string("x_297_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_297 = tile(reps = x_297_reps_0, x = var_1798)[name = string("x_297")];
            tensor<int32, [4]> var_1801 = const()[name = string("op_1801"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_17 = reshape(shape = var_1801, x = x_297)[name = string("value_17")];
            bool var_1806_transpose_x_1 = const()[name = string("op_1806_transpose_x_1"), val = bool(false)];
            bool var_1806_transpose_y_1 = const()[name = string("op_1806_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_1806_cast_fp16 = matmul(transpose_x = var_1806_transpose_x_1, transpose_y = var_1806_transpose_y_1, x = query_17, y = key_17)[name = string("op_1806_cast_fp16")];
            fp16 var_1807_to_fp16 = const()[name = string("op_1807_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_49_cast_fp16 = mul(x = var_1806_cast_fp16, y = var_1807_to_fp16)[name = string("attn_weights_49_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_51_cast_fp16 = add(x = attn_weights_49_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_51_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_1811_cast_fp16 = softmax(axis = var_18, x = attn_weights_51_cast_fp16)[name = string("op_1811_cast_fp16")];
            bool var_1815_transpose_x_0 = const()[name = string("op_1815_transpose_x_0"), val = bool(false)];
            bool var_1815_transpose_y_0 = const()[name = string("op_1815_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1815_cast_fp16 = matmul(transpose_x = var_1815_transpose_x_0, transpose_y = var_1815_transpose_y_0, x = var_1811_cast_fp16, y = value_17)[name = string("op_1815_cast_fp16")];
            tensor<int32, [4]> var_1817 = const()[name = string("op_1817"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1820 = const()[name = string("op_1820"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_1818 = transpose(perm = var_1817, x = var_1815_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [1, 512, 2048]> attn_out_51 = reshape(shape = var_1820, x = var_1818)[name = string("attn_out_51")];
            tensor<int32, [3]> var_1822 = const()[name = string("op_1822"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_8 = const()[name = string("squeeze_8"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1124925248)))];
            string var_1831_pad_type_0 = const()[name = string("op_1831_pad_type_0"), val = string("valid")];
            int32 var_1831_groups_0 = const()[name = string("op_1831_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_1831_strides_0 = const()[name = string("op_1831_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_1831_pad_0 = const()[name = string("op_1831_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_1831_dilations_0 = const()[name = string("op_1831_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_1823 = transpose(perm = var_1822, x = attn_out_51)[name = string("transpose_174")];
            tensor<fp16, [1, 1024, 512]> var_1831 = conv(dilations = var_1831_dilations_0, groups = var_1831_groups_0, pad = var_1831_pad_0, pad_type = var_1831_pad_type_0, strides = var_1831_strides_0, weight = squeeze_8, x = var_1823)[name = string("op_1831")];
            tensor<int32, [3]> var_1832 = const()[name = string("op_1832"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_53 = transpose(perm = var_1832, x = var_1831)[name = string("transpose_173")];
            tensor<fp16, [1, 512, 1024]> x_299_cast_fp16 = add(x = hidden_states_17_cast_fp16, y = attn_out_53)[name = string("x_299_cast_fp16")];
            fp16 var_6_promoted_35_to_fp16 = const()[name = string("op_6_promoted_35_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1838_cast_fp16 = pow(x = x_299_cast_fp16, y = var_6_promoted_35_to_fp16)[name = string("op_1838_cast_fp16")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = string("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_71_keep_dims_0 = const()[name = string("var_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_71_cast_fp16 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_1838_cast_fp16)[name = string("var_71_cast_fp16")];
            fp16 var_1841_to_fp16 = const()[name = string("op_1841_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1842_cast_fp16 = add(x = var_71_cast_fp16, y = var_1841_to_fp16)[name = string("op_1842_cast_fp16")];
            fp32 var_1843_epsilon_0 = const()[name = string("op_1843_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1843_cast_fp16 = rsqrt(epsilon = var_1843_epsilon_0, x = var_1842_cast_fp16)[name = string("op_1843_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_303_cast_fp16 = mul(x = x_299_cast_fp16, y = var_1843_cast_fp16)[name = string("x_303_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129119616)))];
            tensor<fp16, [1, 512, 1024]> var_1846_cast_fp16 = mul(x = x_303_cast_fp16, y = encoder_layers_8_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_1846_cast_fp16")];
            tensor<int32, [3]> var_1851 = const()[name = string("op_1851"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_85_axes_0 = const()[name = string("input_85_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1852 = transpose(perm = var_1851, x = var_1846_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 1024, 1, 512]> input_85 = expand_dims(axes = input_85_axes_0, x = var_1852)[name = string("input_85")];
            string input_87_pad_type_0 = const()[name = string("input_87_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_87_strides_0 = const()[name = string("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_87_pad_0 = const()[name = string("input_87_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_87_dilations_0 = const()[name = string("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_87_groups_0 = const()[name = string("input_87_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_87 = conv(dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = encoder_layers_8_mlp_gate_proj_weight, x = input_85)[name = string("input_87")];
            string up_17_pad_type_0 = const()[name = string("up_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_17_strides_0 = const()[name = string("up_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_17_pad_0 = const()[name = string("up_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_17_dilations_0 = const()[name = string("up_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_17_groups_0 = const()[name = string("up_17_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_17 = conv(dilations = up_17_dilations_0, groups = up_17_groups_0, pad = up_17_pad_0, pad_type = up_17_pad_type_0, strides = up_17_strides_0, weight = encoder_layers_8_mlp_up_proj_weight, x = input_85)[name = string("up_17")];
            tensor<fp16, [1, 3072, 1, 512]> var_1866 = silu(x = input_87)[name = string("op_1866")];
            tensor<fp16, [1, 3072, 1, 512]> input_89 = mul(x = var_1866, y = up_17)[name = string("input_89")];
            string var_1873_pad_type_0 = const()[name = string("op_1873_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1873_strides_0 = const()[name = string("op_1873_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1873_pad_0 = const()[name = string("op_1873_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1873_dilations_0 = const()[name = string("op_1873_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1873_groups_0 = const()[name = string("op_1873_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1873 = conv(dilations = var_1873_dilations_0, groups = var_1873_groups_0, pad = var_1873_pad_0, pad_type = var_1873_pad_type_0, strides = var_1873_strides_0, weight = encoder_layers_8_mlp_down_proj_weight, x = input_89)[name = string("op_1873")];
            tensor<int32, [1]> var_1874_axes_0 = const()[name = string("op_1874_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1874 = squeeze(axes = var_1874_axes_0, x = var_1873)[name = string("op_1874")];
            tensor<int32, [3]> var_1875 = const()[name = string("op_1875"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_17 = transpose(perm = var_1875, x = var_1874)[name = string("transpose_171")];
            tensor<fp16, [1, 512, 1024]> hidden_states_19_cast_fp16 = add(x = x_299_cast_fp16, y = mlp_out_17)[name = string("hidden_states_19_cast_fp16")];
            fp16 var_6_promoted_36_to_fp16 = const()[name = string("op_6_promoted_36_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_1902_cast_fp16 = pow(x = hidden_states_19_cast_fp16, y = var_6_promoted_36_to_fp16)[name = string("op_1902_cast_fp16")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = string("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_73_keep_dims_0 = const()[name = string("var_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_73_cast_fp16 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_1902_cast_fp16)[name = string("var_73_cast_fp16")];
            fp16 var_1905_to_fp16 = const()[name = string("op_1905_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_1906_cast_fp16 = add(x = var_73_cast_fp16, y = var_1905_to_fp16)[name = string("op_1906_cast_fp16")];
            fp32 var_1907_epsilon_0 = const()[name = string("op_1907_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_1907_cast_fp16 = rsqrt(epsilon = var_1907_epsilon_0, x = var_1906_cast_fp16)[name = string("op_1907_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_309_cast_fp16 = mul(x = hidden_states_19_cast_fp16, y = var_1907_cast_fp16)[name = string("x_309_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_9_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_9_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129121728)))];
            tensor<fp16, [1, 512, 1024]> var_1910_cast_fp16 = mul(x = x_309_cast_fp16, y = encoder_layers_9_input_layernorm_weight_promoted_to_fp16)[name = string("op_1910_cast_fp16")];
            tensor<int32, [3]> var_1915 = const()[name = string("op_1915"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_1916 = transpose(perm = var_1915, x = var_1910_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 1024, 1, 512]> input_91 = expand_dims(axes = input_91_axes_0, x = var_1916)[name = string("input_91")];
            string var_1923_pad_type_0 = const()[name = string("op_1923_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1923_strides_0 = const()[name = string("op_1923_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1923_pad_0 = const()[name = string("op_1923_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1923_dilations_0 = const()[name = string("op_1923_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1923_groups_0 = const()[name = string("op_1923_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_1923 = conv(dilations = var_1923_dilations_0, groups = var_1923_groups_0, pad = var_1923_pad_0, pad_type = var_1923_pad_type_0, strides = var_1923_strides_0, weight = encoder_layers_9_self_attn_q_proj_weight, x = input_91)[name = string("op_1923")];
            tensor<int32, [4]> var_1924 = const()[name = string("op_1924"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_1925 = reshape(shape = var_1924, x = var_1923)[name = string("op_1925")];
            tensor<int32, [4]> var_1926 = const()[name = string("op_1926"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1933_pad_type_0 = const()[name = string("op_1933_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1933_strides_0 = const()[name = string("op_1933_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1933_pad_0 = const()[name = string("op_1933_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1933_dilations_0 = const()[name = string("op_1933_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1933_groups_0 = const()[name = string("op_1933_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1933 = conv(dilations = var_1933_dilations_0, groups = var_1933_groups_0, pad = var_1933_pad_0, pad_type = var_1933_pad_type_0, strides = var_1933_strides_0, weight = encoder_layers_9_self_attn_k_proj_weight, x = input_91)[name = string("op_1933")];
            tensor<int32, [4]> var_1934 = const()[name = string("op_1934"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1935 = reshape(shape = var_1934, x = var_1933)[name = string("op_1935")];
            tensor<int32, [4]> var_1936 = const()[name = string("op_1936"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_1943_pad_type_0 = const()[name = string("op_1943_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_1943_strides_0 = const()[name = string("op_1943_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_1943_pad_0 = const()[name = string("op_1943_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_1943_dilations_0 = const()[name = string("op_1943_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_1943_groups_0 = const()[name = string("op_1943_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_1943 = conv(dilations = var_1943_dilations_0, groups = var_1943_groups_0, pad = var_1943_pad_0, pad_type = var_1943_pad_type_0, strides = var_1943_strides_0, weight = encoder_layers_9_self_attn_v_proj_weight, x = input_91)[name = string("op_1943")];
            tensor<int32, [4]> var_1944 = const()[name = string("op_1944"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_1945 = reshape(shape = var_1944, x = var_1943)[name = string("op_1945")];
            tensor<int32, [4]> var_1946 = const()[name = string("op_1946"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_37_to_fp16 = const()[name = string("op_6_promoted_37_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_55 = transpose(perm = var_1926, x = var_1925)[name = string("transpose_169")];
            tensor<fp16, [1, 16, 512, 128]> var_1952_cast_fp16 = pow(x = q_55, y = var_6_promoted_37_to_fp16)[name = string("op_1952_cast_fp16")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = string("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_75_keep_dims_0 = const()[name = string("var_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_75_cast_fp16 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_1952_cast_fp16)[name = string("var_75_cast_fp16")];
            fp16 var_1955_to_fp16 = const()[name = string("op_1955_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_1956_cast_fp16 = add(x = var_75_cast_fp16, y = var_1955_to_fp16)[name = string("op_1956_cast_fp16")];
            fp32 var_1957_epsilon_0 = const()[name = string("op_1957_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_1957_cast_fp16 = rsqrt(epsilon = var_1957_epsilon_0, x = var_1956_cast_fp16)[name = string("op_1957_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_317_cast_fp16 = mul(x = q_55, y = var_1957_cast_fp16)[name = string("x_317_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_57 = mul(x = x_317_cast_fp16, y = encoder_layers_9_self_attn_q_norm_weight)[name = string("q_57")];
            fp16 var_6_promoted_38_to_fp16 = const()[name = string("op_6_promoted_38_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_55 = transpose(perm = var_1936, x = var_1935)[name = string("transpose_168")];
            tensor<fp16, [1, 8, 512, 128]> var_1965_cast_fp16 = pow(x = k_55, y = var_6_promoted_38_to_fp16)[name = string("op_1965_cast_fp16")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = string("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_77_keep_dims_0 = const()[name = string("var_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_77_cast_fp16 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_1965_cast_fp16)[name = string("var_77_cast_fp16")];
            fp16 var_1968_to_fp16 = const()[name = string("op_1968_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_1969_cast_fp16 = add(x = var_77_cast_fp16, y = var_1968_to_fp16)[name = string("op_1969_cast_fp16")];
            fp32 var_1970_epsilon_0 = const()[name = string("op_1970_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_1970_cast_fp16 = rsqrt(epsilon = var_1970_epsilon_0, x = var_1969_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_323_cast_fp16 = mul(x = k_55, y = var_1970_cast_fp16)[name = string("x_323_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_57 = mul(x = x_323_cast_fp16, y = encoder_layers_9_self_attn_k_norm_weight)[name = string("k_57")];
            tensor<fp16, [1, 16, 512, 128]> var_1974 = mul(x = q_57, y = cos)[name = string("op_1974")];
            tensor<int32, [2]> var_1975_split_sizes_0 = const()[name = string("op_1975_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1975_axis_0 = const()[name = string("op_1975_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_1975_0, tensor<fp16, [1, 16, 512, 64]> var_1975_1 = split(axis = var_1975_axis_0, split_sizes = var_1975_split_sizes_0, x = q_57)[name = string("op_1975")];
            fp16 const_30_promoted = const()[name = string("const_30_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_1977 = mul(x = var_1975_1, y = const_30_promoted)[name = string("op_1977")];
            bool var_1979_interleave_0 = const()[name = string("op_1979_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_1979 = concat(axis = var_18, interleave = var_1979_interleave_0, values = (var_1977, var_1975_0))[name = string("op_1979")];
            tensor<fp16, [1, 16, 512, 128]> var_1980 = mul(x = var_1979, y = sin)[name = string("op_1980")];
            tensor<fp16, [1, 16, 512, 128]> query_19 = add(x = var_1974, y = var_1980)[name = string("query_19")];
            tensor<fp16, [1, 8, 512, 128]> var_1982 = mul(x = k_57, y = cos)[name = string("op_1982")];
            tensor<int32, [2]> var_1983_split_sizes_0 = const()[name = string("op_1983_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_1983_axis_0 = const()[name = string("op_1983_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_1983_0, tensor<fp16, [1, 8, 512, 64]> var_1983_1 = split(axis = var_1983_axis_0, split_sizes = var_1983_split_sizes_0, x = k_57)[name = string("op_1983")];
            fp16 const_31_promoted = const()[name = string("const_31_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_1985 = mul(x = var_1983_1, y = const_31_promoted)[name = string("op_1985")];
            bool var_1987_interleave_0 = const()[name = string("op_1987_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_1987 = concat(axis = var_18, interleave = var_1987_interleave_0, values = (var_1985, var_1983_0))[name = string("op_1987")];
            tensor<fp16, [1, 8, 512, 128]> var_1988 = mul(x = var_1987, y = sin)[name = string("op_1988")];
            tensor<fp16, [1, 8, 512, 128]> x_325 = add(x = var_1982, y = var_1988)[name = string("x_325")];
            tensor<int32, [1]> var_1990_axes_0 = const()[name = string("op_1990_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1990 = expand_dims(axes = var_1990_axes_0, x = x_325)[name = string("op_1990")];
            tensor<int32, [5]> x_327_reps_0 = const()[name = string("x_327_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_327 = tile(reps = x_327_reps_0, x = var_1990)[name = string("x_327")];
            tensor<int32, [4]> var_1993 = const()[name = string("op_1993"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_19 = reshape(shape = var_1993, x = x_327)[name = string("key_19")];
            tensor<int32, [1]> var_1995_axes_0 = const()[name = string("op_1995_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_329 = transpose(perm = var_1946, x = var_1945)[name = string("transpose_167")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_1995 = expand_dims(axes = var_1995_axes_0, x = x_329)[name = string("op_1995")];
            tensor<int32, [5]> x_331_reps_0 = const()[name = string("x_331_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_331 = tile(reps = x_331_reps_0, x = var_1995)[name = string("x_331")];
            tensor<int32, [4]> var_1998 = const()[name = string("op_1998"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_19 = reshape(shape = var_1998, x = x_331)[name = string("value_19")];
            bool var_2003_transpose_x_1 = const()[name = string("op_2003_transpose_x_1"), val = bool(false)];
            bool var_2003_transpose_y_1 = const()[name = string("op_2003_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2003_cast_fp16 = matmul(transpose_x = var_2003_transpose_x_1, transpose_y = var_2003_transpose_y_1, x = query_19, y = key_19)[name = string("op_2003_cast_fp16")];
            fp16 var_2004_to_fp16 = const()[name = string("op_2004_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_55_cast_fp16 = mul(x = var_2003_cast_fp16, y = var_2004_to_fp16)[name = string("attn_weights_55_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_57_cast_fp16 = add(x = attn_weights_55_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_57_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2008_cast_fp16 = softmax(axis = var_18, x = attn_weights_57_cast_fp16)[name = string("op_2008_cast_fp16")];
            bool var_2012_transpose_x_0 = const()[name = string("op_2012_transpose_x_0"), val = bool(false)];
            bool var_2012_transpose_y_0 = const()[name = string("op_2012_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2012_cast_fp16 = matmul(transpose_x = var_2012_transpose_x_0, transpose_y = var_2012_transpose_y_0, x = var_2008_cast_fp16, y = value_19)[name = string("op_2012_cast_fp16")];
            tensor<int32, [4]> var_2014 = const()[name = string("op_2014"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2017 = const()[name = string("op_2017"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_2015 = transpose(perm = var_2014, x = var_2012_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [1, 512, 2048]> attn_out_57 = reshape(shape = var_2017, x = var_2015)[name = string("attn_out_57")];
            tensor<int32, [3]> var_2019 = const()[name = string("op_2019"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_9 = const()[name = string("squeeze_9"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1129123840)))];
            string var_2028_pad_type_0 = const()[name = string("op_2028_pad_type_0"), val = string("valid")];
            int32 var_2028_groups_0 = const()[name = string("op_2028_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2028_strides_0 = const()[name = string("op_2028_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2028_pad_0 = const()[name = string("op_2028_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2028_dilations_0 = const()[name = string("op_2028_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_2020 = transpose(perm = var_2019, x = attn_out_57)[name = string("transpose_165")];
            tensor<fp16, [1, 1024, 512]> var_2028 = conv(dilations = var_2028_dilations_0, groups = var_2028_groups_0, pad = var_2028_pad_0, pad_type = var_2028_pad_type_0, strides = var_2028_strides_0, weight = squeeze_9, x = var_2020)[name = string("op_2028")];
            tensor<int32, [3]> var_2029 = const()[name = string("op_2029"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_59 = transpose(perm = var_2029, x = var_2028)[name = string("transpose_164")];
            tensor<fp16, [1, 512, 1024]> x_333_cast_fp16 = add(x = hidden_states_19_cast_fp16, y = attn_out_59)[name = string("x_333_cast_fp16")];
            fp16 var_6_promoted_39_to_fp16 = const()[name = string("op_6_promoted_39_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2035_cast_fp16 = pow(x = x_333_cast_fp16, y = var_6_promoted_39_to_fp16)[name = string("op_2035_cast_fp16")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = string("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_79_keep_dims_0 = const()[name = string("var_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_79_cast_fp16 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_2035_cast_fp16)[name = string("var_79_cast_fp16")];
            fp16 var_2038_to_fp16 = const()[name = string("op_2038_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2039_cast_fp16 = add(x = var_79_cast_fp16, y = var_2038_to_fp16)[name = string("op_2039_cast_fp16")];
            fp32 var_2040_epsilon_0 = const()[name = string("op_2040_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2040_cast_fp16 = rsqrt(epsilon = var_2040_epsilon_0, x = var_2039_cast_fp16)[name = string("op_2040_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_337_cast_fp16 = mul(x = x_333_cast_fp16, y = var_2040_cast_fp16)[name = string("x_337_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133318208)))];
            tensor<fp16, [1, 512, 1024]> var_2043_cast_fp16 = mul(x = x_337_cast_fp16, y = encoder_layers_9_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [3]> var_2048 = const()[name = string("op_2048"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_95_axes_0 = const()[name = string("input_95_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2049 = transpose(perm = var_2048, x = var_2043_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 1024, 1, 512]> input_95 = expand_dims(axes = input_95_axes_0, x = var_2049)[name = string("input_95")];
            string input_97_pad_type_0 = const()[name = string("input_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_97_strides_0 = const()[name = string("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_97_pad_0 = const()[name = string("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_97_dilations_0 = const()[name = string("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_97_groups_0 = const()[name = string("input_97_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_97 = conv(dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = encoder_layers_9_mlp_gate_proj_weight, x = input_95)[name = string("input_97")];
            string up_19_pad_type_0 = const()[name = string("up_19_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_19_strides_0 = const()[name = string("up_19_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_19_pad_0 = const()[name = string("up_19_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_19_dilations_0 = const()[name = string("up_19_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_19_groups_0 = const()[name = string("up_19_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_19 = conv(dilations = up_19_dilations_0, groups = up_19_groups_0, pad = up_19_pad_0, pad_type = up_19_pad_type_0, strides = up_19_strides_0, weight = encoder_layers_9_mlp_up_proj_weight, x = input_95)[name = string("up_19")];
            tensor<fp16, [1, 3072, 1, 512]> var_2063 = silu(x = input_97)[name = string("op_2063")];
            tensor<fp16, [1, 3072, 1, 512]> input_99 = mul(x = var_2063, y = up_19)[name = string("input_99")];
            string var_2070_pad_type_0 = const()[name = string("op_2070_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2070_strides_0 = const()[name = string("op_2070_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2070_pad_0 = const()[name = string("op_2070_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2070_dilations_0 = const()[name = string("op_2070_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2070_groups_0 = const()[name = string("op_2070_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2070 = conv(dilations = var_2070_dilations_0, groups = var_2070_groups_0, pad = var_2070_pad_0, pad_type = var_2070_pad_type_0, strides = var_2070_strides_0, weight = encoder_layers_9_mlp_down_proj_weight, x = input_99)[name = string("op_2070")];
            tensor<int32, [1]> var_2071_axes_0 = const()[name = string("op_2071_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2071 = squeeze(axes = var_2071_axes_0, x = var_2070)[name = string("op_2071")];
            tensor<int32, [3]> var_2072 = const()[name = string("op_2072"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_19 = transpose(perm = var_2072, x = var_2071)[name = string("transpose_162")];
            tensor<fp16, [1, 512, 1024]> hidden_states_21_cast_fp16 = add(x = x_333_cast_fp16, y = mlp_out_19)[name = string("hidden_states_21_cast_fp16")];
            fp16 var_6_promoted_40_to_fp16 = const()[name = string("op_6_promoted_40_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2099_cast_fp16 = pow(x = hidden_states_21_cast_fp16, y = var_6_promoted_40_to_fp16)[name = string("op_2099_cast_fp16")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = string("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_81_keep_dims_0 = const()[name = string("var_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_81_cast_fp16 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_2099_cast_fp16)[name = string("var_81_cast_fp16")];
            fp16 var_2102_to_fp16 = const()[name = string("op_2102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2103_cast_fp16 = add(x = var_81_cast_fp16, y = var_2102_to_fp16)[name = string("op_2103_cast_fp16")];
            fp32 var_2104_epsilon_0 = const()[name = string("op_2104_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2104_cast_fp16 = rsqrt(epsilon = var_2104_epsilon_0, x = var_2103_cast_fp16)[name = string("op_2104_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_343_cast_fp16 = mul(x = hidden_states_21_cast_fp16, y = var_2104_cast_fp16)[name = string("x_343_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_10_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_10_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133320320)))];
            tensor<fp16, [1, 512, 1024]> var_2107_cast_fp16 = mul(x = x_343_cast_fp16, y = encoder_layers_10_input_layernorm_weight_promoted_to_fp16)[name = string("op_2107_cast_fp16")];
            tensor<int32, [3]> var_2112 = const()[name = string("op_2112"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2113 = transpose(perm = var_2112, x = var_2107_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 1024, 1, 512]> input_101 = expand_dims(axes = input_101_axes_0, x = var_2113)[name = string("input_101")];
            string var_2120_pad_type_0 = const()[name = string("op_2120_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2120_strides_0 = const()[name = string("op_2120_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2120_pad_0 = const()[name = string("op_2120_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2120_dilations_0 = const()[name = string("op_2120_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2120_groups_0 = const()[name = string("op_2120_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_2120 = conv(dilations = var_2120_dilations_0, groups = var_2120_groups_0, pad = var_2120_pad_0, pad_type = var_2120_pad_type_0, strides = var_2120_strides_0, weight = encoder_layers_10_self_attn_q_proj_weight, x = input_101)[name = string("op_2120")];
            tensor<int32, [4]> var_2121 = const()[name = string("op_2121"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_2122 = reshape(shape = var_2121, x = var_2120)[name = string("op_2122")];
            tensor<int32, [4]> var_2123 = const()[name = string("op_2123"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2130_pad_type_0 = const()[name = string("op_2130_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2130_strides_0 = const()[name = string("op_2130_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2130_pad_0 = const()[name = string("op_2130_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2130_dilations_0 = const()[name = string("op_2130_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2130_groups_0 = const()[name = string("op_2130_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2130 = conv(dilations = var_2130_dilations_0, groups = var_2130_groups_0, pad = var_2130_pad_0, pad_type = var_2130_pad_type_0, strides = var_2130_strides_0, weight = encoder_layers_10_self_attn_k_proj_weight, x = input_101)[name = string("op_2130")];
            tensor<int32, [4]> var_2131 = const()[name = string("op_2131"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2132 = reshape(shape = var_2131, x = var_2130)[name = string("op_2132")];
            tensor<int32, [4]> var_2133 = const()[name = string("op_2133"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2140_pad_type_0 = const()[name = string("op_2140_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2140_strides_0 = const()[name = string("op_2140_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2140_pad_0 = const()[name = string("op_2140_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2140_dilations_0 = const()[name = string("op_2140_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2140_groups_0 = const()[name = string("op_2140_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2140 = conv(dilations = var_2140_dilations_0, groups = var_2140_groups_0, pad = var_2140_pad_0, pad_type = var_2140_pad_type_0, strides = var_2140_strides_0, weight = encoder_layers_10_self_attn_v_proj_weight, x = input_101)[name = string("op_2140")];
            tensor<int32, [4]> var_2141 = const()[name = string("op_2141"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2142 = reshape(shape = var_2141, x = var_2140)[name = string("op_2142")];
            tensor<int32, [4]> var_2143 = const()[name = string("op_2143"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_41_to_fp16 = const()[name = string("op_6_promoted_41_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_61 = transpose(perm = var_2123, x = var_2122)[name = string("transpose_160")];
            tensor<fp16, [1, 16, 512, 128]> var_2149_cast_fp16 = pow(x = q_61, y = var_6_promoted_41_to_fp16)[name = string("op_2149_cast_fp16")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = string("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_83_keep_dims_0 = const()[name = string("var_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_83_cast_fp16_0 = reduce_mean(axes = var_83_axes_0, keep_dims = var_83_keep_dims_0, x = var_2149_cast_fp16)[name = string("var_83_cast_fp16")];
            fp16 var_2152_to_fp16 = const()[name = string("op_2152_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_2153_cast_fp16 = add(x = var_83_cast_fp16_0, y = var_2152_to_fp16)[name = string("op_2153_cast_fp16")];
            fp32 var_2154_epsilon_0 = const()[name = string("op_2154_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_2154_cast_fp16 = rsqrt(epsilon = var_2154_epsilon_0, x = var_2153_cast_fp16)[name = string("op_2154_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_351_cast_fp16 = mul(x = q_61, y = var_2154_cast_fp16)[name = string("x_351_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_63 = mul(x = x_351_cast_fp16, y = encoder_layers_10_self_attn_q_norm_weight)[name = string("q_63")];
            fp16 var_6_promoted_42_to_fp16 = const()[name = string("op_6_promoted_42_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_61 = transpose(perm = var_2133, x = var_2132)[name = string("transpose_159")];
            tensor<fp16, [1, 8, 512, 128]> var_2162_cast_fp16 = pow(x = k_61, y = var_6_promoted_42_to_fp16)[name = string("op_2162_cast_fp16")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = string("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_85_keep_dims_0 = const()[name = string("var_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_85_cast_fp16 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_2162_cast_fp16)[name = string("var_85_cast_fp16")];
            fp16 var_2165_to_fp16 = const()[name = string("op_2165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_2166_cast_fp16 = add(x = var_85_cast_fp16, y = var_2165_to_fp16)[name = string("op_2166_cast_fp16")];
            fp32 var_2167_epsilon_0 = const()[name = string("op_2167_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_2167_cast_fp16 = rsqrt(epsilon = var_2167_epsilon_0, x = var_2166_cast_fp16)[name = string("op_2167_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_357_cast_fp16 = mul(x = k_61, y = var_2167_cast_fp16)[name = string("x_357_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_63 = mul(x = x_357_cast_fp16, y = encoder_layers_10_self_attn_k_norm_weight)[name = string("k_63")];
            tensor<fp16, [1, 16, 512, 128]> var_2171 = mul(x = q_63, y = cos)[name = string("op_2171")];
            tensor<int32, [2]> var_2172_split_sizes_0 = const()[name = string("op_2172_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2172_axis_0 = const()[name = string("op_2172_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_2172_0, tensor<fp16, [1, 16, 512, 64]> var_2172_1 = split(axis = var_2172_axis_0, split_sizes = var_2172_split_sizes_0, x = q_63)[name = string("op_2172")];
            fp16 const_33_promoted = const()[name = string("const_33_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_2174 = mul(x = var_2172_1, y = const_33_promoted)[name = string("op_2174")];
            bool var_2176_interleave_0 = const()[name = string("op_2176_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2176 = concat(axis = var_18, interleave = var_2176_interleave_0, values = (var_2174, var_2172_0))[name = string("op_2176")];
            tensor<fp16, [1, 16, 512, 128]> var_2177 = mul(x = var_2176, y = sin)[name = string("op_2177")];
            tensor<fp16, [1, 16, 512, 128]> query_21 = add(x = var_2171, y = var_2177)[name = string("query_21")];
            tensor<fp16, [1, 8, 512, 128]> var_2179 = mul(x = k_63, y = cos)[name = string("op_2179")];
            tensor<int32, [2]> var_2180_split_sizes_0 = const()[name = string("op_2180_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2180_axis_0 = const()[name = string("op_2180_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_2180_0, tensor<fp16, [1, 8, 512, 64]> var_2180_1 = split(axis = var_2180_axis_0, split_sizes = var_2180_split_sizes_0, x = k_63)[name = string("op_2180")];
            fp16 const_34_promoted = const()[name = string("const_34_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_2182 = mul(x = var_2180_1, y = const_34_promoted)[name = string("op_2182")];
            bool var_2184_interleave_0 = const()[name = string("op_2184_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_2184 = concat(axis = var_18, interleave = var_2184_interleave_0, values = (var_2182, var_2180_0))[name = string("op_2184")];
            tensor<fp16, [1, 8, 512, 128]> var_2185 = mul(x = var_2184, y = sin)[name = string("op_2185")];
            tensor<fp16, [1, 8, 512, 128]> x_359 = add(x = var_2179, y = var_2185)[name = string("x_359")];
            tensor<int32, [1]> var_2187_axes_0 = const()[name = string("op_2187_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2187 = expand_dims(axes = var_2187_axes_0, x = x_359)[name = string("op_2187")];
            tensor<int32, [5]> x_361_reps_0 = const()[name = string("x_361_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_361 = tile(reps = x_361_reps_0, x = var_2187)[name = string("x_361")];
            tensor<int32, [4]> var_2190 = const()[name = string("op_2190"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_21 = reshape(shape = var_2190, x = x_361)[name = string("key_21")];
            tensor<int32, [1]> var_2192_axes_0 = const()[name = string("op_2192_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_363 = transpose(perm = var_2143, x = var_2142)[name = string("transpose_158")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2192 = expand_dims(axes = var_2192_axes_0, x = x_363)[name = string("op_2192")];
            tensor<int32, [5]> x_365_reps_0 = const()[name = string("x_365_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_365 = tile(reps = x_365_reps_0, x = var_2192)[name = string("x_365")];
            tensor<int32, [4]> var_2195 = const()[name = string("op_2195"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_21 = reshape(shape = var_2195, x = x_365)[name = string("value_21")];
            bool var_2200_transpose_x_1 = const()[name = string("op_2200_transpose_x_1"), val = bool(false)];
            bool var_2200_transpose_y_1 = const()[name = string("op_2200_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2200_cast_fp16 = matmul(transpose_x = var_2200_transpose_x_1, transpose_y = var_2200_transpose_y_1, x = query_21, y = key_21)[name = string("op_2200_cast_fp16")];
            fp16 var_2201_to_fp16 = const()[name = string("op_2201_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_61_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = string("attn_weights_61_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_63_cast_fp16 = add(x = attn_weights_61_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_63_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2205_cast_fp16 = softmax(axis = var_18, x = attn_weights_63_cast_fp16)[name = string("op_2205_cast_fp16")];
            bool var_2209_transpose_x_0 = const()[name = string("op_2209_transpose_x_0"), val = bool(false)];
            bool var_2209_transpose_y_0 = const()[name = string("op_2209_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2209_cast_fp16 = matmul(transpose_x = var_2209_transpose_x_0, transpose_y = var_2209_transpose_y_0, x = var_2205_cast_fp16, y = value_21)[name = string("op_2209_cast_fp16")];
            tensor<int32, [4]> var_2211 = const()[name = string("op_2211"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2214 = const()[name = string("op_2214"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_2212 = transpose(perm = var_2211, x = var_2209_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 512, 2048]> attn_out_63 = reshape(shape = var_2214, x = var_2212)[name = string("attn_out_63")];
            tensor<int32, [3]> var_2216 = const()[name = string("op_2216"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_10 = const()[name = string("squeeze_10"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1133322432)))];
            string var_2225_pad_type_0 = const()[name = string("op_2225_pad_type_0"), val = string("valid")];
            int32 var_2225_groups_0 = const()[name = string("op_2225_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2225_strides_0 = const()[name = string("op_2225_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2225_pad_0 = const()[name = string("op_2225_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2225_dilations_0 = const()[name = string("op_2225_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_2217 = transpose(perm = var_2216, x = attn_out_63)[name = string("transpose_156")];
            tensor<fp16, [1, 1024, 512]> var_2225 = conv(dilations = var_2225_dilations_0, groups = var_2225_groups_0, pad = var_2225_pad_0, pad_type = var_2225_pad_type_0, strides = var_2225_strides_0, weight = squeeze_10, x = var_2217)[name = string("op_2225")];
            tensor<int32, [3]> var_2226 = const()[name = string("op_2226"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_65 = transpose(perm = var_2226, x = var_2225)[name = string("transpose_155")];
            tensor<fp16, [1, 512, 1024]> x_367_cast_fp16 = add(x = hidden_states_21_cast_fp16, y = attn_out_65)[name = string("x_367_cast_fp16")];
            fp16 var_6_promoted_43_to_fp16 = const()[name = string("op_6_promoted_43_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2232_cast_fp16 = pow(x = x_367_cast_fp16, y = var_6_promoted_43_to_fp16)[name = string("op_2232_cast_fp16")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = string("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_87_keep_dims_0 = const()[name = string("var_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_87_cast_fp16 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_2232_cast_fp16)[name = string("var_87_cast_fp16")];
            fp16 var_2235_to_fp16 = const()[name = string("op_2235_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2236_cast_fp16 = add(x = var_87_cast_fp16, y = var_2235_to_fp16)[name = string("op_2236_cast_fp16")];
            fp32 var_2237_epsilon_0 = const()[name = string("op_2237_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2237_cast_fp16 = rsqrt(epsilon = var_2237_epsilon_0, x = var_2236_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_371_cast_fp16 = mul(x = x_367_cast_fp16, y = var_2237_cast_fp16)[name = string("x_371_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137516800)))];
            tensor<fp16, [1, 512, 1024]> var_2240_cast_fp16 = mul(x = x_371_cast_fp16, y = encoder_layers_10_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2240_cast_fp16")];
            tensor<int32, [3]> var_2245 = const()[name = string("op_2245"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_105_axes_0 = const()[name = string("input_105_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2246 = transpose(perm = var_2245, x = var_2240_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 1024, 1, 512]> input_105 = expand_dims(axes = input_105_axes_0, x = var_2246)[name = string("input_105")];
            string input_107_pad_type_0 = const()[name = string("input_107_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_107_strides_0 = const()[name = string("input_107_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_107_pad_0 = const()[name = string("input_107_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_107_dilations_0 = const()[name = string("input_107_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_107_groups_0 = const()[name = string("input_107_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_107 = conv(dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = encoder_layers_10_mlp_gate_proj_weight, x = input_105)[name = string("input_107")];
            string up_21_pad_type_0 = const()[name = string("up_21_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_21_strides_0 = const()[name = string("up_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_21_pad_0 = const()[name = string("up_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_21_dilations_0 = const()[name = string("up_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_21_groups_0 = const()[name = string("up_21_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_21 = conv(dilations = up_21_dilations_0, groups = up_21_groups_0, pad = up_21_pad_0, pad_type = up_21_pad_type_0, strides = up_21_strides_0, weight = encoder_layers_10_mlp_up_proj_weight, x = input_105)[name = string("up_21")];
            tensor<fp16, [1, 3072, 1, 512]> var_2260 = silu(x = input_107)[name = string("op_2260")];
            tensor<fp16, [1, 3072, 1, 512]> input_109 = mul(x = var_2260, y = up_21)[name = string("input_109")];
            string var_2267_pad_type_0 = const()[name = string("op_2267_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2267_strides_0 = const()[name = string("op_2267_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2267_pad_0 = const()[name = string("op_2267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2267_dilations_0 = const()[name = string("op_2267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2267_groups_0 = const()[name = string("op_2267_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2267 = conv(dilations = var_2267_dilations_0, groups = var_2267_groups_0, pad = var_2267_pad_0, pad_type = var_2267_pad_type_0, strides = var_2267_strides_0, weight = encoder_layers_10_mlp_down_proj_weight, x = input_109)[name = string("op_2267")];
            tensor<int32, [1]> var_2268_axes_0 = const()[name = string("op_2268_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2268 = squeeze(axes = var_2268_axes_0, x = var_2267)[name = string("op_2268")];
            tensor<int32, [3]> var_2269 = const()[name = string("op_2269"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_21 = transpose(perm = var_2269, x = var_2268)[name = string("transpose_153")];
            tensor<fp16, [1, 512, 1024]> hidden_states_23_cast_fp16 = add(x = x_367_cast_fp16, y = mlp_out_21)[name = string("hidden_states_23_cast_fp16")];
            fp16 var_6_promoted_44_to_fp16 = const()[name = string("op_6_promoted_44_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2296_cast_fp16 = pow(x = hidden_states_23_cast_fp16, y = var_6_promoted_44_to_fp16)[name = string("op_2296_cast_fp16")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = string("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_89_keep_dims_0 = const()[name = string("var_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_89_cast_fp16 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_2296_cast_fp16)[name = string("var_89_cast_fp16")];
            fp16 var_2299_to_fp16 = const()[name = string("op_2299_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2300_cast_fp16 = add(x = var_89_cast_fp16, y = var_2299_to_fp16)[name = string("op_2300_cast_fp16")];
            fp32 var_2301_epsilon_0 = const()[name = string("op_2301_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2301_cast_fp16 = rsqrt(epsilon = var_2301_epsilon_0, x = var_2300_cast_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_377_cast_fp16 = mul(x = hidden_states_23_cast_fp16, y = var_2301_cast_fp16)[name = string("x_377_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_11_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_11_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137518912)))];
            tensor<fp16, [1, 512, 1024]> var_2304_cast_fp16 = mul(x = x_377_cast_fp16, y = encoder_layers_11_input_layernorm_weight_promoted_to_fp16)[name = string("op_2304_cast_fp16")];
            tensor<int32, [3]> var_2309 = const()[name = string("op_2309"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2310 = transpose(perm = var_2309, x = var_2304_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 1024, 1, 512]> input_111 = expand_dims(axes = input_111_axes_0, x = var_2310)[name = string("input_111")];
            string var_2317_pad_type_0 = const()[name = string("op_2317_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2317_strides_0 = const()[name = string("op_2317_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2317_pad_0 = const()[name = string("op_2317_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2317_dilations_0 = const()[name = string("op_2317_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2317_groups_0 = const()[name = string("op_2317_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_2317 = conv(dilations = var_2317_dilations_0, groups = var_2317_groups_0, pad = var_2317_pad_0, pad_type = var_2317_pad_type_0, strides = var_2317_strides_0, weight = encoder_layers_11_self_attn_q_proj_weight, x = input_111)[name = string("op_2317")];
            tensor<int32, [4]> var_2318 = const()[name = string("op_2318"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_2319 = reshape(shape = var_2318, x = var_2317)[name = string("op_2319")];
            tensor<int32, [4]> var_2320 = const()[name = string("op_2320"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2327_pad_type_0 = const()[name = string("op_2327_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2327_strides_0 = const()[name = string("op_2327_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2327_pad_0 = const()[name = string("op_2327_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2327_dilations_0 = const()[name = string("op_2327_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2327_groups_0 = const()[name = string("op_2327_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2327 = conv(dilations = var_2327_dilations_0, groups = var_2327_groups_0, pad = var_2327_pad_0, pad_type = var_2327_pad_type_0, strides = var_2327_strides_0, weight = encoder_layers_11_self_attn_k_proj_weight, x = input_111)[name = string("op_2327")];
            tensor<int32, [4]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2329 = reshape(shape = var_2328, x = var_2327)[name = string("op_2329")];
            tensor<int32, [4]> var_2330 = const()[name = string("op_2330"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2337_pad_type_0 = const()[name = string("op_2337_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2337_strides_0 = const()[name = string("op_2337_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2337_pad_0 = const()[name = string("op_2337_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2337_dilations_0 = const()[name = string("op_2337_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2337_groups_0 = const()[name = string("op_2337_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2337 = conv(dilations = var_2337_dilations_0, groups = var_2337_groups_0, pad = var_2337_pad_0, pad_type = var_2337_pad_type_0, strides = var_2337_strides_0, weight = encoder_layers_11_self_attn_v_proj_weight, x = input_111)[name = string("op_2337")];
            tensor<int32, [4]> var_2338 = const()[name = string("op_2338"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2339 = reshape(shape = var_2338, x = var_2337)[name = string("op_2339")];
            tensor<int32, [4]> var_2340 = const()[name = string("op_2340"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_45_to_fp16 = const()[name = string("op_6_promoted_45_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_67 = transpose(perm = var_2320, x = var_2319)[name = string("transpose_151")];
            tensor<fp16, [1, 16, 512, 128]> var_2346_cast_fp16 = pow(x = q_67, y = var_6_promoted_45_to_fp16)[name = string("op_2346_cast_fp16")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = string("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_91_keep_dims_0 = const()[name = string("var_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_91_cast_fp16 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_2346_cast_fp16)[name = string("var_91_cast_fp16")];
            fp16 var_2349_to_fp16 = const()[name = string("op_2349_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_2350_cast_fp16 = add(x = var_91_cast_fp16, y = var_2349_to_fp16)[name = string("op_2350_cast_fp16")];
            fp32 var_2351_epsilon_0 = const()[name = string("op_2351_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_2351_cast_fp16 = rsqrt(epsilon = var_2351_epsilon_0, x = var_2350_cast_fp16)[name = string("op_2351_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_385_cast_fp16 = mul(x = q_67, y = var_2351_cast_fp16)[name = string("x_385_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_69 = mul(x = x_385_cast_fp16, y = encoder_layers_11_self_attn_q_norm_weight)[name = string("q_69")];
            fp16 var_6_promoted_46_to_fp16 = const()[name = string("op_6_promoted_46_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_67 = transpose(perm = var_2330, x = var_2329)[name = string("transpose_150")];
            tensor<fp16, [1, 8, 512, 128]> var_2359_cast_fp16 = pow(x = k_67, y = var_6_promoted_46_to_fp16)[name = string("op_2359_cast_fp16")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = string("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_93_keep_dims_0 = const()[name = string("var_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_93_cast_fp16 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_2359_cast_fp16)[name = string("var_93_cast_fp16")];
            fp16 var_2362_to_fp16 = const()[name = string("op_2362_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_2363_cast_fp16 = add(x = var_93_cast_fp16, y = var_2362_to_fp16)[name = string("op_2363_cast_fp16")];
            fp32 var_2364_epsilon_0 = const()[name = string("op_2364_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_2364_cast_fp16 = rsqrt(epsilon = var_2364_epsilon_0, x = var_2363_cast_fp16)[name = string("op_2364_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_391_cast_fp16 = mul(x = k_67, y = var_2364_cast_fp16)[name = string("x_391_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_69 = mul(x = x_391_cast_fp16, y = encoder_layers_11_self_attn_k_norm_weight)[name = string("k_69")];
            tensor<fp16, [1, 16, 512, 128]> var_2368 = mul(x = q_69, y = cos)[name = string("op_2368")];
            tensor<int32, [2]> var_2369_split_sizes_0 = const()[name = string("op_2369_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2369_axis_0 = const()[name = string("op_2369_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_2369_0, tensor<fp16, [1, 16, 512, 64]> var_2369_1 = split(axis = var_2369_axis_0, split_sizes = var_2369_split_sizes_0, x = q_69)[name = string("op_2369")];
            fp16 const_36_promoted = const()[name = string("const_36_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_2371 = mul(x = var_2369_1, y = const_36_promoted)[name = string("op_2371")];
            bool var_2373_interleave_0 = const()[name = string("op_2373_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2373 = concat(axis = var_18, interleave = var_2373_interleave_0, values = (var_2371, var_2369_0))[name = string("op_2373")];
            tensor<fp16, [1, 16, 512, 128]> var_2374 = mul(x = var_2373, y = sin)[name = string("op_2374")];
            tensor<fp16, [1, 16, 512, 128]> query_23 = add(x = var_2368, y = var_2374)[name = string("query_23")];
            tensor<fp16, [1, 8, 512, 128]> var_2376 = mul(x = k_69, y = cos)[name = string("op_2376")];
            tensor<int32, [2]> var_2377_split_sizes_0 = const()[name = string("op_2377_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2377_axis_0 = const()[name = string("op_2377_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_2377_0, tensor<fp16, [1, 8, 512, 64]> var_2377_1 = split(axis = var_2377_axis_0, split_sizes = var_2377_split_sizes_0, x = k_69)[name = string("op_2377")];
            fp16 const_37_promoted = const()[name = string("const_37_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_2379 = mul(x = var_2377_1, y = const_37_promoted)[name = string("op_2379")];
            bool var_2381_interleave_0 = const()[name = string("op_2381_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_2381 = concat(axis = var_18, interleave = var_2381_interleave_0, values = (var_2379, var_2377_0))[name = string("op_2381")];
            tensor<fp16, [1, 8, 512, 128]> var_2382 = mul(x = var_2381, y = sin)[name = string("op_2382")];
            tensor<fp16, [1, 8, 512, 128]> x_393 = add(x = var_2376, y = var_2382)[name = string("x_393")];
            tensor<int32, [1]> var_2384_axes_0 = const()[name = string("op_2384_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2384 = expand_dims(axes = var_2384_axes_0, x = x_393)[name = string("op_2384")];
            tensor<int32, [5]> x_395_reps_0 = const()[name = string("x_395_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_395 = tile(reps = x_395_reps_0, x = var_2384)[name = string("x_395")];
            tensor<int32, [4]> var_2387 = const()[name = string("op_2387"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_23 = reshape(shape = var_2387, x = x_395)[name = string("key_23")];
            tensor<int32, [1]> var_2389_axes_0 = const()[name = string("op_2389_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_397 = transpose(perm = var_2340, x = var_2339)[name = string("transpose_149")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2389 = expand_dims(axes = var_2389_axes_0, x = x_397)[name = string("op_2389")];
            tensor<int32, [5]> x_399_reps_0 = const()[name = string("x_399_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_399 = tile(reps = x_399_reps_0, x = var_2389)[name = string("x_399")];
            tensor<int32, [4]> var_2392 = const()[name = string("op_2392"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_23 = reshape(shape = var_2392, x = x_399)[name = string("value_23")];
            bool var_2397_transpose_x_1 = const()[name = string("op_2397_transpose_x_1"), val = bool(false)];
            bool var_2397_transpose_y_1 = const()[name = string("op_2397_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2397_cast_fp16 = matmul(transpose_x = var_2397_transpose_x_1, transpose_y = var_2397_transpose_y_1, x = query_23, y = key_23)[name = string("op_2397_cast_fp16")];
            fp16 var_2398_to_fp16 = const()[name = string("op_2398_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_67_cast_fp16 = mul(x = var_2397_cast_fp16, y = var_2398_to_fp16)[name = string("attn_weights_67_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_69_cast_fp16 = add(x = attn_weights_67_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_69_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2402_cast_fp16 = softmax(axis = var_18, x = attn_weights_69_cast_fp16)[name = string("op_2402_cast_fp16")];
            bool var_2406_transpose_x_0 = const()[name = string("op_2406_transpose_x_0"), val = bool(false)];
            bool var_2406_transpose_y_0 = const()[name = string("op_2406_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2406_cast_fp16 = matmul(transpose_x = var_2406_transpose_x_0, transpose_y = var_2406_transpose_y_0, x = var_2402_cast_fp16, y = value_23)[name = string("op_2406_cast_fp16")];
            tensor<int32, [4]> var_2408 = const()[name = string("op_2408"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2411 = const()[name = string("op_2411"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_2409 = transpose(perm = var_2408, x = var_2406_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 512, 2048]> attn_out_69 = reshape(shape = var_2411, x = var_2409)[name = string("attn_out_69")];
            tensor<int32, [3]> var_2413 = const()[name = string("op_2413"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_11 = const()[name = string("squeeze_11"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1137521024)))];
            string var_2422_pad_type_0 = const()[name = string("op_2422_pad_type_0"), val = string("valid")];
            int32 var_2422_groups_0 = const()[name = string("op_2422_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2422_strides_0 = const()[name = string("op_2422_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2422_pad_0 = const()[name = string("op_2422_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2422_dilations_0 = const()[name = string("op_2422_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_2414 = transpose(perm = var_2413, x = attn_out_69)[name = string("transpose_147")];
            tensor<fp16, [1, 1024, 512]> var_2422 = conv(dilations = var_2422_dilations_0, groups = var_2422_groups_0, pad = var_2422_pad_0, pad_type = var_2422_pad_type_0, strides = var_2422_strides_0, weight = squeeze_11, x = var_2414)[name = string("op_2422")];
            tensor<int32, [3]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_71 = transpose(perm = var_2423, x = var_2422)[name = string("transpose_146")];
            tensor<fp16, [1, 512, 1024]> x_401_cast_fp16 = add(x = hidden_states_23_cast_fp16, y = attn_out_71)[name = string("x_401_cast_fp16")];
            fp16 var_6_promoted_47_to_fp16 = const()[name = string("op_6_promoted_47_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2429_cast_fp16 = pow(x = x_401_cast_fp16, y = var_6_promoted_47_to_fp16)[name = string("op_2429_cast_fp16")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = string("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_95_keep_dims_0 = const()[name = string("var_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_95_cast_fp16 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_2429_cast_fp16)[name = string("var_95_cast_fp16")];
            fp16 var_2432_to_fp16 = const()[name = string("op_2432_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2433_cast_fp16 = add(x = var_95_cast_fp16, y = var_2432_to_fp16)[name = string("op_2433_cast_fp16")];
            fp32 var_2434_epsilon_0 = const()[name = string("op_2434_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_2433_cast_fp16)[name = string("op_2434_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_405_cast_fp16 = mul(x = x_401_cast_fp16, y = var_2434_cast_fp16)[name = string("x_405_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141715392)))];
            tensor<fp16, [1, 512, 1024]> var_2437_cast_fp16 = mul(x = x_405_cast_fp16, y = encoder_layers_11_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2437_cast_fp16")];
            tensor<int32, [3]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2443 = transpose(perm = var_2442, x = var_2437_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 1024, 1, 512]> input_115 = expand_dims(axes = input_115_axes_0, x = var_2443)[name = string("input_115")];
            string input_117_pad_type_0 = const()[name = string("input_117_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_117_strides_0 = const()[name = string("input_117_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_117_pad_0 = const()[name = string("input_117_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_117_dilations_0 = const()[name = string("input_117_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_117_groups_0 = const()[name = string("input_117_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_117 = conv(dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = encoder_layers_11_mlp_gate_proj_weight, x = input_115)[name = string("input_117")];
            string up_23_pad_type_0 = const()[name = string("up_23_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_23_strides_0 = const()[name = string("up_23_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_23_pad_0 = const()[name = string("up_23_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_23_dilations_0 = const()[name = string("up_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_23_groups_0 = const()[name = string("up_23_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_23 = conv(dilations = up_23_dilations_0, groups = up_23_groups_0, pad = up_23_pad_0, pad_type = up_23_pad_type_0, strides = up_23_strides_0, weight = encoder_layers_11_mlp_up_proj_weight, x = input_115)[name = string("up_23")];
            tensor<fp16, [1, 3072, 1, 512]> var_2457 = silu(x = input_117)[name = string("op_2457")];
            tensor<fp16, [1, 3072, 1, 512]> input_119 = mul(x = var_2457, y = up_23)[name = string("input_119")];
            string var_2464_pad_type_0 = const()[name = string("op_2464_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2464_strides_0 = const()[name = string("op_2464_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2464_pad_0 = const()[name = string("op_2464_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2464_dilations_0 = const()[name = string("op_2464_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2464_groups_0 = const()[name = string("op_2464_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2464 = conv(dilations = var_2464_dilations_0, groups = var_2464_groups_0, pad = var_2464_pad_0, pad_type = var_2464_pad_type_0, strides = var_2464_strides_0, weight = encoder_layers_11_mlp_down_proj_weight, x = input_119)[name = string("op_2464")];
            tensor<int32, [1]> var_2465_axes_0 = const()[name = string("op_2465_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2465 = squeeze(axes = var_2465_axes_0, x = var_2464)[name = string("op_2465")];
            tensor<int32, [3]> var_2466 = const()[name = string("op_2466"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_23 = transpose(perm = var_2466, x = var_2465)[name = string("transpose_144")];
            tensor<fp16, [1, 512, 1024]> hidden_states_25_cast_fp16 = add(x = x_401_cast_fp16, y = mlp_out_23)[name = string("hidden_states_25_cast_fp16")];
            fp16 var_6_promoted_48_to_fp16 = const()[name = string("op_6_promoted_48_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2493_cast_fp16 = pow(x = hidden_states_25_cast_fp16, y = var_6_promoted_48_to_fp16)[name = string("op_2493_cast_fp16")];
            tensor<int32, [1]> var_97_axes_0 = const()[name = string("var_97_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_97_keep_dims_0 = const()[name = string("var_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_97_cast_fp16 = reduce_mean(axes = var_97_axes_0, keep_dims = var_97_keep_dims_0, x = var_2493_cast_fp16)[name = string("var_97_cast_fp16")];
            fp16 var_2496_to_fp16 = const()[name = string("op_2496_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2497_cast_fp16 = add(x = var_97_cast_fp16, y = var_2496_to_fp16)[name = string("op_2497_cast_fp16")];
            fp32 var_2498_epsilon_0 = const()[name = string("op_2498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2498_cast_fp16 = rsqrt(epsilon = var_2498_epsilon_0, x = var_2497_cast_fp16)[name = string("op_2498_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_411_cast_fp16 = mul(x = hidden_states_25_cast_fp16, y = var_2498_cast_fp16)[name = string("x_411_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_12_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_12_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141717504)))];
            tensor<fp16, [1, 512, 1024]> var_2501_cast_fp16 = mul(x = x_411_cast_fp16, y = encoder_layers_12_input_layernorm_weight_promoted_to_fp16)[name = string("op_2501_cast_fp16")];
            tensor<int32, [3]> var_2506 = const()[name = string("op_2506"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2507 = transpose(perm = var_2506, x = var_2501_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 1024, 1, 512]> input_121 = expand_dims(axes = input_121_axes_0, x = var_2507)[name = string("input_121")];
            string var_2514_pad_type_0 = const()[name = string("op_2514_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2514_strides_0 = const()[name = string("op_2514_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2514_pad_0 = const()[name = string("op_2514_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2514_dilations_0 = const()[name = string("op_2514_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2514_groups_0 = const()[name = string("op_2514_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_2514 = conv(dilations = var_2514_dilations_0, groups = var_2514_groups_0, pad = var_2514_pad_0, pad_type = var_2514_pad_type_0, strides = var_2514_strides_0, weight = encoder_layers_12_self_attn_q_proj_weight, x = input_121)[name = string("op_2514")];
            tensor<int32, [4]> var_2515 = const()[name = string("op_2515"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_2516 = reshape(shape = var_2515, x = var_2514)[name = string("op_2516")];
            tensor<int32, [4]> var_2517 = const()[name = string("op_2517"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2524_pad_type_0 = const()[name = string("op_2524_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2524_strides_0 = const()[name = string("op_2524_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2524_pad_0 = const()[name = string("op_2524_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2524_dilations_0 = const()[name = string("op_2524_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2524_groups_0 = const()[name = string("op_2524_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2524 = conv(dilations = var_2524_dilations_0, groups = var_2524_groups_0, pad = var_2524_pad_0, pad_type = var_2524_pad_type_0, strides = var_2524_strides_0, weight = encoder_layers_12_self_attn_k_proj_weight, x = input_121)[name = string("op_2524")];
            tensor<int32, [4]> var_2525 = const()[name = string("op_2525"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2526 = reshape(shape = var_2525, x = var_2524)[name = string("op_2526")];
            tensor<int32, [4]> var_2527 = const()[name = string("op_2527"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2534_pad_type_0 = const()[name = string("op_2534_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2534_strides_0 = const()[name = string("op_2534_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2534_pad_0 = const()[name = string("op_2534_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2534_dilations_0 = const()[name = string("op_2534_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2534_groups_0 = const()[name = string("op_2534_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2534 = conv(dilations = var_2534_dilations_0, groups = var_2534_groups_0, pad = var_2534_pad_0, pad_type = var_2534_pad_type_0, strides = var_2534_strides_0, weight = encoder_layers_12_self_attn_v_proj_weight, x = input_121)[name = string("op_2534")];
            tensor<int32, [4]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2536 = reshape(shape = var_2535, x = var_2534)[name = string("op_2536")];
            tensor<int32, [4]> var_2537 = const()[name = string("op_2537"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_49_to_fp16 = const()[name = string("op_6_promoted_49_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_73 = transpose(perm = var_2517, x = var_2516)[name = string("transpose_142")];
            tensor<fp16, [1, 16, 512, 128]> var_2543_cast_fp16 = pow(x = q_73, y = var_6_promoted_49_to_fp16)[name = string("op_2543_cast_fp16")];
            tensor<int32, [1]> var_99_axes_0 = const()[name = string("var_99_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_99_keep_dims_0 = const()[name = string("var_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_99_cast_fp16 = reduce_mean(axes = var_99_axes_0, keep_dims = var_99_keep_dims_0, x = var_2543_cast_fp16)[name = string("var_99_cast_fp16")];
            fp16 var_2546_to_fp16 = const()[name = string("op_2546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_2547_cast_fp16 = add(x = var_99_cast_fp16, y = var_2546_to_fp16)[name = string("op_2547_cast_fp16")];
            fp32 var_2548_epsilon_0 = const()[name = string("op_2548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_2548_cast_fp16 = rsqrt(epsilon = var_2548_epsilon_0, x = var_2547_cast_fp16)[name = string("op_2548_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_419_cast_fp16 = mul(x = q_73, y = var_2548_cast_fp16)[name = string("x_419_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_75 = mul(x = x_419_cast_fp16, y = encoder_layers_12_self_attn_q_norm_weight)[name = string("q_75")];
            fp16 var_6_promoted_50_to_fp16 = const()[name = string("op_6_promoted_50_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_73 = transpose(perm = var_2527, x = var_2526)[name = string("transpose_141")];
            tensor<fp16, [1, 8, 512, 128]> var_2556_cast_fp16 = pow(x = k_73, y = var_6_promoted_50_to_fp16)[name = string("op_2556_cast_fp16")];
            tensor<int32, [1]> var_101_axes_0 = const()[name = string("var_101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_101_keep_dims_0 = const()[name = string("var_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_101_cast_fp16 = reduce_mean(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_2556_cast_fp16)[name = string("var_101_cast_fp16")];
            fp16 var_2559_to_fp16 = const()[name = string("op_2559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_2560_cast_fp16 = add(x = var_101_cast_fp16, y = var_2559_to_fp16)[name = string("op_2560_cast_fp16")];
            fp32 var_2561_epsilon_0 = const()[name = string("op_2561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_2561_cast_fp16 = rsqrt(epsilon = var_2561_epsilon_0, x = var_2560_cast_fp16)[name = string("op_2561_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_425_cast_fp16 = mul(x = k_73, y = var_2561_cast_fp16)[name = string("x_425_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_75 = mul(x = x_425_cast_fp16, y = encoder_layers_12_self_attn_k_norm_weight)[name = string("k_75")];
            tensor<fp16, [1, 16, 512, 128]> var_2565 = mul(x = q_75, y = cos)[name = string("op_2565")];
            tensor<int32, [2]> var_2566_split_sizes_0 = const()[name = string("op_2566_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2566_axis_0 = const()[name = string("op_2566_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_2566_0, tensor<fp16, [1, 16, 512, 64]> var_2566_1 = split(axis = var_2566_axis_0, split_sizes = var_2566_split_sizes_0, x = q_75)[name = string("op_2566")];
            fp16 const_39_promoted = const()[name = string("const_39_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_2568 = mul(x = var_2566_1, y = const_39_promoted)[name = string("op_2568")];
            bool var_2570_interleave_0 = const()[name = string("op_2570_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2570 = concat(axis = var_18, interleave = var_2570_interleave_0, values = (var_2568, var_2566_0))[name = string("op_2570")];
            tensor<fp16, [1, 16, 512, 128]> var_2571 = mul(x = var_2570, y = sin)[name = string("op_2571")];
            tensor<fp16, [1, 16, 512, 128]> query_25 = add(x = var_2565, y = var_2571)[name = string("query_25")];
            tensor<fp16, [1, 8, 512, 128]> var_2573 = mul(x = k_75, y = cos)[name = string("op_2573")];
            tensor<int32, [2]> var_2574_split_sizes_0 = const()[name = string("op_2574_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2574_axis_0 = const()[name = string("op_2574_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_2574_0, tensor<fp16, [1, 8, 512, 64]> var_2574_1 = split(axis = var_2574_axis_0, split_sizes = var_2574_split_sizes_0, x = k_75)[name = string("op_2574")];
            fp16 const_40_promoted = const()[name = string("const_40_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_2576 = mul(x = var_2574_1, y = const_40_promoted)[name = string("op_2576")];
            bool var_2578_interleave_0 = const()[name = string("op_2578_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_2578 = concat(axis = var_18, interleave = var_2578_interleave_0, values = (var_2576, var_2574_0))[name = string("op_2578")];
            tensor<fp16, [1, 8, 512, 128]> var_2579 = mul(x = var_2578, y = sin)[name = string("op_2579")];
            tensor<fp16, [1, 8, 512, 128]> x_427 = add(x = var_2573, y = var_2579)[name = string("x_427")];
            tensor<int32, [1]> var_2581_axes_0 = const()[name = string("op_2581_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2581 = expand_dims(axes = var_2581_axes_0, x = x_427)[name = string("op_2581")];
            tensor<int32, [5]> x_429_reps_0 = const()[name = string("x_429_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_429 = tile(reps = x_429_reps_0, x = var_2581)[name = string("x_429")];
            tensor<int32, [4]> var_2584 = const()[name = string("op_2584"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_25 = reshape(shape = var_2584, x = x_429)[name = string("key_25")];
            tensor<int32, [1]> var_2586_axes_0 = const()[name = string("op_2586_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_431 = transpose(perm = var_2537, x = var_2536)[name = string("transpose_140")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2586 = expand_dims(axes = var_2586_axes_0, x = x_431)[name = string("op_2586")];
            tensor<int32, [5]> x_433_reps_0 = const()[name = string("x_433_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_433 = tile(reps = x_433_reps_0, x = var_2586)[name = string("x_433")];
            tensor<int32, [4]> var_2589 = const()[name = string("op_2589"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_25 = reshape(shape = var_2589, x = x_433)[name = string("value_25")];
            bool var_2594_transpose_x_1 = const()[name = string("op_2594_transpose_x_1"), val = bool(false)];
            bool var_2594_transpose_y_1 = const()[name = string("op_2594_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2594_cast_fp16 = matmul(transpose_x = var_2594_transpose_x_1, transpose_y = var_2594_transpose_y_1, x = query_25, y = key_25)[name = string("op_2594_cast_fp16")];
            fp16 var_2595_to_fp16 = const()[name = string("op_2595_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_73_cast_fp16 = mul(x = var_2594_cast_fp16, y = var_2595_to_fp16)[name = string("attn_weights_73_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_75_cast_fp16 = add(x = attn_weights_73_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_75_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2599_cast_fp16 = softmax(axis = var_18, x = attn_weights_75_cast_fp16)[name = string("op_2599_cast_fp16")];
            bool var_2603_transpose_x_0 = const()[name = string("op_2603_transpose_x_0"), val = bool(false)];
            bool var_2603_transpose_y_0 = const()[name = string("op_2603_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2603_cast_fp16 = matmul(transpose_x = var_2603_transpose_x_0, transpose_y = var_2603_transpose_y_0, x = var_2599_cast_fp16, y = value_25)[name = string("op_2603_cast_fp16")];
            tensor<int32, [4]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2608 = const()[name = string("op_2608"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_2606 = transpose(perm = var_2605, x = var_2603_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 512, 2048]> attn_out_75 = reshape(shape = var_2608, x = var_2606)[name = string("attn_out_75")];
            tensor<int32, [3]> var_2610 = const()[name = string("op_2610"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_12 = const()[name = string("squeeze_12"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1141719616)))];
            string var_2619_pad_type_0 = const()[name = string("op_2619_pad_type_0"), val = string("valid")];
            int32 var_2619_groups_0 = const()[name = string("op_2619_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2619_strides_0 = const()[name = string("op_2619_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2619_pad_0 = const()[name = string("op_2619_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2619_dilations_0 = const()[name = string("op_2619_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_2611 = transpose(perm = var_2610, x = attn_out_75)[name = string("transpose_138")];
            tensor<fp16, [1, 1024, 512]> var_2619 = conv(dilations = var_2619_dilations_0, groups = var_2619_groups_0, pad = var_2619_pad_0, pad_type = var_2619_pad_type_0, strides = var_2619_strides_0, weight = squeeze_12, x = var_2611)[name = string("op_2619")];
            tensor<int32, [3]> var_2620 = const()[name = string("op_2620"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_77 = transpose(perm = var_2620, x = var_2619)[name = string("transpose_137")];
            tensor<fp16, [1, 512, 1024]> x_435_cast_fp16 = add(x = hidden_states_25_cast_fp16, y = attn_out_77)[name = string("x_435_cast_fp16")];
            fp16 var_6_promoted_51_to_fp16 = const()[name = string("op_6_promoted_51_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2626_cast_fp16 = pow(x = x_435_cast_fp16, y = var_6_promoted_51_to_fp16)[name = string("op_2626_cast_fp16")];
            tensor<int32, [1]> var_103_axes_0 = const()[name = string("var_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_103_keep_dims_0 = const()[name = string("var_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_103_cast_fp16_0 = reduce_mean(axes = var_103_axes_0, keep_dims = var_103_keep_dims_0, x = var_2626_cast_fp16)[name = string("var_103_cast_fp16")];
            fp16 var_2629_to_fp16 = const()[name = string("op_2629_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2630_cast_fp16 = add(x = var_103_cast_fp16_0, y = var_2629_to_fp16)[name = string("op_2630_cast_fp16")];
            fp32 var_2631_epsilon_0 = const()[name = string("op_2631_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2631_cast_fp16 = rsqrt(epsilon = var_2631_epsilon_0, x = var_2630_cast_fp16)[name = string("op_2631_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_439_cast_fp16 = mul(x = x_435_cast_fp16, y = var_2631_cast_fp16)[name = string("x_439_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145913984)))];
            tensor<fp16, [1, 512, 1024]> var_2634_cast_fp16 = mul(x = x_439_cast_fp16, y = encoder_layers_12_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2634_cast_fp16")];
            tensor<int32, [3]> var_2639 = const()[name = string("op_2639"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2640 = transpose(perm = var_2639, x = var_2634_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 1024, 1, 512]> input_125 = expand_dims(axes = input_125_axes_0, x = var_2640)[name = string("input_125")];
            string input_127_pad_type_0 = const()[name = string("input_127_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_127_strides_0 = const()[name = string("input_127_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_127_pad_0 = const()[name = string("input_127_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_127_dilations_0 = const()[name = string("input_127_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_127_groups_0 = const()[name = string("input_127_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_127 = conv(dilations = input_127_dilations_0, groups = input_127_groups_0, pad = input_127_pad_0, pad_type = input_127_pad_type_0, strides = input_127_strides_0, weight = encoder_layers_12_mlp_gate_proj_weight, x = input_125)[name = string("input_127")];
            string up_25_pad_type_0 = const()[name = string("up_25_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_25_strides_0 = const()[name = string("up_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_25_pad_0 = const()[name = string("up_25_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_25_dilations_0 = const()[name = string("up_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_25_groups_0 = const()[name = string("up_25_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_25 = conv(dilations = up_25_dilations_0, groups = up_25_groups_0, pad = up_25_pad_0, pad_type = up_25_pad_type_0, strides = up_25_strides_0, weight = encoder_layers_12_mlp_up_proj_weight, x = input_125)[name = string("up_25")];
            tensor<fp16, [1, 3072, 1, 512]> var_2654 = silu(x = input_127)[name = string("op_2654")];
            tensor<fp16, [1, 3072, 1, 512]> input_129 = mul(x = var_2654, y = up_25)[name = string("input_129")];
            string var_2661_pad_type_0 = const()[name = string("op_2661_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2661_strides_0 = const()[name = string("op_2661_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2661_pad_0 = const()[name = string("op_2661_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2661_dilations_0 = const()[name = string("op_2661_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2661_groups_0 = const()[name = string("op_2661_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2661 = conv(dilations = var_2661_dilations_0, groups = var_2661_groups_0, pad = var_2661_pad_0, pad_type = var_2661_pad_type_0, strides = var_2661_strides_0, weight = encoder_layers_12_mlp_down_proj_weight, x = input_129)[name = string("op_2661")];
            tensor<int32, [1]> var_2662_axes_0 = const()[name = string("op_2662_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2662 = squeeze(axes = var_2662_axes_0, x = var_2661)[name = string("op_2662")];
            tensor<int32, [3]> var_2663 = const()[name = string("op_2663"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_25 = transpose(perm = var_2663, x = var_2662)[name = string("transpose_135")];
            tensor<fp16, [1, 512, 1024]> hidden_states_27_cast_fp16 = add(x = x_435_cast_fp16, y = mlp_out_25)[name = string("hidden_states_27_cast_fp16")];
            fp16 var_6_promoted_52_to_fp16 = const()[name = string("op_6_promoted_52_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2690_cast_fp16 = pow(x = hidden_states_27_cast_fp16, y = var_6_promoted_52_to_fp16)[name = string("op_2690_cast_fp16")];
            tensor<int32, [1]> var_105_axes_0 = const()[name = string("var_105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_105_keep_dims_0 = const()[name = string("var_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_105_cast_fp16 = reduce_mean(axes = var_105_axes_0, keep_dims = var_105_keep_dims_0, x = var_2690_cast_fp16)[name = string("var_105_cast_fp16")];
            fp16 var_2693_to_fp16 = const()[name = string("op_2693_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2694_cast_fp16 = add(x = var_105_cast_fp16, y = var_2693_to_fp16)[name = string("op_2694_cast_fp16")];
            fp32 var_2695_epsilon_0 = const()[name = string("op_2695_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2695_cast_fp16 = rsqrt(epsilon = var_2695_epsilon_0, x = var_2694_cast_fp16)[name = string("op_2695_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_445_cast_fp16 = mul(x = hidden_states_27_cast_fp16, y = var_2695_cast_fp16)[name = string("x_445_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_13_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_13_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145916096)))];
            tensor<fp16, [1, 512, 1024]> var_2698_cast_fp16 = mul(x = x_445_cast_fp16, y = encoder_layers_13_input_layernorm_weight_promoted_to_fp16)[name = string("op_2698_cast_fp16")];
            tensor<int32, [3]> var_2703 = const()[name = string("op_2703"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2704 = transpose(perm = var_2703, x = var_2698_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 1024, 1, 512]> input_131 = expand_dims(axes = input_131_axes_0, x = var_2704)[name = string("input_131")];
            string var_2711_pad_type_0 = const()[name = string("op_2711_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2711_strides_0 = const()[name = string("op_2711_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2711_pad_0 = const()[name = string("op_2711_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2711_dilations_0 = const()[name = string("op_2711_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2711_groups_0 = const()[name = string("op_2711_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_2711 = conv(dilations = var_2711_dilations_0, groups = var_2711_groups_0, pad = var_2711_pad_0, pad_type = var_2711_pad_type_0, strides = var_2711_strides_0, weight = encoder_layers_13_self_attn_q_proj_weight, x = input_131)[name = string("op_2711")];
            tensor<int32, [4]> var_2712 = const()[name = string("op_2712"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_2713 = reshape(shape = var_2712, x = var_2711)[name = string("op_2713")];
            tensor<int32, [4]> var_2714 = const()[name = string("op_2714"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2721_pad_type_0 = const()[name = string("op_2721_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2721_strides_0 = const()[name = string("op_2721_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2721_pad_0 = const()[name = string("op_2721_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2721_dilations_0 = const()[name = string("op_2721_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2721_groups_0 = const()[name = string("op_2721_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2721 = conv(dilations = var_2721_dilations_0, groups = var_2721_groups_0, pad = var_2721_pad_0, pad_type = var_2721_pad_type_0, strides = var_2721_strides_0, weight = encoder_layers_13_self_attn_k_proj_weight, x = input_131)[name = string("op_2721")];
            tensor<int32, [4]> var_2722 = const()[name = string("op_2722"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2723 = reshape(shape = var_2722, x = var_2721)[name = string("op_2723")];
            tensor<int32, [4]> var_2724 = const()[name = string("op_2724"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2731_pad_type_0 = const()[name = string("op_2731_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2731_strides_0 = const()[name = string("op_2731_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2731_pad_0 = const()[name = string("op_2731_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2731_dilations_0 = const()[name = string("op_2731_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2731_groups_0 = const()[name = string("op_2731_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2731 = conv(dilations = var_2731_dilations_0, groups = var_2731_groups_0, pad = var_2731_pad_0, pad_type = var_2731_pad_type_0, strides = var_2731_strides_0, weight = encoder_layers_13_self_attn_v_proj_weight, x = input_131)[name = string("op_2731")];
            tensor<int32, [4]> var_2732 = const()[name = string("op_2732"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2733 = reshape(shape = var_2732, x = var_2731)[name = string("op_2733")];
            tensor<int32, [4]> var_2734 = const()[name = string("op_2734"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_53_to_fp16 = const()[name = string("op_6_promoted_53_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_79 = transpose(perm = var_2714, x = var_2713)[name = string("transpose_133")];
            tensor<fp16, [1, 16, 512, 128]> var_2740_cast_fp16 = pow(x = q_79, y = var_6_promoted_53_to_fp16)[name = string("op_2740_cast_fp16")];
            tensor<int32, [1]> var_107_axes_0 = const()[name = string("var_107_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_107_keep_dims_0 = const()[name = string("var_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_107_cast_fp16 = reduce_mean(axes = var_107_axes_0, keep_dims = var_107_keep_dims_0, x = var_2740_cast_fp16)[name = string("var_107_cast_fp16")];
            fp16 var_2743_to_fp16 = const()[name = string("op_2743_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_2744_cast_fp16 = add(x = var_107_cast_fp16, y = var_2743_to_fp16)[name = string("op_2744_cast_fp16")];
            fp32 var_2745_epsilon_0 = const()[name = string("op_2745_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_2745_cast_fp16 = rsqrt(epsilon = var_2745_epsilon_0, x = var_2744_cast_fp16)[name = string("op_2745_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_453_cast_fp16 = mul(x = q_79, y = var_2745_cast_fp16)[name = string("x_453_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_81 = mul(x = x_453_cast_fp16, y = encoder_layers_13_self_attn_q_norm_weight)[name = string("q_81")];
            fp16 var_6_promoted_54_to_fp16 = const()[name = string("op_6_promoted_54_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_79 = transpose(perm = var_2724, x = var_2723)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 512, 128]> var_2753_cast_fp16 = pow(x = k_79, y = var_6_promoted_54_to_fp16)[name = string("op_2753_cast_fp16")];
            tensor<int32, [1]> var_109_axes_0 = const()[name = string("var_109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_109_keep_dims_0 = const()[name = string("var_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_109_cast_fp16 = reduce_mean(axes = var_109_axes_0, keep_dims = var_109_keep_dims_0, x = var_2753_cast_fp16)[name = string("var_109_cast_fp16")];
            fp16 var_2756_to_fp16 = const()[name = string("op_2756_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_2757_cast_fp16 = add(x = var_109_cast_fp16, y = var_2756_to_fp16)[name = string("op_2757_cast_fp16")];
            fp32 var_2758_epsilon_0 = const()[name = string("op_2758_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_2758_cast_fp16 = rsqrt(epsilon = var_2758_epsilon_0, x = var_2757_cast_fp16)[name = string("op_2758_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_459_cast_fp16 = mul(x = k_79, y = var_2758_cast_fp16)[name = string("x_459_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_81 = mul(x = x_459_cast_fp16, y = encoder_layers_13_self_attn_k_norm_weight)[name = string("k_81")];
            tensor<fp16, [1, 16, 512, 128]> var_2762 = mul(x = q_81, y = cos)[name = string("op_2762")];
            tensor<int32, [2]> var_2763_split_sizes_0 = const()[name = string("op_2763_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2763_axis_0 = const()[name = string("op_2763_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_2763_0, tensor<fp16, [1, 16, 512, 64]> var_2763_1 = split(axis = var_2763_axis_0, split_sizes = var_2763_split_sizes_0, x = q_81)[name = string("op_2763")];
            fp16 const_42_promoted = const()[name = string("const_42_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_2765 = mul(x = var_2763_1, y = const_42_promoted)[name = string("op_2765")];
            bool var_2767_interleave_0 = const()[name = string("op_2767_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2767 = concat(axis = var_18, interleave = var_2767_interleave_0, values = (var_2765, var_2763_0))[name = string("op_2767")];
            tensor<fp16, [1, 16, 512, 128]> var_2768 = mul(x = var_2767, y = sin)[name = string("op_2768")];
            tensor<fp16, [1, 16, 512, 128]> query_27 = add(x = var_2762, y = var_2768)[name = string("query_27")];
            tensor<fp16, [1, 8, 512, 128]> var_2770 = mul(x = k_81, y = cos)[name = string("op_2770")];
            tensor<int32, [2]> var_2771_split_sizes_0 = const()[name = string("op_2771_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2771_axis_0 = const()[name = string("op_2771_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_2771_0, tensor<fp16, [1, 8, 512, 64]> var_2771_1 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = k_81)[name = string("op_2771")];
            fp16 const_43_promoted = const()[name = string("const_43_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_2773 = mul(x = var_2771_1, y = const_43_promoted)[name = string("op_2773")];
            bool var_2775_interleave_0 = const()[name = string("op_2775_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_2775 = concat(axis = var_18, interleave = var_2775_interleave_0, values = (var_2773, var_2771_0))[name = string("op_2775")];
            tensor<fp16, [1, 8, 512, 128]> var_2776 = mul(x = var_2775, y = sin)[name = string("op_2776")];
            tensor<fp16, [1, 8, 512, 128]> x_461 = add(x = var_2770, y = var_2776)[name = string("x_461")];
            tensor<int32, [1]> var_2778_axes_0 = const()[name = string("op_2778_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2778 = expand_dims(axes = var_2778_axes_0, x = x_461)[name = string("op_2778")];
            tensor<int32, [5]> x_463_reps_0 = const()[name = string("x_463_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_463 = tile(reps = x_463_reps_0, x = var_2778)[name = string("x_463")];
            tensor<int32, [4]> var_2781 = const()[name = string("op_2781"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_27 = reshape(shape = var_2781, x = x_463)[name = string("key_27")];
            tensor<int32, [1]> var_2783_axes_0 = const()[name = string("op_2783_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_465 = transpose(perm = var_2734, x = var_2733)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2783 = expand_dims(axes = var_2783_axes_0, x = x_465)[name = string("op_2783")];
            tensor<int32, [5]> x_467_reps_0 = const()[name = string("x_467_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_467 = tile(reps = x_467_reps_0, x = var_2783)[name = string("x_467")];
            tensor<int32, [4]> var_2786 = const()[name = string("op_2786"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_27 = reshape(shape = var_2786, x = x_467)[name = string("value_27")];
            bool var_2791_transpose_x_1 = const()[name = string("op_2791_transpose_x_1"), val = bool(false)];
            bool var_2791_transpose_y_1 = const()[name = string("op_2791_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2791_cast_fp16 = matmul(transpose_x = var_2791_transpose_x_1, transpose_y = var_2791_transpose_y_1, x = query_27, y = key_27)[name = string("op_2791_cast_fp16")];
            fp16 var_2792_to_fp16 = const()[name = string("op_2792_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_79_cast_fp16 = mul(x = var_2791_cast_fp16, y = var_2792_to_fp16)[name = string("attn_weights_79_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_81_cast_fp16 = add(x = attn_weights_79_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_81_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2796_cast_fp16 = softmax(axis = var_18, x = attn_weights_81_cast_fp16)[name = string("op_2796_cast_fp16")];
            bool var_2800_transpose_x_0 = const()[name = string("op_2800_transpose_x_0"), val = bool(false)];
            bool var_2800_transpose_y_0 = const()[name = string("op_2800_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2800_cast_fp16 = matmul(transpose_x = var_2800_transpose_x_0, transpose_y = var_2800_transpose_y_0, x = var_2796_cast_fp16, y = value_27)[name = string("op_2800_cast_fp16")];
            tensor<int32, [4]> var_2802 = const()[name = string("op_2802"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2805 = const()[name = string("op_2805"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_2803 = transpose(perm = var_2802, x = var_2800_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 512, 2048]> attn_out_81 = reshape(shape = var_2805, x = var_2803)[name = string("attn_out_81")];
            tensor<int32, [3]> var_2807 = const()[name = string("op_2807"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_13 = const()[name = string("squeeze_13"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1145918208)))];
            string var_2816_pad_type_0 = const()[name = string("op_2816_pad_type_0"), val = string("valid")];
            int32 var_2816_groups_0 = const()[name = string("op_2816_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_2816_strides_0 = const()[name = string("op_2816_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_2816_pad_0 = const()[name = string("op_2816_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_2816_dilations_0 = const()[name = string("op_2816_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_2808 = transpose(perm = var_2807, x = attn_out_81)[name = string("transpose_129")];
            tensor<fp16, [1, 1024, 512]> var_2816 = conv(dilations = var_2816_dilations_0, groups = var_2816_groups_0, pad = var_2816_pad_0, pad_type = var_2816_pad_type_0, strides = var_2816_strides_0, weight = squeeze_13, x = var_2808)[name = string("op_2816")];
            tensor<int32, [3]> var_2817 = const()[name = string("op_2817"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_83 = transpose(perm = var_2817, x = var_2816)[name = string("transpose_128")];
            tensor<fp16, [1, 512, 1024]> x_469_cast_fp16 = add(x = hidden_states_27_cast_fp16, y = attn_out_83)[name = string("x_469_cast_fp16")];
            fp16 var_6_promoted_55_to_fp16 = const()[name = string("op_6_promoted_55_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2823_cast_fp16 = pow(x = x_469_cast_fp16, y = var_6_promoted_55_to_fp16)[name = string("op_2823_cast_fp16")];
            tensor<int32, [1]> var_111_axes_0 = const()[name = string("var_111_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_111_keep_dims_0 = const()[name = string("var_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_111_cast_fp16 = reduce_mean(axes = var_111_axes_0, keep_dims = var_111_keep_dims_0, x = var_2823_cast_fp16)[name = string("var_111_cast_fp16")];
            fp16 var_2826_to_fp16 = const()[name = string("op_2826_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2827_cast_fp16 = add(x = var_111_cast_fp16, y = var_2826_to_fp16)[name = string("op_2827_cast_fp16")];
            fp32 var_2828_epsilon_0 = const()[name = string("op_2828_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2828_cast_fp16 = rsqrt(epsilon = var_2828_epsilon_0, x = var_2827_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_473_cast_fp16 = mul(x = x_469_cast_fp16, y = var_2828_cast_fp16)[name = string("x_473_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150112576)))];
            tensor<fp16, [1, 512, 1024]> var_2831_cast_fp16 = mul(x = x_473_cast_fp16, y = encoder_layers_13_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_2831_cast_fp16")];
            tensor<int32, [3]> var_2836 = const()[name = string("op_2836"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_135_axes_0 = const()[name = string("input_135_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2837 = transpose(perm = var_2836, x = var_2831_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1024, 1, 512]> input_135 = expand_dims(axes = input_135_axes_0, x = var_2837)[name = string("input_135")];
            string input_137_pad_type_0 = const()[name = string("input_137_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_137_strides_0 = const()[name = string("input_137_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_137_pad_0 = const()[name = string("input_137_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_137_dilations_0 = const()[name = string("input_137_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_137_groups_0 = const()[name = string("input_137_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_137 = conv(dilations = input_137_dilations_0, groups = input_137_groups_0, pad = input_137_pad_0, pad_type = input_137_pad_type_0, strides = input_137_strides_0, weight = encoder_layers_13_mlp_gate_proj_weight, x = input_135)[name = string("input_137")];
            string up_27_pad_type_0 = const()[name = string("up_27_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_27_strides_0 = const()[name = string("up_27_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_27_pad_0 = const()[name = string("up_27_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_27_dilations_0 = const()[name = string("up_27_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_27_groups_0 = const()[name = string("up_27_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_27 = conv(dilations = up_27_dilations_0, groups = up_27_groups_0, pad = up_27_pad_0, pad_type = up_27_pad_type_0, strides = up_27_strides_0, weight = encoder_layers_13_mlp_up_proj_weight, x = input_135)[name = string("up_27")];
            tensor<fp16, [1, 3072, 1, 512]> var_2851 = silu(x = input_137)[name = string("op_2851")];
            tensor<fp16, [1, 3072, 1, 512]> input_139 = mul(x = var_2851, y = up_27)[name = string("input_139")];
            string var_2858_pad_type_0 = const()[name = string("op_2858_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2858_strides_0 = const()[name = string("op_2858_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2858_pad_0 = const()[name = string("op_2858_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2858_dilations_0 = const()[name = string("op_2858_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2858_groups_0 = const()[name = string("op_2858_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2858 = conv(dilations = var_2858_dilations_0, groups = var_2858_groups_0, pad = var_2858_pad_0, pad_type = var_2858_pad_type_0, strides = var_2858_strides_0, weight = encoder_layers_13_mlp_down_proj_weight, x = input_139)[name = string("op_2858")];
            tensor<int32, [1]> var_2859_axes_0 = const()[name = string("op_2859_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2859 = squeeze(axes = var_2859_axes_0, x = var_2858)[name = string("op_2859")];
            tensor<int32, [3]> var_2860 = const()[name = string("op_2860"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_27 = transpose(perm = var_2860, x = var_2859)[name = string("transpose_126")];
            tensor<fp16, [1, 512, 1024]> hidden_states_29_cast_fp16 = add(x = x_469_cast_fp16, y = mlp_out_27)[name = string("hidden_states_29_cast_fp16")];
            fp16 var_6_promoted_56_to_fp16 = const()[name = string("op_6_promoted_56_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_2887_cast_fp16 = pow(x = hidden_states_29_cast_fp16, y = var_6_promoted_56_to_fp16)[name = string("op_2887_cast_fp16")];
            tensor<int32, [1]> var_113_axes_0 = const()[name = string("var_113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_113_keep_dims_0 = const()[name = string("var_113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_113_cast_fp16 = reduce_mean(axes = var_113_axes_0, keep_dims = var_113_keep_dims_0, x = var_2887_cast_fp16)[name = string("var_113_cast_fp16")];
            fp16 var_2890_to_fp16 = const()[name = string("op_2890_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_2891_cast_fp16 = add(x = var_113_cast_fp16, y = var_2890_to_fp16)[name = string("op_2891_cast_fp16")];
            fp32 var_2892_epsilon_0 = const()[name = string("op_2892_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_2892_cast_fp16 = rsqrt(epsilon = var_2892_epsilon_0, x = var_2891_cast_fp16)[name = string("op_2892_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_479_cast_fp16 = mul(x = hidden_states_29_cast_fp16, y = var_2892_cast_fp16)[name = string("x_479_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_14_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_14_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150114688)))];
            tensor<fp16, [1, 512, 1024]> var_2895_cast_fp16 = mul(x = x_479_cast_fp16, y = encoder_layers_14_input_layernorm_weight_promoted_to_fp16)[name = string("op_2895_cast_fp16")];
            tensor<int32, [3]> var_2900 = const()[name = string("op_2900"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_2901 = transpose(perm = var_2900, x = var_2895_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 1024, 1, 512]> input_141 = expand_dims(axes = input_141_axes_0, x = var_2901)[name = string("input_141")];
            string var_2908_pad_type_0 = const()[name = string("op_2908_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2908_strides_0 = const()[name = string("op_2908_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2908_pad_0 = const()[name = string("op_2908_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2908_dilations_0 = const()[name = string("op_2908_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2908_groups_0 = const()[name = string("op_2908_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_2908 = conv(dilations = var_2908_dilations_0, groups = var_2908_groups_0, pad = var_2908_pad_0, pad_type = var_2908_pad_type_0, strides = var_2908_strides_0, weight = encoder_layers_14_self_attn_q_proj_weight, x = input_141)[name = string("op_2908")];
            tensor<int32, [4]> var_2909 = const()[name = string("op_2909"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_2910 = reshape(shape = var_2909, x = var_2908)[name = string("op_2910")];
            tensor<int32, [4]> var_2911 = const()[name = string("op_2911"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2918_pad_type_0 = const()[name = string("op_2918_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2918_strides_0 = const()[name = string("op_2918_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2918_pad_0 = const()[name = string("op_2918_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2918_dilations_0 = const()[name = string("op_2918_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2918_groups_0 = const()[name = string("op_2918_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2918 = conv(dilations = var_2918_dilations_0, groups = var_2918_groups_0, pad = var_2918_pad_0, pad_type = var_2918_pad_type_0, strides = var_2918_strides_0, weight = encoder_layers_14_self_attn_k_proj_weight, x = input_141)[name = string("op_2918")];
            tensor<int32, [4]> var_2919 = const()[name = string("op_2919"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2920 = reshape(shape = var_2919, x = var_2918)[name = string("op_2920")];
            tensor<int32, [4]> var_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_2928_pad_type_0 = const()[name = string("op_2928_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_2928_strides_0 = const()[name = string("op_2928_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_2928_pad_0 = const()[name = string("op_2928_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_2928_dilations_0 = const()[name = string("op_2928_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_2928_groups_0 = const()[name = string("op_2928_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_2928 = conv(dilations = var_2928_dilations_0, groups = var_2928_groups_0, pad = var_2928_pad_0, pad_type = var_2928_pad_type_0, strides = var_2928_strides_0, weight = encoder_layers_14_self_attn_v_proj_weight, x = input_141)[name = string("op_2928")];
            tensor<int32, [4]> var_2929 = const()[name = string("op_2929"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_2930 = reshape(shape = var_2929, x = var_2928)[name = string("op_2930")];
            tensor<int32, [4]> var_2931 = const()[name = string("op_2931"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_57_to_fp16 = const()[name = string("op_6_promoted_57_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_85 = transpose(perm = var_2911, x = var_2910)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 512, 128]> var_2937_cast_fp16 = pow(x = q_85, y = var_6_promoted_57_to_fp16)[name = string("op_2937_cast_fp16")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = string("var_115_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_115_keep_dims_0 = const()[name = string("var_115_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_115_cast_fp16 = reduce_mean(axes = var_115_axes_0, keep_dims = var_115_keep_dims_0, x = var_2937_cast_fp16)[name = string("var_115_cast_fp16")];
            fp16 var_2940_to_fp16 = const()[name = string("op_2940_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_2941_cast_fp16 = add(x = var_115_cast_fp16, y = var_2940_to_fp16)[name = string("op_2941_cast_fp16")];
            fp32 var_2942_epsilon_0 = const()[name = string("op_2942_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_2942_cast_fp16 = rsqrt(epsilon = var_2942_epsilon_0, x = var_2941_cast_fp16)[name = string("op_2942_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_487_cast_fp16 = mul(x = q_85, y = var_2942_cast_fp16)[name = string("x_487_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_87 = mul(x = x_487_cast_fp16, y = encoder_layers_14_self_attn_q_norm_weight)[name = string("q_87")];
            fp16 var_6_promoted_58_to_fp16 = const()[name = string("op_6_promoted_58_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_85 = transpose(perm = var_2921, x = var_2920)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 512, 128]> var_2950_cast_fp16 = pow(x = k_85, y = var_6_promoted_58_to_fp16)[name = string("op_2950_cast_fp16")];
            tensor<int32, [1]> var_117_axes_0 = const()[name = string("var_117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_117_keep_dims_0 = const()[name = string("var_117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_117_cast_fp16 = reduce_mean(axes = var_117_axes_0, keep_dims = var_117_keep_dims_0, x = var_2950_cast_fp16)[name = string("var_117_cast_fp16")];
            fp16 var_2953_to_fp16 = const()[name = string("op_2953_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_2954_cast_fp16 = add(x = var_117_cast_fp16, y = var_2953_to_fp16)[name = string("op_2954_cast_fp16")];
            fp32 var_2955_epsilon_0 = const()[name = string("op_2955_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_2955_cast_fp16 = rsqrt(epsilon = var_2955_epsilon_0, x = var_2954_cast_fp16)[name = string("op_2955_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_493_cast_fp16 = mul(x = k_85, y = var_2955_cast_fp16)[name = string("x_493_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_87 = mul(x = x_493_cast_fp16, y = encoder_layers_14_self_attn_k_norm_weight)[name = string("k_87")];
            tensor<fp16, [1, 16, 512, 128]> var_2959 = mul(x = q_87, y = cos)[name = string("op_2959")];
            tensor<int32, [2]> var_2960_split_sizes_0 = const()[name = string("op_2960_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2960_axis_0 = const()[name = string("op_2960_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_2960_0, tensor<fp16, [1, 16, 512, 64]> var_2960_1 = split(axis = var_2960_axis_0, split_sizes = var_2960_split_sizes_0, x = q_87)[name = string("op_2960")];
            fp16 const_45_promoted = const()[name = string("const_45_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_2962 = mul(x = var_2960_1, y = const_45_promoted)[name = string("op_2962")];
            bool var_2964_interleave_0 = const()[name = string("op_2964_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2964 = concat(axis = var_18, interleave = var_2964_interleave_0, values = (var_2962, var_2960_0))[name = string("op_2964")];
            tensor<fp16, [1, 16, 512, 128]> var_2965 = mul(x = var_2964, y = sin)[name = string("op_2965")];
            tensor<fp16, [1, 16, 512, 128]> query_29 = add(x = var_2959, y = var_2965)[name = string("query_29")];
            tensor<fp16, [1, 8, 512, 128]> var_2967 = mul(x = k_87, y = cos)[name = string("op_2967")];
            tensor<int32, [2]> var_2968_split_sizes_0 = const()[name = string("op_2968_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_2968_axis_0 = const()[name = string("op_2968_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_2968_0, tensor<fp16, [1, 8, 512, 64]> var_2968_1 = split(axis = var_2968_axis_0, split_sizes = var_2968_split_sizes_0, x = k_87)[name = string("op_2968")];
            fp16 const_46_promoted = const()[name = string("const_46_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_2970 = mul(x = var_2968_1, y = const_46_promoted)[name = string("op_2970")];
            bool var_2972_interleave_0 = const()[name = string("op_2972_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_2972 = concat(axis = var_18, interleave = var_2972_interleave_0, values = (var_2970, var_2968_0))[name = string("op_2972")];
            tensor<fp16, [1, 8, 512, 128]> var_2973 = mul(x = var_2972, y = sin)[name = string("op_2973")];
            tensor<fp16, [1, 8, 512, 128]> x_495 = add(x = var_2967, y = var_2973)[name = string("x_495")];
            tensor<int32, [1]> var_2975_axes_0 = const()[name = string("op_2975_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2975 = expand_dims(axes = var_2975_axes_0, x = x_495)[name = string("op_2975")];
            tensor<int32, [5]> x_497_reps_0 = const()[name = string("x_497_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_497 = tile(reps = x_497_reps_0, x = var_2975)[name = string("x_497")];
            tensor<int32, [4]> var_2978 = const()[name = string("op_2978"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_29 = reshape(shape = var_2978, x = x_497)[name = string("key_29")];
            tensor<int32, [1]> var_2980_axes_0 = const()[name = string("op_2980_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_499 = transpose(perm = var_2931, x = var_2930)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_2980 = expand_dims(axes = var_2980_axes_0, x = x_499)[name = string("op_2980")];
            tensor<int32, [5]> x_501_reps_0 = const()[name = string("x_501_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_501 = tile(reps = x_501_reps_0, x = var_2980)[name = string("x_501")];
            tensor<int32, [4]> var_2983 = const()[name = string("op_2983"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_29 = reshape(shape = var_2983, x = x_501)[name = string("value_29")];
            bool var_2988_transpose_x_1 = const()[name = string("op_2988_transpose_x_1"), val = bool(false)];
            bool var_2988_transpose_y_1 = const()[name = string("op_2988_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_2988_cast_fp16 = matmul(transpose_x = var_2988_transpose_x_1, transpose_y = var_2988_transpose_y_1, x = query_29, y = key_29)[name = string("op_2988_cast_fp16")];
            fp16 var_2989_to_fp16 = const()[name = string("op_2989_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_85_cast_fp16 = mul(x = var_2988_cast_fp16, y = var_2989_to_fp16)[name = string("attn_weights_85_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_87_cast_fp16 = add(x = attn_weights_85_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_87_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_2993_cast_fp16 = softmax(axis = var_18, x = attn_weights_87_cast_fp16)[name = string("op_2993_cast_fp16")];
            bool var_2997_transpose_x_0 = const()[name = string("op_2997_transpose_x_0"), val = bool(false)];
            bool var_2997_transpose_y_0 = const()[name = string("op_2997_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_2997_cast_fp16 = matmul(transpose_x = var_2997_transpose_x_0, transpose_y = var_2997_transpose_y_0, x = var_2993_cast_fp16, y = value_29)[name = string("op_2997_cast_fp16")];
            tensor<int32, [4]> var_2999 = const()[name = string("op_2999"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3002 = const()[name = string("op_3002"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3000 = transpose(perm = var_2999, x = var_2997_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 512, 2048]> attn_out_87 = reshape(shape = var_3002, x = var_3000)[name = string("attn_out_87")];
            tensor<int32, [3]> var_3004 = const()[name = string("op_3004"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_14 = const()[name = string("squeeze_14"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1150116800)))];
            string var_3013_pad_type_0 = const()[name = string("op_3013_pad_type_0"), val = string("valid")];
            int32 var_3013_groups_0 = const()[name = string("op_3013_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3013_strides_0 = const()[name = string("op_3013_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3013_pad_0 = const()[name = string("op_3013_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3013_dilations_0 = const()[name = string("op_3013_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3005 = transpose(perm = var_3004, x = attn_out_87)[name = string("transpose_120")];
            tensor<fp16, [1, 1024, 512]> var_3013 = conv(dilations = var_3013_dilations_0, groups = var_3013_groups_0, pad = var_3013_pad_0, pad_type = var_3013_pad_type_0, strides = var_3013_strides_0, weight = squeeze_14, x = var_3005)[name = string("op_3013")];
            tensor<int32, [3]> var_3014 = const()[name = string("op_3014"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_89 = transpose(perm = var_3014, x = var_3013)[name = string("transpose_119")];
            tensor<fp16, [1, 512, 1024]> x_503_cast_fp16 = add(x = hidden_states_29_cast_fp16, y = attn_out_89)[name = string("x_503_cast_fp16")];
            fp16 var_6_promoted_59_to_fp16 = const()[name = string("op_6_promoted_59_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3020_cast_fp16 = pow(x = x_503_cast_fp16, y = var_6_promoted_59_to_fp16)[name = string("op_3020_cast_fp16")];
            tensor<int32, [1]> var_119_axes_0 = const()[name = string("var_119_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_119_keep_dims_0 = const()[name = string("var_119_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_119_cast_fp16 = reduce_mean(axes = var_119_axes_0, keep_dims = var_119_keep_dims_0, x = var_3020_cast_fp16)[name = string("var_119_cast_fp16")];
            fp16 var_3023_to_fp16 = const()[name = string("op_3023_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3024_cast_fp16 = add(x = var_119_cast_fp16, y = var_3023_to_fp16)[name = string("op_3024_cast_fp16")];
            fp32 var_3025_epsilon_0 = const()[name = string("op_3025_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3025_cast_fp16 = rsqrt(epsilon = var_3025_epsilon_0, x = var_3024_cast_fp16)[name = string("op_3025_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_507_cast_fp16 = mul(x = x_503_cast_fp16, y = var_3025_cast_fp16)[name = string("x_507_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154311168)))];
            tensor<fp16, [1, 512, 1024]> var_3028_cast_fp16 = mul(x = x_507_cast_fp16, y = encoder_layers_14_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3028_cast_fp16")];
            tensor<int32, [3]> var_3033 = const()[name = string("op_3033"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_145_axes_0 = const()[name = string("input_145_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3034 = transpose(perm = var_3033, x = var_3028_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 1024, 1, 512]> input_145 = expand_dims(axes = input_145_axes_0, x = var_3034)[name = string("input_145")];
            string input_147_pad_type_0 = const()[name = string("input_147_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_147_strides_0 = const()[name = string("input_147_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_147_pad_0 = const()[name = string("input_147_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_147_dilations_0 = const()[name = string("input_147_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_147_groups_0 = const()[name = string("input_147_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_147 = conv(dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = encoder_layers_14_mlp_gate_proj_weight, x = input_145)[name = string("input_147")];
            string up_29_pad_type_0 = const()[name = string("up_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_29_strides_0 = const()[name = string("up_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_29_pad_0 = const()[name = string("up_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_29_dilations_0 = const()[name = string("up_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_29_groups_0 = const()[name = string("up_29_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_29 = conv(dilations = up_29_dilations_0, groups = up_29_groups_0, pad = up_29_pad_0, pad_type = up_29_pad_type_0, strides = up_29_strides_0, weight = encoder_layers_14_mlp_up_proj_weight, x = input_145)[name = string("up_29")];
            tensor<fp16, [1, 3072, 1, 512]> var_3048 = silu(x = input_147)[name = string("op_3048")];
            tensor<fp16, [1, 3072, 1, 512]> input_149 = mul(x = var_3048, y = up_29)[name = string("input_149")];
            string var_3055_pad_type_0 = const()[name = string("op_3055_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3055_strides_0 = const()[name = string("op_3055_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3055_pad_0 = const()[name = string("op_3055_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3055_dilations_0 = const()[name = string("op_3055_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3055_groups_0 = const()[name = string("op_3055_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3055 = conv(dilations = var_3055_dilations_0, groups = var_3055_groups_0, pad = var_3055_pad_0, pad_type = var_3055_pad_type_0, strides = var_3055_strides_0, weight = encoder_layers_14_mlp_down_proj_weight, x = input_149)[name = string("op_3055")];
            tensor<int32, [1]> var_3056_axes_0 = const()[name = string("op_3056_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3056 = squeeze(axes = var_3056_axes_0, x = var_3055)[name = string("op_3056")];
            tensor<int32, [3]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_29 = transpose(perm = var_3057, x = var_3056)[name = string("transpose_117")];
            tensor<fp16, [1, 512, 1024]> hidden_states_31_cast_fp16 = add(x = x_503_cast_fp16, y = mlp_out_29)[name = string("hidden_states_31_cast_fp16")];
            fp16 var_6_promoted_60_to_fp16 = const()[name = string("op_6_promoted_60_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3084_cast_fp16 = pow(x = hidden_states_31_cast_fp16, y = var_6_promoted_60_to_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [1]> var_121_axes_0 = const()[name = string("var_121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_121_keep_dims_0 = const()[name = string("var_121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_121_cast_fp16 = reduce_mean(axes = var_121_axes_0, keep_dims = var_121_keep_dims_0, x = var_3084_cast_fp16)[name = string("var_121_cast_fp16")];
            fp16 var_3087_to_fp16 = const()[name = string("op_3087_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3088_cast_fp16 = add(x = var_121_cast_fp16, y = var_3087_to_fp16)[name = string("op_3088_cast_fp16")];
            fp32 var_3089_epsilon_0 = const()[name = string("op_3089_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3089_cast_fp16 = rsqrt(epsilon = var_3089_epsilon_0, x = var_3088_cast_fp16)[name = string("op_3089_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_513_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = var_3089_cast_fp16)[name = string("x_513_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_15_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_15_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154313280)))];
            tensor<fp16, [1, 512, 1024]> var_3092_cast_fp16 = mul(x = x_513_cast_fp16, y = encoder_layers_15_input_layernorm_weight_promoted_to_fp16)[name = string("op_3092_cast_fp16")];
            tensor<int32, [3]> var_3097 = const()[name = string("op_3097"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_151_axes_0 = const()[name = string("input_151_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3098 = transpose(perm = var_3097, x = var_3092_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 1024, 1, 512]> input_151 = expand_dims(axes = input_151_axes_0, x = var_3098)[name = string("input_151")];
            string var_3105_pad_type_0 = const()[name = string("op_3105_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3105_strides_0 = const()[name = string("op_3105_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3105_pad_0 = const()[name = string("op_3105_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3105_dilations_0 = const()[name = string("op_3105_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3105_groups_0 = const()[name = string("op_3105_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_3105 = conv(dilations = var_3105_dilations_0, groups = var_3105_groups_0, pad = var_3105_pad_0, pad_type = var_3105_pad_type_0, strides = var_3105_strides_0, weight = encoder_layers_15_self_attn_q_proj_weight, x = input_151)[name = string("op_3105")];
            tensor<int32, [4]> var_3106 = const()[name = string("op_3106"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_3107 = reshape(shape = var_3106, x = var_3105)[name = string("op_3107")];
            tensor<int32, [4]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3115_pad_type_0 = const()[name = string("op_3115_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3115_strides_0 = const()[name = string("op_3115_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3115_pad_0 = const()[name = string("op_3115_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3115_dilations_0 = const()[name = string("op_3115_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3115_groups_0 = const()[name = string("op_3115_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3115 = conv(dilations = var_3115_dilations_0, groups = var_3115_groups_0, pad = var_3115_pad_0, pad_type = var_3115_pad_type_0, strides = var_3115_strides_0, weight = encoder_layers_15_self_attn_k_proj_weight, x = input_151)[name = string("op_3115")];
            tensor<int32, [4]> var_3116 = const()[name = string("op_3116"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3117 = reshape(shape = var_3116, x = var_3115)[name = string("op_3117")];
            tensor<int32, [4]> var_3118 = const()[name = string("op_3118"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3125_pad_type_0 = const()[name = string("op_3125_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3125_strides_0 = const()[name = string("op_3125_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3125_pad_0 = const()[name = string("op_3125_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3125_dilations_0 = const()[name = string("op_3125_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3125_groups_0 = const()[name = string("op_3125_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3125 = conv(dilations = var_3125_dilations_0, groups = var_3125_groups_0, pad = var_3125_pad_0, pad_type = var_3125_pad_type_0, strides = var_3125_strides_0, weight = encoder_layers_15_self_attn_v_proj_weight, x = input_151)[name = string("op_3125")];
            tensor<int32, [4]> var_3126 = const()[name = string("op_3126"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3127 = reshape(shape = var_3126, x = var_3125)[name = string("op_3127")];
            tensor<int32, [4]> var_3128 = const()[name = string("op_3128"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_61_to_fp16 = const()[name = string("op_6_promoted_61_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_91 = transpose(perm = var_3108, x = var_3107)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 512, 128]> var_3134_cast_fp16 = pow(x = q_91, y = var_6_promoted_61_to_fp16)[name = string("op_3134_cast_fp16")];
            tensor<int32, [1]> var_123_axes_0 = const()[name = string("var_123_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_123_keep_dims_0 = const()[name = string("var_123_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_123_cast_fp16 = reduce_mean(axes = var_123_axes_0, keep_dims = var_123_keep_dims_0, x = var_3134_cast_fp16)[name = string("var_123_cast_fp16")];
            fp16 var_3137_to_fp16 = const()[name = string("op_3137_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_3138_cast_fp16 = add(x = var_123_cast_fp16, y = var_3137_to_fp16)[name = string("op_3138_cast_fp16")];
            fp32 var_3139_epsilon_0 = const()[name = string("op_3139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_3139_cast_fp16 = rsqrt(epsilon = var_3139_epsilon_0, x = var_3138_cast_fp16)[name = string("op_3139_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_521_cast_fp16 = mul(x = q_91, y = var_3139_cast_fp16)[name = string("x_521_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_93 = mul(x = x_521_cast_fp16, y = encoder_layers_15_self_attn_q_norm_weight)[name = string("q_93")];
            fp16 var_6_promoted_62_to_fp16 = const()[name = string("op_6_promoted_62_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_91 = transpose(perm = var_3118, x = var_3117)[name = string("transpose_114")];
            tensor<fp16, [1, 8, 512, 128]> var_3147_cast_fp16 = pow(x = k_91, y = var_6_promoted_62_to_fp16)[name = string("op_3147_cast_fp16")];
            tensor<int32, [1]> var_125_axes_0 = const()[name = string("var_125_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_125_keep_dims_0 = const()[name = string("var_125_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_125_cast_fp16 = reduce_mean(axes = var_125_axes_0, keep_dims = var_125_keep_dims_0, x = var_3147_cast_fp16)[name = string("var_125_cast_fp16")];
            fp16 var_3150_to_fp16 = const()[name = string("op_3150_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_3151_cast_fp16 = add(x = var_125_cast_fp16, y = var_3150_to_fp16)[name = string("op_3151_cast_fp16")];
            fp32 var_3152_epsilon_0 = const()[name = string("op_3152_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_3152_cast_fp16 = rsqrt(epsilon = var_3152_epsilon_0, x = var_3151_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_527_cast_fp16 = mul(x = k_91, y = var_3152_cast_fp16)[name = string("x_527_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_93 = mul(x = x_527_cast_fp16, y = encoder_layers_15_self_attn_k_norm_weight)[name = string("k_93")];
            tensor<fp16, [1, 16, 512, 128]> var_3156 = mul(x = q_93, y = cos)[name = string("op_3156")];
            tensor<int32, [2]> var_3157_split_sizes_0 = const()[name = string("op_3157_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3157_axis_0 = const()[name = string("op_3157_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_3157_0, tensor<fp16, [1, 16, 512, 64]> var_3157_1 = split(axis = var_3157_axis_0, split_sizes = var_3157_split_sizes_0, x = q_93)[name = string("op_3157")];
            fp16 const_48_promoted = const()[name = string("const_48_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_3159 = mul(x = var_3157_1, y = const_48_promoted)[name = string("op_3159")];
            bool var_3161_interleave_0 = const()[name = string("op_3161_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3161 = concat(axis = var_18, interleave = var_3161_interleave_0, values = (var_3159, var_3157_0))[name = string("op_3161")];
            tensor<fp16, [1, 16, 512, 128]> var_3162 = mul(x = var_3161, y = sin)[name = string("op_3162")];
            tensor<fp16, [1, 16, 512, 128]> query_31 = add(x = var_3156, y = var_3162)[name = string("query_31")];
            tensor<fp16, [1, 8, 512, 128]> var_3164 = mul(x = k_93, y = cos)[name = string("op_3164")];
            tensor<int32, [2]> var_3165_split_sizes_0 = const()[name = string("op_3165_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3165_axis_0 = const()[name = string("op_3165_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_3165_0, tensor<fp16, [1, 8, 512, 64]> var_3165_1 = split(axis = var_3165_axis_0, split_sizes = var_3165_split_sizes_0, x = k_93)[name = string("op_3165")];
            fp16 const_49_promoted = const()[name = string("const_49_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_3167 = mul(x = var_3165_1, y = const_49_promoted)[name = string("op_3167")];
            bool var_3169_interleave_0 = const()[name = string("op_3169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_3169 = concat(axis = var_18, interleave = var_3169_interleave_0, values = (var_3167, var_3165_0))[name = string("op_3169")];
            tensor<fp16, [1, 8, 512, 128]> var_3170 = mul(x = var_3169, y = sin)[name = string("op_3170")];
            tensor<fp16, [1, 8, 512, 128]> x_529 = add(x = var_3164, y = var_3170)[name = string("x_529")];
            tensor<int32, [1]> var_3172_axes_0 = const()[name = string("op_3172_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3172 = expand_dims(axes = var_3172_axes_0, x = x_529)[name = string("op_3172")];
            tensor<int32, [5]> x_531_reps_0 = const()[name = string("x_531_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_531 = tile(reps = x_531_reps_0, x = var_3172)[name = string("x_531")];
            tensor<int32, [4]> var_3175 = const()[name = string("op_3175"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_31 = reshape(shape = var_3175, x = x_531)[name = string("key_31")];
            tensor<int32, [1]> var_3177_axes_0 = const()[name = string("op_3177_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_533 = transpose(perm = var_3128, x = var_3127)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3177 = expand_dims(axes = var_3177_axes_0, x = x_533)[name = string("op_3177")];
            tensor<int32, [5]> x_535_reps_0 = const()[name = string("x_535_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_535 = tile(reps = x_535_reps_0, x = var_3177)[name = string("x_535")];
            tensor<int32, [4]> var_3180 = const()[name = string("op_3180"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_31 = reshape(shape = var_3180, x = x_535)[name = string("value_31")];
            bool var_3185_transpose_x_1 = const()[name = string("op_3185_transpose_x_1"), val = bool(false)];
            bool var_3185_transpose_y_1 = const()[name = string("op_3185_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_3185_cast_fp16 = matmul(transpose_x = var_3185_transpose_x_1, transpose_y = var_3185_transpose_y_1, x = query_31, y = key_31)[name = string("op_3185_cast_fp16")];
            fp16 var_3186_to_fp16 = const()[name = string("op_3186_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_91_cast_fp16 = mul(x = var_3185_cast_fp16, y = var_3186_to_fp16)[name = string("attn_weights_91_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_93_cast_fp16 = add(x = attn_weights_91_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_93_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_3190_cast_fp16 = softmax(axis = var_18, x = attn_weights_93_cast_fp16)[name = string("op_3190_cast_fp16")];
            bool var_3194_transpose_x_0 = const()[name = string("op_3194_transpose_x_0"), val = bool(false)];
            bool var_3194_transpose_y_0 = const()[name = string("op_3194_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3194_cast_fp16 = matmul(transpose_x = var_3194_transpose_x_0, transpose_y = var_3194_transpose_y_0, x = var_3190_cast_fp16, y = value_31)[name = string("op_3194_cast_fp16")];
            tensor<int32, [4]> var_3196 = const()[name = string("op_3196"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3197 = transpose(perm = var_3196, x = var_3194_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 512, 2048]> attn_out_93 = reshape(shape = var_3199, x = var_3197)[name = string("attn_out_93")];
            tensor<int32, [3]> var_3201 = const()[name = string("op_3201"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_15 = const()[name = string("squeeze_15"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1154315392)))];
            string var_3210_pad_type_0 = const()[name = string("op_3210_pad_type_0"), val = string("valid")];
            int32 var_3210_groups_0 = const()[name = string("op_3210_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3210_strides_0 = const()[name = string("op_3210_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3210_pad_0 = const()[name = string("op_3210_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3210_dilations_0 = const()[name = string("op_3210_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3202 = transpose(perm = var_3201, x = attn_out_93)[name = string("transpose_111")];
            tensor<fp16, [1, 1024, 512]> var_3210 = conv(dilations = var_3210_dilations_0, groups = var_3210_groups_0, pad = var_3210_pad_0, pad_type = var_3210_pad_type_0, strides = var_3210_strides_0, weight = squeeze_15, x = var_3202)[name = string("op_3210")];
            tensor<int32, [3]> var_3211 = const()[name = string("op_3211"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_95 = transpose(perm = var_3211, x = var_3210)[name = string("transpose_110")];
            tensor<fp16, [1, 512, 1024]> x_537_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = attn_out_95)[name = string("x_537_cast_fp16")];
            fp16 var_6_promoted_63_to_fp16 = const()[name = string("op_6_promoted_63_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3217_cast_fp16 = pow(x = x_537_cast_fp16, y = var_6_promoted_63_to_fp16)[name = string("op_3217_cast_fp16")];
            tensor<int32, [1]> var_127_axes_0 = const()[name = string("var_127_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_127_keep_dims_0 = const()[name = string("var_127_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_127_cast_fp16 = reduce_mean(axes = var_127_axes_0, keep_dims = var_127_keep_dims_0, x = var_3217_cast_fp16)[name = string("var_127_cast_fp16")];
            fp16 var_3220_to_fp16 = const()[name = string("op_3220_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3221_cast_fp16 = add(x = var_127_cast_fp16, y = var_3220_to_fp16)[name = string("op_3221_cast_fp16")];
            fp32 var_3222_epsilon_0 = const()[name = string("op_3222_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3222_cast_fp16 = rsqrt(epsilon = var_3222_epsilon_0, x = var_3221_cast_fp16)[name = string("op_3222_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_541_cast_fp16 = mul(x = x_537_cast_fp16, y = var_3222_cast_fp16)[name = string("x_541_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158509760)))];
            tensor<fp16, [1, 512, 1024]> var_3225_cast_fp16 = mul(x = x_541_cast_fp16, y = encoder_layers_15_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3225_cast_fp16")];
            tensor<int32, [3]> var_3230 = const()[name = string("op_3230"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_155_axes_0 = const()[name = string("input_155_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3231 = transpose(perm = var_3230, x = var_3225_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 1024, 1, 512]> input_155 = expand_dims(axes = input_155_axes_0, x = var_3231)[name = string("input_155")];
            string input_157_pad_type_0 = const()[name = string("input_157_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_157_strides_0 = const()[name = string("input_157_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_157_pad_0 = const()[name = string("input_157_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_157_dilations_0 = const()[name = string("input_157_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_157_groups_0 = const()[name = string("input_157_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_157 = conv(dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = encoder_layers_15_mlp_gate_proj_weight, x = input_155)[name = string("input_157")];
            string up_31_pad_type_0 = const()[name = string("up_31_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_31_strides_0 = const()[name = string("up_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_31_pad_0 = const()[name = string("up_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_31_dilations_0 = const()[name = string("up_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_31_groups_0 = const()[name = string("up_31_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_31 = conv(dilations = up_31_dilations_0, groups = up_31_groups_0, pad = up_31_pad_0, pad_type = up_31_pad_type_0, strides = up_31_strides_0, weight = encoder_layers_15_mlp_up_proj_weight, x = input_155)[name = string("up_31")];
            tensor<fp16, [1, 3072, 1, 512]> var_3245 = silu(x = input_157)[name = string("op_3245")];
            tensor<fp16, [1, 3072, 1, 512]> input_159 = mul(x = var_3245, y = up_31)[name = string("input_159")];
            string var_3252_pad_type_0 = const()[name = string("op_3252_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3252_strides_0 = const()[name = string("op_3252_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3252_pad_0 = const()[name = string("op_3252_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3252_dilations_0 = const()[name = string("op_3252_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3252_groups_0 = const()[name = string("op_3252_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3252 = conv(dilations = var_3252_dilations_0, groups = var_3252_groups_0, pad = var_3252_pad_0, pad_type = var_3252_pad_type_0, strides = var_3252_strides_0, weight = encoder_layers_15_mlp_down_proj_weight, x = input_159)[name = string("op_3252")];
            tensor<int32, [1]> var_3253_axes_0 = const()[name = string("op_3253_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3253 = squeeze(axes = var_3253_axes_0, x = var_3252)[name = string("op_3253")];
            tensor<int32, [3]> var_3254 = const()[name = string("op_3254"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_31 = transpose(perm = var_3254, x = var_3253)[name = string("transpose_108")];
            tensor<fp16, [1, 512, 1024]> hidden_states_33_cast_fp16 = add(x = x_537_cast_fp16, y = mlp_out_31)[name = string("hidden_states_33_cast_fp16")];
            fp16 var_6_promoted_64_to_fp16 = const()[name = string("op_6_promoted_64_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3281_cast_fp16 = pow(x = hidden_states_33_cast_fp16, y = var_6_promoted_64_to_fp16)[name = string("op_3281_cast_fp16")];
            tensor<int32, [1]> var_129_axes_0 = const()[name = string("var_129_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_129_keep_dims_0 = const()[name = string("var_129_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_129_cast_fp16_0 = reduce_mean(axes = var_129_axes_0, keep_dims = var_129_keep_dims_0, x = var_3281_cast_fp16)[name = string("var_129_cast_fp16")];
            fp16 var_3284_to_fp16 = const()[name = string("op_3284_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3285_cast_fp16 = add(x = var_129_cast_fp16_0, y = var_3284_to_fp16)[name = string("op_3285_cast_fp16")];
            fp32 var_3286_epsilon_0 = const()[name = string("op_3286_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3286_cast_fp16 = rsqrt(epsilon = var_3286_epsilon_0, x = var_3285_cast_fp16)[name = string("op_3286_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_547_cast_fp16 = mul(x = hidden_states_33_cast_fp16, y = var_3286_cast_fp16)[name = string("x_547_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_16_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_16_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158511872)))];
            tensor<fp16, [1, 512, 1024]> var_3289_cast_fp16 = mul(x = x_547_cast_fp16, y = encoder_layers_16_input_layernorm_weight_promoted_to_fp16)[name = string("op_3289_cast_fp16")];
            tensor<int32, [3]> var_3294 = const()[name = string("op_3294"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3295 = transpose(perm = var_3294, x = var_3289_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 1024, 1, 512]> input_161 = expand_dims(axes = input_161_axes_0, x = var_3295)[name = string("input_161")];
            string var_3302_pad_type_0 = const()[name = string("op_3302_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3302_strides_0 = const()[name = string("op_3302_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3302_pad_0 = const()[name = string("op_3302_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3302_dilations_0 = const()[name = string("op_3302_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3302_groups_0 = const()[name = string("op_3302_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_3302 = conv(dilations = var_3302_dilations_0, groups = var_3302_groups_0, pad = var_3302_pad_0, pad_type = var_3302_pad_type_0, strides = var_3302_strides_0, weight = encoder_layers_16_self_attn_q_proj_weight, x = input_161)[name = string("op_3302")];
            tensor<int32, [4]> var_3303 = const()[name = string("op_3303"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_3304 = reshape(shape = var_3303, x = var_3302)[name = string("op_3304")];
            tensor<int32, [4]> var_3305 = const()[name = string("op_3305"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3312_pad_type_0 = const()[name = string("op_3312_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3312_strides_0 = const()[name = string("op_3312_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3312_pad_0 = const()[name = string("op_3312_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3312_dilations_0 = const()[name = string("op_3312_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3312_groups_0 = const()[name = string("op_3312_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3312 = conv(dilations = var_3312_dilations_0, groups = var_3312_groups_0, pad = var_3312_pad_0, pad_type = var_3312_pad_type_0, strides = var_3312_strides_0, weight = encoder_layers_16_self_attn_k_proj_weight, x = input_161)[name = string("op_3312")];
            tensor<int32, [4]> var_3313 = const()[name = string("op_3313"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3314 = reshape(shape = var_3313, x = var_3312)[name = string("op_3314")];
            tensor<int32, [4]> var_3315 = const()[name = string("op_3315"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3322_pad_type_0 = const()[name = string("op_3322_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3322_strides_0 = const()[name = string("op_3322_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3322_pad_0 = const()[name = string("op_3322_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3322_dilations_0 = const()[name = string("op_3322_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3322_groups_0 = const()[name = string("op_3322_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3322 = conv(dilations = var_3322_dilations_0, groups = var_3322_groups_0, pad = var_3322_pad_0, pad_type = var_3322_pad_type_0, strides = var_3322_strides_0, weight = encoder_layers_16_self_attn_v_proj_weight, x = input_161)[name = string("op_3322")];
            tensor<int32, [4]> var_3323 = const()[name = string("op_3323"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3324 = reshape(shape = var_3323, x = var_3322)[name = string("op_3324")];
            tensor<int32, [4]> var_3325 = const()[name = string("op_3325"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_65_to_fp16 = const()[name = string("op_6_promoted_65_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_97 = transpose(perm = var_3305, x = var_3304)[name = string("transpose_106")];
            tensor<fp16, [1, 16, 512, 128]> var_3331_cast_fp16 = pow(x = q_97, y = var_6_promoted_65_to_fp16)[name = string("op_3331_cast_fp16")];
            tensor<int32, [1]> var_131_axes_0 = const()[name = string("var_131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_131_keep_dims_0 = const()[name = string("var_131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_131_cast_fp16 = reduce_mean(axes = var_131_axes_0, keep_dims = var_131_keep_dims_0, x = var_3331_cast_fp16)[name = string("var_131_cast_fp16")];
            fp16 var_3334_to_fp16 = const()[name = string("op_3334_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_3335_cast_fp16 = add(x = var_131_cast_fp16, y = var_3334_to_fp16)[name = string("op_3335_cast_fp16")];
            fp32 var_3336_epsilon_0 = const()[name = string("op_3336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_3336_cast_fp16 = rsqrt(epsilon = var_3336_epsilon_0, x = var_3335_cast_fp16)[name = string("op_3336_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_555_cast_fp16 = mul(x = q_97, y = var_3336_cast_fp16)[name = string("x_555_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_99 = mul(x = x_555_cast_fp16, y = encoder_layers_16_self_attn_q_norm_weight)[name = string("q_99")];
            fp16 var_6_promoted_66_to_fp16 = const()[name = string("op_6_promoted_66_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_97 = transpose(perm = var_3315, x = var_3314)[name = string("transpose_105")];
            tensor<fp16, [1, 8, 512, 128]> var_3344_cast_fp16 = pow(x = k_97, y = var_6_promoted_66_to_fp16)[name = string("op_3344_cast_fp16")];
            tensor<int32, [1]> var_133_axes_0 = const()[name = string("var_133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_133_keep_dims_0 = const()[name = string("var_133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_133_cast_fp16_0 = reduce_mean(axes = var_133_axes_0, keep_dims = var_133_keep_dims_0, x = var_3344_cast_fp16)[name = string("var_133_cast_fp16")];
            fp16 var_3347_to_fp16 = const()[name = string("op_3347_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_3348_cast_fp16 = add(x = var_133_cast_fp16_0, y = var_3347_to_fp16)[name = string("op_3348_cast_fp16")];
            fp32 var_3349_epsilon_0 = const()[name = string("op_3349_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_3349_cast_fp16 = rsqrt(epsilon = var_3349_epsilon_0, x = var_3348_cast_fp16)[name = string("op_3349_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_561_cast_fp16 = mul(x = k_97, y = var_3349_cast_fp16)[name = string("x_561_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_99 = mul(x = x_561_cast_fp16, y = encoder_layers_16_self_attn_k_norm_weight)[name = string("k_99")];
            tensor<fp16, [1, 16, 512, 128]> var_3353 = mul(x = q_99, y = cos)[name = string("op_3353")];
            tensor<int32, [2]> var_3354_split_sizes_0 = const()[name = string("op_3354_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3354_axis_0 = const()[name = string("op_3354_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_3354_0, tensor<fp16, [1, 16, 512, 64]> var_3354_1 = split(axis = var_3354_axis_0, split_sizes = var_3354_split_sizes_0, x = q_99)[name = string("op_3354")];
            fp16 const_51_promoted = const()[name = string("const_51_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_3356 = mul(x = var_3354_1, y = const_51_promoted)[name = string("op_3356")];
            bool var_3358_interleave_0 = const()[name = string("op_3358_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3358 = concat(axis = var_18, interleave = var_3358_interleave_0, values = (var_3356, var_3354_0))[name = string("op_3358")];
            tensor<fp16, [1, 16, 512, 128]> var_3359 = mul(x = var_3358, y = sin)[name = string("op_3359")];
            tensor<fp16, [1, 16, 512, 128]> query_33 = add(x = var_3353, y = var_3359)[name = string("query_33")];
            tensor<fp16, [1, 8, 512, 128]> var_3361 = mul(x = k_99, y = cos)[name = string("op_3361")];
            tensor<int32, [2]> var_3362_split_sizes_0 = const()[name = string("op_3362_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3362_axis_0 = const()[name = string("op_3362_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_3362_0, tensor<fp16, [1, 8, 512, 64]> var_3362_1 = split(axis = var_3362_axis_0, split_sizes = var_3362_split_sizes_0, x = k_99)[name = string("op_3362")];
            fp16 const_52_promoted = const()[name = string("const_52_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_3364 = mul(x = var_3362_1, y = const_52_promoted)[name = string("op_3364")];
            bool var_3366_interleave_0 = const()[name = string("op_3366_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_3366 = concat(axis = var_18, interleave = var_3366_interleave_0, values = (var_3364, var_3362_0))[name = string("op_3366")];
            tensor<fp16, [1, 8, 512, 128]> var_3367 = mul(x = var_3366, y = sin)[name = string("op_3367")];
            tensor<fp16, [1, 8, 512, 128]> x_563 = add(x = var_3361, y = var_3367)[name = string("x_563")];
            tensor<int32, [1]> var_3369_axes_0 = const()[name = string("op_3369_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3369 = expand_dims(axes = var_3369_axes_0, x = x_563)[name = string("op_3369")];
            tensor<int32, [5]> x_565_reps_0 = const()[name = string("x_565_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_565 = tile(reps = x_565_reps_0, x = var_3369)[name = string("x_565")];
            tensor<int32, [4]> var_3372 = const()[name = string("op_3372"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_33 = reshape(shape = var_3372, x = x_565)[name = string("key_33")];
            tensor<int32, [1]> var_3374_axes_0 = const()[name = string("op_3374_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_567 = transpose(perm = var_3325, x = var_3324)[name = string("transpose_104")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3374 = expand_dims(axes = var_3374_axes_0, x = x_567)[name = string("op_3374")];
            tensor<int32, [5]> x_569_reps_0 = const()[name = string("x_569_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_569 = tile(reps = x_569_reps_0, x = var_3374)[name = string("x_569")];
            tensor<int32, [4]> var_3377 = const()[name = string("op_3377"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_33 = reshape(shape = var_3377, x = x_569)[name = string("value_33")];
            bool var_3382_transpose_x_1 = const()[name = string("op_3382_transpose_x_1"), val = bool(false)];
            bool var_3382_transpose_y_1 = const()[name = string("op_3382_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_3382_cast_fp16 = matmul(transpose_x = var_3382_transpose_x_1, transpose_y = var_3382_transpose_y_1, x = query_33, y = key_33)[name = string("op_3382_cast_fp16")];
            fp16 var_3383_to_fp16 = const()[name = string("op_3383_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_97_cast_fp16 = mul(x = var_3382_cast_fp16, y = var_3383_to_fp16)[name = string("attn_weights_97_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_99_cast_fp16 = add(x = attn_weights_97_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_99_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_3387_cast_fp16 = softmax(axis = var_18, x = attn_weights_99_cast_fp16)[name = string("op_3387_cast_fp16")];
            bool var_3391_transpose_x_0 = const()[name = string("op_3391_transpose_x_0"), val = bool(false)];
            bool var_3391_transpose_y_0 = const()[name = string("op_3391_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3391_cast_fp16 = matmul(transpose_x = var_3391_transpose_x_0, transpose_y = var_3391_transpose_y_0, x = var_3387_cast_fp16, y = value_33)[name = string("op_3391_cast_fp16")];
            tensor<int32, [4]> var_3393 = const()[name = string("op_3393"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3396 = const()[name = string("op_3396"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3394 = transpose(perm = var_3393, x = var_3391_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 512, 2048]> attn_out_99 = reshape(shape = var_3396, x = var_3394)[name = string("attn_out_99")];
            tensor<int32, [3]> var_3398 = const()[name = string("op_3398"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_16 = const()[name = string("squeeze_16"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1158513984)))];
            string var_3407_pad_type_0 = const()[name = string("op_3407_pad_type_0"), val = string("valid")];
            int32 var_3407_groups_0 = const()[name = string("op_3407_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3407_strides_0 = const()[name = string("op_3407_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3407_pad_0 = const()[name = string("op_3407_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3407_dilations_0 = const()[name = string("op_3407_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3399 = transpose(perm = var_3398, x = attn_out_99)[name = string("transpose_102")];
            tensor<fp16, [1, 1024, 512]> var_3407 = conv(dilations = var_3407_dilations_0, groups = var_3407_groups_0, pad = var_3407_pad_0, pad_type = var_3407_pad_type_0, strides = var_3407_strides_0, weight = squeeze_16, x = var_3399)[name = string("op_3407")];
            tensor<int32, [3]> var_3408 = const()[name = string("op_3408"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_101 = transpose(perm = var_3408, x = var_3407)[name = string("transpose_101")];
            tensor<fp16, [1, 512, 1024]> x_571_cast_fp16 = add(x = hidden_states_33_cast_fp16, y = attn_out_101)[name = string("x_571_cast_fp16")];
            fp16 var_6_promoted_67_to_fp16 = const()[name = string("op_6_promoted_67_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3414_cast_fp16 = pow(x = x_571_cast_fp16, y = var_6_promoted_67_to_fp16)[name = string("op_3414_cast_fp16")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = string("var_135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_135_keep_dims_0 = const()[name = string("var_135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_135_cast_fp16 = reduce_mean(axes = var_135_axes_0, keep_dims = var_135_keep_dims_0, x = var_3414_cast_fp16)[name = string("var_135_cast_fp16")];
            fp16 var_3417_to_fp16 = const()[name = string("op_3417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3418_cast_fp16 = add(x = var_135_cast_fp16, y = var_3417_to_fp16)[name = string("op_3418_cast_fp16")];
            fp32 var_3419_epsilon_0 = const()[name = string("op_3419_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3419_cast_fp16 = rsqrt(epsilon = var_3419_epsilon_0, x = var_3418_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_575_cast_fp16 = mul(x = x_571_cast_fp16, y = var_3419_cast_fp16)[name = string("x_575_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162708352)))];
            tensor<fp16, [1, 512, 1024]> var_3422_cast_fp16 = mul(x = x_575_cast_fp16, y = encoder_layers_16_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3422_cast_fp16")];
            tensor<int32, [3]> var_3427 = const()[name = string("op_3427"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_165_axes_0 = const()[name = string("input_165_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3428 = transpose(perm = var_3427, x = var_3422_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 1024, 1, 512]> input_165 = expand_dims(axes = input_165_axes_0, x = var_3428)[name = string("input_165")];
            string input_167_pad_type_0 = const()[name = string("input_167_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_167_strides_0 = const()[name = string("input_167_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_167_pad_0 = const()[name = string("input_167_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_167_dilations_0 = const()[name = string("input_167_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_167_groups_0 = const()[name = string("input_167_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_167 = conv(dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = encoder_layers_16_mlp_gate_proj_weight, x = input_165)[name = string("input_167")];
            string up_33_pad_type_0 = const()[name = string("up_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_33_strides_0 = const()[name = string("up_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_33_pad_0 = const()[name = string("up_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_33_dilations_0 = const()[name = string("up_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_33_groups_0 = const()[name = string("up_33_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_33 = conv(dilations = up_33_dilations_0, groups = up_33_groups_0, pad = up_33_pad_0, pad_type = up_33_pad_type_0, strides = up_33_strides_0, weight = encoder_layers_16_mlp_up_proj_weight, x = input_165)[name = string("up_33")];
            tensor<fp16, [1, 3072, 1, 512]> var_3442 = silu(x = input_167)[name = string("op_3442")];
            tensor<fp16, [1, 3072, 1, 512]> input_169 = mul(x = var_3442, y = up_33)[name = string("input_169")];
            string var_3449_pad_type_0 = const()[name = string("op_3449_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3449_strides_0 = const()[name = string("op_3449_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3449_pad_0 = const()[name = string("op_3449_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3449_dilations_0 = const()[name = string("op_3449_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3449_groups_0 = const()[name = string("op_3449_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3449 = conv(dilations = var_3449_dilations_0, groups = var_3449_groups_0, pad = var_3449_pad_0, pad_type = var_3449_pad_type_0, strides = var_3449_strides_0, weight = encoder_layers_16_mlp_down_proj_weight, x = input_169)[name = string("op_3449")];
            tensor<int32, [1]> var_3450_axes_0 = const()[name = string("op_3450_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3450 = squeeze(axes = var_3450_axes_0, x = var_3449)[name = string("op_3450")];
            tensor<int32, [3]> var_3451 = const()[name = string("op_3451"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_33 = transpose(perm = var_3451, x = var_3450)[name = string("transpose_99")];
            tensor<fp16, [1, 512, 1024]> hidden_states_35_cast_fp16 = add(x = x_571_cast_fp16, y = mlp_out_33)[name = string("hidden_states_35_cast_fp16")];
            fp16 var_6_promoted_68_to_fp16 = const()[name = string("op_6_promoted_68_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3478_cast_fp16 = pow(x = hidden_states_35_cast_fp16, y = var_6_promoted_68_to_fp16)[name = string("op_3478_cast_fp16")];
            tensor<int32, [1]> var_137_axes_0 = const()[name = string("var_137_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_137_keep_dims_0 = const()[name = string("var_137_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_137_cast_fp16_0 = reduce_mean(axes = var_137_axes_0, keep_dims = var_137_keep_dims_0, x = var_3478_cast_fp16)[name = string("var_137_cast_fp16")];
            fp16 var_3481_to_fp16 = const()[name = string("op_3481_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3482_cast_fp16 = add(x = var_137_cast_fp16_0, y = var_3481_to_fp16)[name = string("op_3482_cast_fp16")];
            fp32 var_3483_epsilon_0 = const()[name = string("op_3483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3483_cast_fp16 = rsqrt(epsilon = var_3483_epsilon_0, x = var_3482_cast_fp16)[name = string("op_3483_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_581_cast_fp16 = mul(x = hidden_states_35_cast_fp16, y = var_3483_cast_fp16)[name = string("x_581_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_17_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_17_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162710464)))];
            tensor<fp16, [1, 512, 1024]> var_3486_cast_fp16 = mul(x = x_581_cast_fp16, y = encoder_layers_17_input_layernorm_weight_promoted_to_fp16)[name = string("op_3486_cast_fp16")];
            tensor<int32, [3]> var_3491 = const()[name = string("op_3491"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_171_axes_0 = const()[name = string("input_171_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3492 = transpose(perm = var_3491, x = var_3486_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 1024, 1, 512]> input_171 = expand_dims(axes = input_171_axes_0, x = var_3492)[name = string("input_171")];
            string var_3499_pad_type_0 = const()[name = string("op_3499_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3499_strides_0 = const()[name = string("op_3499_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3499_pad_0 = const()[name = string("op_3499_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3499_dilations_0 = const()[name = string("op_3499_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3499_groups_0 = const()[name = string("op_3499_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_3499 = conv(dilations = var_3499_dilations_0, groups = var_3499_groups_0, pad = var_3499_pad_0, pad_type = var_3499_pad_type_0, strides = var_3499_strides_0, weight = encoder_layers_17_self_attn_q_proj_weight, x = input_171)[name = string("op_3499")];
            tensor<int32, [4]> var_3500 = const()[name = string("op_3500"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_3501 = reshape(shape = var_3500, x = var_3499)[name = string("op_3501")];
            tensor<int32, [4]> var_3502 = const()[name = string("op_3502"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3509_pad_type_0 = const()[name = string("op_3509_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3509_strides_0 = const()[name = string("op_3509_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3509_pad_0 = const()[name = string("op_3509_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3509_dilations_0 = const()[name = string("op_3509_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3509_groups_0 = const()[name = string("op_3509_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3509 = conv(dilations = var_3509_dilations_0, groups = var_3509_groups_0, pad = var_3509_pad_0, pad_type = var_3509_pad_type_0, strides = var_3509_strides_0, weight = encoder_layers_17_self_attn_k_proj_weight, x = input_171)[name = string("op_3509")];
            tensor<int32, [4]> var_3510 = const()[name = string("op_3510"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3511 = reshape(shape = var_3510, x = var_3509)[name = string("op_3511")];
            tensor<int32, [4]> var_3512 = const()[name = string("op_3512"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3519_pad_type_0 = const()[name = string("op_3519_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3519_strides_0 = const()[name = string("op_3519_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3519_pad_0 = const()[name = string("op_3519_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3519_dilations_0 = const()[name = string("op_3519_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3519_groups_0 = const()[name = string("op_3519_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3519 = conv(dilations = var_3519_dilations_0, groups = var_3519_groups_0, pad = var_3519_pad_0, pad_type = var_3519_pad_type_0, strides = var_3519_strides_0, weight = encoder_layers_17_self_attn_v_proj_weight, x = input_171)[name = string("op_3519")];
            tensor<int32, [4]> var_3520 = const()[name = string("op_3520"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3521 = reshape(shape = var_3520, x = var_3519)[name = string("op_3521")];
            tensor<int32, [4]> var_3522 = const()[name = string("op_3522"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_69_to_fp16 = const()[name = string("op_6_promoted_69_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_103 = transpose(perm = var_3502, x = var_3501)[name = string("transpose_97")];
            tensor<fp16, [1, 16, 512, 128]> var_3528_cast_fp16 = pow(x = q_103, y = var_6_promoted_69_to_fp16)[name = string("op_3528_cast_fp16")];
            tensor<int32, [1]> var_139_axes_0 = const()[name = string("var_139_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_139_keep_dims_0 = const()[name = string("var_139_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_139_cast_fp16 = reduce_mean(axes = var_139_axes_0, keep_dims = var_139_keep_dims_0, x = var_3528_cast_fp16)[name = string("var_139_cast_fp16")];
            fp16 var_3531_to_fp16 = const()[name = string("op_3531_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_3532_cast_fp16 = add(x = var_139_cast_fp16, y = var_3531_to_fp16)[name = string("op_3532_cast_fp16")];
            fp32 var_3533_epsilon_0 = const()[name = string("op_3533_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_3533_cast_fp16 = rsqrt(epsilon = var_3533_epsilon_0, x = var_3532_cast_fp16)[name = string("op_3533_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_589_cast_fp16 = mul(x = q_103, y = var_3533_cast_fp16)[name = string("x_589_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_105 = mul(x = x_589_cast_fp16, y = encoder_layers_17_self_attn_q_norm_weight)[name = string("q_105")];
            fp16 var_6_promoted_70_to_fp16 = const()[name = string("op_6_promoted_70_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_103 = transpose(perm = var_3512, x = var_3511)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 512, 128]> var_3541_cast_fp16 = pow(x = k_103, y = var_6_promoted_70_to_fp16)[name = string("op_3541_cast_fp16")];
            tensor<int32, [1]> var_141_axes_0 = const()[name = string("var_141_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_141_keep_dims_0 = const()[name = string("var_141_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_141_cast_fp16 = reduce_mean(axes = var_141_axes_0, keep_dims = var_141_keep_dims_0, x = var_3541_cast_fp16)[name = string("var_141_cast_fp16")];
            fp16 var_3544_to_fp16 = const()[name = string("op_3544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_3545_cast_fp16 = add(x = var_141_cast_fp16, y = var_3544_to_fp16)[name = string("op_3545_cast_fp16")];
            fp32 var_3546_epsilon_0 = const()[name = string("op_3546_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_3546_cast_fp16 = rsqrt(epsilon = var_3546_epsilon_0, x = var_3545_cast_fp16)[name = string("op_3546_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_595_cast_fp16 = mul(x = k_103, y = var_3546_cast_fp16)[name = string("x_595_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_105 = mul(x = x_595_cast_fp16, y = encoder_layers_17_self_attn_k_norm_weight)[name = string("k_105")];
            tensor<fp16, [1, 16, 512, 128]> var_3550 = mul(x = q_105, y = cos)[name = string("op_3550")];
            tensor<int32, [2]> var_3551_split_sizes_0 = const()[name = string("op_3551_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3551_axis_0 = const()[name = string("op_3551_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_3551_0, tensor<fp16, [1, 16, 512, 64]> var_3551_1 = split(axis = var_3551_axis_0, split_sizes = var_3551_split_sizes_0, x = q_105)[name = string("op_3551")];
            fp16 const_54_promoted = const()[name = string("const_54_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_3553 = mul(x = var_3551_1, y = const_54_promoted)[name = string("op_3553")];
            bool var_3555_interleave_0 = const()[name = string("op_3555_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3555 = concat(axis = var_18, interleave = var_3555_interleave_0, values = (var_3553, var_3551_0))[name = string("op_3555")];
            tensor<fp16, [1, 16, 512, 128]> var_3556 = mul(x = var_3555, y = sin)[name = string("op_3556")];
            tensor<fp16, [1, 16, 512, 128]> query_35 = add(x = var_3550, y = var_3556)[name = string("query_35")];
            tensor<fp16, [1, 8, 512, 128]> var_3558 = mul(x = k_105, y = cos)[name = string("op_3558")];
            tensor<int32, [2]> var_3559_split_sizes_0 = const()[name = string("op_3559_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3559_axis_0 = const()[name = string("op_3559_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_3559_0, tensor<fp16, [1, 8, 512, 64]> var_3559_1 = split(axis = var_3559_axis_0, split_sizes = var_3559_split_sizes_0, x = k_105)[name = string("op_3559")];
            fp16 const_55_promoted = const()[name = string("const_55_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_3561 = mul(x = var_3559_1, y = const_55_promoted)[name = string("op_3561")];
            bool var_3563_interleave_0 = const()[name = string("op_3563_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_3563 = concat(axis = var_18, interleave = var_3563_interleave_0, values = (var_3561, var_3559_0))[name = string("op_3563")];
            tensor<fp16, [1, 8, 512, 128]> var_3564 = mul(x = var_3563, y = sin)[name = string("op_3564")];
            tensor<fp16, [1, 8, 512, 128]> x_597 = add(x = var_3558, y = var_3564)[name = string("x_597")];
            tensor<int32, [1]> var_3566_axes_0 = const()[name = string("op_3566_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3566 = expand_dims(axes = var_3566_axes_0, x = x_597)[name = string("op_3566")];
            tensor<int32, [5]> x_599_reps_0 = const()[name = string("x_599_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_599 = tile(reps = x_599_reps_0, x = var_3566)[name = string("x_599")];
            tensor<int32, [4]> var_3569 = const()[name = string("op_3569"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_35 = reshape(shape = var_3569, x = x_599)[name = string("key_35")];
            tensor<int32, [1]> var_3571_axes_0 = const()[name = string("op_3571_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_601 = transpose(perm = var_3522, x = var_3521)[name = string("transpose_95")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3571 = expand_dims(axes = var_3571_axes_0, x = x_601)[name = string("op_3571")];
            tensor<int32, [5]> x_603_reps_0 = const()[name = string("x_603_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_603 = tile(reps = x_603_reps_0, x = var_3571)[name = string("x_603")];
            tensor<int32, [4]> var_3574 = const()[name = string("op_3574"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_35 = reshape(shape = var_3574, x = x_603)[name = string("value_35")];
            bool var_3579_transpose_x_1 = const()[name = string("op_3579_transpose_x_1"), val = bool(false)];
            bool var_3579_transpose_y_1 = const()[name = string("op_3579_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_3579_cast_fp16 = matmul(transpose_x = var_3579_transpose_x_1, transpose_y = var_3579_transpose_y_1, x = query_35, y = key_35)[name = string("op_3579_cast_fp16")];
            fp16 var_3580_to_fp16 = const()[name = string("op_3580_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_103_cast_fp16 = mul(x = var_3579_cast_fp16, y = var_3580_to_fp16)[name = string("attn_weights_103_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_105_cast_fp16 = add(x = attn_weights_103_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_105_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_3584_cast_fp16 = softmax(axis = var_18, x = attn_weights_105_cast_fp16)[name = string("op_3584_cast_fp16")];
            bool var_3588_transpose_x_0 = const()[name = string("op_3588_transpose_x_0"), val = bool(false)];
            bool var_3588_transpose_y_0 = const()[name = string("op_3588_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3588_cast_fp16 = matmul(transpose_x = var_3588_transpose_x_0, transpose_y = var_3588_transpose_y_0, x = var_3584_cast_fp16, y = value_35)[name = string("op_3588_cast_fp16")];
            tensor<int32, [4]> var_3590 = const()[name = string("op_3590"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3593 = const()[name = string("op_3593"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3591 = transpose(perm = var_3590, x = var_3588_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 512, 2048]> attn_out_105 = reshape(shape = var_3593, x = var_3591)[name = string("attn_out_105")];
            tensor<int32, [3]> var_3595 = const()[name = string("op_3595"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_17 = const()[name = string("squeeze_17"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1162712576)))];
            string var_3604_pad_type_0 = const()[name = string("op_3604_pad_type_0"), val = string("valid")];
            int32 var_3604_groups_0 = const()[name = string("op_3604_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3604_strides_0 = const()[name = string("op_3604_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3604_pad_0 = const()[name = string("op_3604_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3604_dilations_0 = const()[name = string("op_3604_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3596 = transpose(perm = var_3595, x = attn_out_105)[name = string("transpose_93")];
            tensor<fp16, [1, 1024, 512]> var_3604 = conv(dilations = var_3604_dilations_0, groups = var_3604_groups_0, pad = var_3604_pad_0, pad_type = var_3604_pad_type_0, strides = var_3604_strides_0, weight = squeeze_17, x = var_3596)[name = string("op_3604")];
            tensor<int32, [3]> var_3605 = const()[name = string("op_3605"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_107 = transpose(perm = var_3605, x = var_3604)[name = string("transpose_92")];
            tensor<fp16, [1, 512, 1024]> x_605_cast_fp16 = add(x = hidden_states_35_cast_fp16, y = attn_out_107)[name = string("x_605_cast_fp16")];
            fp16 var_6_promoted_71_to_fp16 = const()[name = string("op_6_promoted_71_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3611_cast_fp16 = pow(x = x_605_cast_fp16, y = var_6_promoted_71_to_fp16)[name = string("op_3611_cast_fp16")];
            tensor<int32, [1]> var_143_axes_0 = const()[name = string("var_143_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_143_keep_dims_0 = const()[name = string("var_143_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_143_cast_fp16 = reduce_mean(axes = var_143_axes_0, keep_dims = var_143_keep_dims_0, x = var_3611_cast_fp16)[name = string("var_143_cast_fp16")];
            fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3615_cast_fp16 = add(x = var_143_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")];
            fp32 var_3616_epsilon_0 = const()[name = string("op_3616_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3616_cast_fp16 = rsqrt(epsilon = var_3616_epsilon_0, x = var_3615_cast_fp16)[name = string("op_3616_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_609_cast_fp16 = mul(x = x_605_cast_fp16, y = var_3616_cast_fp16)[name = string("x_609_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166906944)))];
            tensor<fp16, [1, 512, 1024]> var_3619_cast_fp16 = mul(x = x_609_cast_fp16, y = encoder_layers_17_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3619_cast_fp16")];
            tensor<int32, [3]> var_3624 = const()[name = string("op_3624"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3625 = transpose(perm = var_3624, x = var_3619_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 1024, 1, 512]> input_175 = expand_dims(axes = input_175_axes_0, x = var_3625)[name = string("input_175")];
            string input_177_pad_type_0 = const()[name = string("input_177_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_177_strides_0 = const()[name = string("input_177_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_177_pad_0 = const()[name = string("input_177_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_177_dilations_0 = const()[name = string("input_177_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_177_groups_0 = const()[name = string("input_177_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_177 = conv(dilations = input_177_dilations_0, groups = input_177_groups_0, pad = input_177_pad_0, pad_type = input_177_pad_type_0, strides = input_177_strides_0, weight = encoder_layers_17_mlp_gate_proj_weight, x = input_175)[name = string("input_177")];
            string up_35_pad_type_0 = const()[name = string("up_35_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_35_strides_0 = const()[name = string("up_35_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_35_pad_0 = const()[name = string("up_35_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_35_dilations_0 = const()[name = string("up_35_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_35_groups_0 = const()[name = string("up_35_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_35 = conv(dilations = up_35_dilations_0, groups = up_35_groups_0, pad = up_35_pad_0, pad_type = up_35_pad_type_0, strides = up_35_strides_0, weight = encoder_layers_17_mlp_up_proj_weight, x = input_175)[name = string("up_35")];
            tensor<fp16, [1, 3072, 1, 512]> var_3639 = silu(x = input_177)[name = string("op_3639")];
            tensor<fp16, [1, 3072, 1, 512]> input_179 = mul(x = var_3639, y = up_35)[name = string("input_179")];
            string var_3646_pad_type_0 = const()[name = string("op_3646_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3646_strides_0 = const()[name = string("op_3646_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3646_pad_0 = const()[name = string("op_3646_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3646_dilations_0 = const()[name = string("op_3646_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3646_groups_0 = const()[name = string("op_3646_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3646 = conv(dilations = var_3646_dilations_0, groups = var_3646_groups_0, pad = var_3646_pad_0, pad_type = var_3646_pad_type_0, strides = var_3646_strides_0, weight = encoder_layers_17_mlp_down_proj_weight, x = input_179)[name = string("op_3646")];
            tensor<int32, [1]> var_3647_axes_0 = const()[name = string("op_3647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3647 = squeeze(axes = var_3647_axes_0, x = var_3646)[name = string("op_3647")];
            tensor<int32, [3]> var_3648 = const()[name = string("op_3648"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_35 = transpose(perm = var_3648, x = var_3647)[name = string("transpose_90")];
            tensor<fp16, [1, 512, 1024]> hidden_states_37_cast_fp16 = add(x = x_605_cast_fp16, y = mlp_out_35)[name = string("hidden_states_37_cast_fp16")];
            fp16 var_6_promoted_72_to_fp16 = const()[name = string("op_6_promoted_72_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3675_cast_fp16 = pow(x = hidden_states_37_cast_fp16, y = var_6_promoted_72_to_fp16)[name = string("op_3675_cast_fp16")];
            tensor<int32, [1]> var_145_axes_0 = const()[name = string("var_145_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_145_keep_dims_0 = const()[name = string("var_145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_145_cast_fp16 = reduce_mean(axes = var_145_axes_0, keep_dims = var_145_keep_dims_0, x = var_3675_cast_fp16)[name = string("var_145_cast_fp16")];
            fp16 var_3678_to_fp16 = const()[name = string("op_3678_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3679_cast_fp16 = add(x = var_145_cast_fp16, y = var_3678_to_fp16)[name = string("op_3679_cast_fp16")];
            fp32 var_3680_epsilon_0 = const()[name = string("op_3680_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3680_cast_fp16 = rsqrt(epsilon = var_3680_epsilon_0, x = var_3679_cast_fp16)[name = string("op_3680_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_615_cast_fp16 = mul(x = hidden_states_37_cast_fp16, y = var_3680_cast_fp16)[name = string("x_615_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_18_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_18_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166909056)))];
            tensor<fp16, [1, 512, 1024]> var_3683_cast_fp16 = mul(x = x_615_cast_fp16, y = encoder_layers_18_input_layernorm_weight_promoted_to_fp16)[name = string("op_3683_cast_fp16")];
            tensor<int32, [3]> var_3688 = const()[name = string("op_3688"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3689 = transpose(perm = var_3688, x = var_3683_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 1024, 1, 512]> input_181 = expand_dims(axes = input_181_axes_0, x = var_3689)[name = string("input_181")];
            string var_3696_pad_type_0 = const()[name = string("op_3696_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3696_strides_0 = const()[name = string("op_3696_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3696_pad_0 = const()[name = string("op_3696_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3696_dilations_0 = const()[name = string("op_3696_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3696_groups_0 = const()[name = string("op_3696_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_3696 = conv(dilations = var_3696_dilations_0, groups = var_3696_groups_0, pad = var_3696_pad_0, pad_type = var_3696_pad_type_0, strides = var_3696_strides_0, weight = encoder_layers_18_self_attn_q_proj_weight, x = input_181)[name = string("op_3696")];
            tensor<int32, [4]> var_3697 = const()[name = string("op_3697"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_3698 = reshape(shape = var_3697, x = var_3696)[name = string("op_3698")];
            tensor<int32, [4]> var_3699 = const()[name = string("op_3699"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3706_pad_type_0 = const()[name = string("op_3706_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3706_strides_0 = const()[name = string("op_3706_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3706_pad_0 = const()[name = string("op_3706_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3706_dilations_0 = const()[name = string("op_3706_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3706_groups_0 = const()[name = string("op_3706_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3706 = conv(dilations = var_3706_dilations_0, groups = var_3706_groups_0, pad = var_3706_pad_0, pad_type = var_3706_pad_type_0, strides = var_3706_strides_0, weight = encoder_layers_18_self_attn_k_proj_weight, x = input_181)[name = string("op_3706")];
            tensor<int32, [4]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3708 = reshape(shape = var_3707, x = var_3706)[name = string("op_3708")];
            tensor<int32, [4]> var_3709 = const()[name = string("op_3709"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3716_pad_type_0 = const()[name = string("op_3716_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3716_strides_0 = const()[name = string("op_3716_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3716_pad_0 = const()[name = string("op_3716_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3716_dilations_0 = const()[name = string("op_3716_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3716_groups_0 = const()[name = string("op_3716_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3716 = conv(dilations = var_3716_dilations_0, groups = var_3716_groups_0, pad = var_3716_pad_0, pad_type = var_3716_pad_type_0, strides = var_3716_strides_0, weight = encoder_layers_18_self_attn_v_proj_weight, x = input_181)[name = string("op_3716")];
            tensor<int32, [4]> var_3717 = const()[name = string("op_3717"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3718 = reshape(shape = var_3717, x = var_3716)[name = string("op_3718")];
            tensor<int32, [4]> var_3719 = const()[name = string("op_3719"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_73_to_fp16 = const()[name = string("op_6_promoted_73_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_109 = transpose(perm = var_3699, x = var_3698)[name = string("transpose_88")];
            tensor<fp16, [1, 16, 512, 128]> var_3725_cast_fp16 = pow(x = q_109, y = var_6_promoted_73_to_fp16)[name = string("op_3725_cast_fp16")];
            tensor<int32, [1]> var_147_axes_0 = const()[name = string("var_147_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_147_keep_dims_0 = const()[name = string("var_147_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_147_cast_fp16 = reduce_mean(axes = var_147_axes_0, keep_dims = var_147_keep_dims_0, x = var_3725_cast_fp16)[name = string("var_147_cast_fp16")];
            fp16 var_3728_to_fp16 = const()[name = string("op_3728_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_3729_cast_fp16 = add(x = var_147_cast_fp16, y = var_3728_to_fp16)[name = string("op_3729_cast_fp16")];
            fp32 var_3730_epsilon_0 = const()[name = string("op_3730_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_3730_cast_fp16 = rsqrt(epsilon = var_3730_epsilon_0, x = var_3729_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_623_cast_fp16 = mul(x = q_109, y = var_3730_cast_fp16)[name = string("x_623_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_111 = mul(x = x_623_cast_fp16, y = encoder_layers_18_self_attn_q_norm_weight)[name = string("q_111")];
            fp16 var_6_promoted_74_to_fp16 = const()[name = string("op_6_promoted_74_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_109 = transpose(perm = var_3709, x = var_3708)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 512, 128]> var_3738_cast_fp16 = pow(x = k_109, y = var_6_promoted_74_to_fp16)[name = string("op_3738_cast_fp16")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = string("var_149_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_149_keep_dims_0 = const()[name = string("var_149_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_149_cast_fp16 = reduce_mean(axes = var_149_axes_0, keep_dims = var_149_keep_dims_0, x = var_3738_cast_fp16)[name = string("var_149_cast_fp16")];
            fp16 var_3741_to_fp16 = const()[name = string("op_3741_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_3742_cast_fp16 = add(x = var_149_cast_fp16, y = var_3741_to_fp16)[name = string("op_3742_cast_fp16")];
            fp32 var_3743_epsilon_0 = const()[name = string("op_3743_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_3743_cast_fp16 = rsqrt(epsilon = var_3743_epsilon_0, x = var_3742_cast_fp16)[name = string("op_3743_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_629_cast_fp16 = mul(x = k_109, y = var_3743_cast_fp16)[name = string("x_629_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_111 = mul(x = x_629_cast_fp16, y = encoder_layers_18_self_attn_k_norm_weight)[name = string("k_111")];
            tensor<fp16, [1, 16, 512, 128]> var_3747 = mul(x = q_111, y = cos)[name = string("op_3747")];
            tensor<int32, [2]> var_3748_split_sizes_0 = const()[name = string("op_3748_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3748_axis_0 = const()[name = string("op_3748_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_3748_0, tensor<fp16, [1, 16, 512, 64]> var_3748_1 = split(axis = var_3748_axis_0, split_sizes = var_3748_split_sizes_0, x = q_111)[name = string("op_3748")];
            fp16 const_57_promoted = const()[name = string("const_57_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_3750 = mul(x = var_3748_1, y = const_57_promoted)[name = string("op_3750")];
            bool var_3752_interleave_0 = const()[name = string("op_3752_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3752 = concat(axis = var_18, interleave = var_3752_interleave_0, values = (var_3750, var_3748_0))[name = string("op_3752")];
            tensor<fp16, [1, 16, 512, 128]> var_3753 = mul(x = var_3752, y = sin)[name = string("op_3753")];
            tensor<fp16, [1, 16, 512, 128]> query_37 = add(x = var_3747, y = var_3753)[name = string("query_37")];
            tensor<fp16, [1, 8, 512, 128]> var_3755 = mul(x = k_111, y = cos)[name = string("op_3755")];
            tensor<int32, [2]> var_3756_split_sizes_0 = const()[name = string("op_3756_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3756_axis_0 = const()[name = string("op_3756_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_3756_0, tensor<fp16, [1, 8, 512, 64]> var_3756_1 = split(axis = var_3756_axis_0, split_sizes = var_3756_split_sizes_0, x = k_111)[name = string("op_3756")];
            fp16 const_58_promoted = const()[name = string("const_58_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_3758 = mul(x = var_3756_1, y = const_58_promoted)[name = string("op_3758")];
            bool var_3760_interleave_0 = const()[name = string("op_3760_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_3760 = concat(axis = var_18, interleave = var_3760_interleave_0, values = (var_3758, var_3756_0))[name = string("op_3760")];
            tensor<fp16, [1, 8, 512, 128]> var_3761 = mul(x = var_3760, y = sin)[name = string("op_3761")];
            tensor<fp16, [1, 8, 512, 128]> x_631 = add(x = var_3755, y = var_3761)[name = string("x_631")];
            tensor<int32, [1]> var_3763_axes_0 = const()[name = string("op_3763_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3763 = expand_dims(axes = var_3763_axes_0, x = x_631)[name = string("op_3763")];
            tensor<int32, [5]> x_633_reps_0 = const()[name = string("x_633_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_633 = tile(reps = x_633_reps_0, x = var_3763)[name = string("x_633")];
            tensor<int32, [4]> var_3766 = const()[name = string("op_3766"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_37 = reshape(shape = var_3766, x = x_633)[name = string("key_37")];
            tensor<int32, [1]> var_3768_axes_0 = const()[name = string("op_3768_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_635 = transpose(perm = var_3719, x = var_3718)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3768 = expand_dims(axes = var_3768_axes_0, x = x_635)[name = string("op_3768")];
            tensor<int32, [5]> x_637_reps_0 = const()[name = string("x_637_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_637 = tile(reps = x_637_reps_0, x = var_3768)[name = string("x_637")];
            tensor<int32, [4]> var_3771 = const()[name = string("op_3771"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_37 = reshape(shape = var_3771, x = x_637)[name = string("value_37")];
            bool var_3776_transpose_x_1 = const()[name = string("op_3776_transpose_x_1"), val = bool(false)];
            bool var_3776_transpose_y_1 = const()[name = string("op_3776_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_3776_cast_fp16 = matmul(transpose_x = var_3776_transpose_x_1, transpose_y = var_3776_transpose_y_1, x = query_37, y = key_37)[name = string("op_3776_cast_fp16")];
            fp16 var_3777_to_fp16 = const()[name = string("op_3777_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_109_cast_fp16 = mul(x = var_3776_cast_fp16, y = var_3777_to_fp16)[name = string("attn_weights_109_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_111_cast_fp16 = add(x = attn_weights_109_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_111_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_3781_cast_fp16 = softmax(axis = var_18, x = attn_weights_111_cast_fp16)[name = string("op_3781_cast_fp16")];
            bool var_3785_transpose_x_0 = const()[name = string("op_3785_transpose_x_0"), val = bool(false)];
            bool var_3785_transpose_y_0 = const()[name = string("op_3785_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3785_cast_fp16 = matmul(transpose_x = var_3785_transpose_x_0, transpose_y = var_3785_transpose_y_0, x = var_3781_cast_fp16, y = value_37)[name = string("op_3785_cast_fp16")];
            tensor<int32, [4]> var_3787 = const()[name = string("op_3787"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3790 = const()[name = string("op_3790"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3788 = transpose(perm = var_3787, x = var_3785_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 512, 2048]> attn_out_111 = reshape(shape = var_3790, x = var_3788)[name = string("attn_out_111")];
            tensor<int32, [3]> var_3792 = const()[name = string("op_3792"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_18 = const()[name = string("squeeze_18"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1166911168)))];
            string var_3801_pad_type_0 = const()[name = string("op_3801_pad_type_0"), val = string("valid")];
            int32 var_3801_groups_0 = const()[name = string("op_3801_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3801_strides_0 = const()[name = string("op_3801_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3801_pad_0 = const()[name = string("op_3801_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3801_dilations_0 = const()[name = string("op_3801_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3793 = transpose(perm = var_3792, x = attn_out_111)[name = string("transpose_84")];
            tensor<fp16, [1, 1024, 512]> var_3801 = conv(dilations = var_3801_dilations_0, groups = var_3801_groups_0, pad = var_3801_pad_0, pad_type = var_3801_pad_type_0, strides = var_3801_strides_0, weight = squeeze_18, x = var_3793)[name = string("op_3801")];
            tensor<int32, [3]> var_3802 = const()[name = string("op_3802"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_113 = transpose(perm = var_3802, x = var_3801)[name = string("transpose_83")];
            tensor<fp16, [1, 512, 1024]> x_639_cast_fp16 = add(x = hidden_states_37_cast_fp16, y = attn_out_113)[name = string("x_639_cast_fp16")];
            fp16 var_6_promoted_75_to_fp16 = const()[name = string("op_6_promoted_75_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3808_cast_fp16 = pow(x = x_639_cast_fp16, y = var_6_promoted_75_to_fp16)[name = string("op_3808_cast_fp16")];
            tensor<int32, [1]> var_151_axes_0 = const()[name = string("var_151_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_151_keep_dims_0 = const()[name = string("var_151_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_151_cast_fp16 = reduce_mean(axes = var_151_axes_0, keep_dims = var_151_keep_dims_0, x = var_3808_cast_fp16)[name = string("var_151_cast_fp16")];
            fp16 var_3811_to_fp16 = const()[name = string("op_3811_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3812_cast_fp16 = add(x = var_151_cast_fp16, y = var_3811_to_fp16)[name = string("op_3812_cast_fp16")];
            fp32 var_3813_epsilon_0 = const()[name = string("op_3813_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3813_cast_fp16 = rsqrt(epsilon = var_3813_epsilon_0, x = var_3812_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_643_cast_fp16 = mul(x = x_639_cast_fp16, y = var_3813_cast_fp16)[name = string("x_643_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171105536)))];
            tensor<fp16, [1, 512, 1024]> var_3816_cast_fp16 = mul(x = x_643_cast_fp16, y = encoder_layers_18_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_3816_cast_fp16")];
            tensor<int32, [3]> var_3821 = const()[name = string("op_3821"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_185_axes_0 = const()[name = string("input_185_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3822 = transpose(perm = var_3821, x = var_3816_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 1024, 1, 512]> input_185 = expand_dims(axes = input_185_axes_0, x = var_3822)[name = string("input_185")];
            string input_187_pad_type_0 = const()[name = string("input_187_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_187_strides_0 = const()[name = string("input_187_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_187_pad_0 = const()[name = string("input_187_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_187_dilations_0 = const()[name = string("input_187_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_187_groups_0 = const()[name = string("input_187_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_187 = conv(dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = encoder_layers_18_mlp_gate_proj_weight, x = input_185)[name = string("input_187")];
            string up_37_pad_type_0 = const()[name = string("up_37_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_37_strides_0 = const()[name = string("up_37_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_37_pad_0 = const()[name = string("up_37_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_37_dilations_0 = const()[name = string("up_37_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_37_groups_0 = const()[name = string("up_37_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_37 = conv(dilations = up_37_dilations_0, groups = up_37_groups_0, pad = up_37_pad_0, pad_type = up_37_pad_type_0, strides = up_37_strides_0, weight = encoder_layers_18_mlp_up_proj_weight, x = input_185)[name = string("up_37")];
            tensor<fp16, [1, 3072, 1, 512]> var_3836 = silu(x = input_187)[name = string("op_3836")];
            tensor<fp16, [1, 3072, 1, 512]> input_189 = mul(x = var_3836, y = up_37)[name = string("input_189")];
            string var_3843_pad_type_0 = const()[name = string("op_3843_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3843_strides_0 = const()[name = string("op_3843_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3843_pad_0 = const()[name = string("op_3843_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3843_dilations_0 = const()[name = string("op_3843_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3843_groups_0 = const()[name = string("op_3843_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3843 = conv(dilations = var_3843_dilations_0, groups = var_3843_groups_0, pad = var_3843_pad_0, pad_type = var_3843_pad_type_0, strides = var_3843_strides_0, weight = encoder_layers_18_mlp_down_proj_weight, x = input_189)[name = string("op_3843")];
            tensor<int32, [1]> var_3844_axes_0 = const()[name = string("op_3844_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3844 = squeeze(axes = var_3844_axes_0, x = var_3843)[name = string("op_3844")];
            tensor<int32, [3]> var_3845 = const()[name = string("op_3845"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_37 = transpose(perm = var_3845, x = var_3844)[name = string("transpose_81")];
            tensor<fp16, [1, 512, 1024]> hidden_states_39_cast_fp16 = add(x = x_639_cast_fp16, y = mlp_out_37)[name = string("hidden_states_39_cast_fp16")];
            fp16 var_6_promoted_76_to_fp16 = const()[name = string("op_6_promoted_76_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_3872_cast_fp16 = pow(x = hidden_states_39_cast_fp16, y = var_6_promoted_76_to_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [1]> var_153_axes_0 = const()[name = string("var_153_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_153_keep_dims_0 = const()[name = string("var_153_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_153_cast_fp16 = reduce_mean(axes = var_153_axes_0, keep_dims = var_153_keep_dims_0, x = var_3872_cast_fp16)[name = string("var_153_cast_fp16")];
            fp16 var_3875_to_fp16 = const()[name = string("op_3875_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_3876_cast_fp16 = add(x = var_153_cast_fp16, y = var_3875_to_fp16)[name = string("op_3876_cast_fp16")];
            fp32 var_3877_epsilon_0 = const()[name = string("op_3877_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_3877_cast_fp16 = rsqrt(epsilon = var_3877_epsilon_0, x = var_3876_cast_fp16)[name = string("op_3877_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_649_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = var_3877_cast_fp16)[name = string("x_649_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_19_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_19_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171107648)))];
            tensor<fp16, [1, 512, 1024]> var_3880_cast_fp16 = mul(x = x_649_cast_fp16, y = encoder_layers_19_input_layernorm_weight_promoted_to_fp16)[name = string("op_3880_cast_fp16")];
            tensor<int32, [3]> var_3885 = const()[name = string("op_3885"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_3886 = transpose(perm = var_3885, x = var_3880_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 1024, 1, 512]> input_191 = expand_dims(axes = input_191_axes_0, x = var_3886)[name = string("input_191")];
            string var_3893_pad_type_0 = const()[name = string("op_3893_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3893_strides_0 = const()[name = string("op_3893_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3893_pad_0 = const()[name = string("op_3893_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3893_dilations_0 = const()[name = string("op_3893_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3893_groups_0 = const()[name = string("op_3893_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_3893 = conv(dilations = var_3893_dilations_0, groups = var_3893_groups_0, pad = var_3893_pad_0, pad_type = var_3893_pad_type_0, strides = var_3893_strides_0, weight = encoder_layers_19_self_attn_q_proj_weight, x = input_191)[name = string("op_3893")];
            tensor<int32, [4]> var_3894 = const()[name = string("op_3894"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_3895 = reshape(shape = var_3894, x = var_3893)[name = string("op_3895")];
            tensor<int32, [4]> var_3896 = const()[name = string("op_3896"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3903_pad_type_0 = const()[name = string("op_3903_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3903_strides_0 = const()[name = string("op_3903_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3903_pad_0 = const()[name = string("op_3903_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3903_dilations_0 = const()[name = string("op_3903_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3903_groups_0 = const()[name = string("op_3903_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3903 = conv(dilations = var_3903_dilations_0, groups = var_3903_groups_0, pad = var_3903_pad_0, pad_type = var_3903_pad_type_0, strides = var_3903_strides_0, weight = encoder_layers_19_self_attn_k_proj_weight, x = input_191)[name = string("op_3903")];
            tensor<int32, [4]> var_3904 = const()[name = string("op_3904"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3905 = reshape(shape = var_3904, x = var_3903)[name = string("op_3905")];
            tensor<int32, [4]> var_3906 = const()[name = string("op_3906"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_3913_pad_type_0 = const()[name = string("op_3913_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_3913_strides_0 = const()[name = string("op_3913_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_3913_pad_0 = const()[name = string("op_3913_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_3913_dilations_0 = const()[name = string("op_3913_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_3913_groups_0 = const()[name = string("op_3913_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_3913 = conv(dilations = var_3913_dilations_0, groups = var_3913_groups_0, pad = var_3913_pad_0, pad_type = var_3913_pad_type_0, strides = var_3913_strides_0, weight = encoder_layers_19_self_attn_v_proj_weight, x = input_191)[name = string("op_3913")];
            tensor<int32, [4]> var_3914 = const()[name = string("op_3914"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_3915 = reshape(shape = var_3914, x = var_3913)[name = string("op_3915")];
            tensor<int32, [4]> var_3916 = const()[name = string("op_3916"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_77_to_fp16 = const()[name = string("op_6_promoted_77_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_115 = transpose(perm = var_3896, x = var_3895)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 512, 128]> var_3922_cast_fp16 = pow(x = q_115, y = var_6_promoted_77_to_fp16)[name = string("op_3922_cast_fp16")];
            tensor<int32, [1]> var_155_axes_0 = const()[name = string("var_155_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_155_keep_dims_0 = const()[name = string("var_155_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_155_cast_fp16 = reduce_mean(axes = var_155_axes_0, keep_dims = var_155_keep_dims_0, x = var_3922_cast_fp16)[name = string("var_155_cast_fp16")];
            fp16 var_3925_to_fp16 = const()[name = string("op_3925_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_3926_cast_fp16 = add(x = var_155_cast_fp16, y = var_3925_to_fp16)[name = string("op_3926_cast_fp16")];
            fp32 var_3927_epsilon_0 = const()[name = string("op_3927_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_3927_cast_fp16 = rsqrt(epsilon = var_3927_epsilon_0, x = var_3926_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_657_cast_fp16 = mul(x = q_115, y = var_3927_cast_fp16)[name = string("x_657_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_117 = mul(x = x_657_cast_fp16, y = encoder_layers_19_self_attn_q_norm_weight)[name = string("q_117")];
            fp16 var_6_promoted_78_to_fp16 = const()[name = string("op_6_promoted_78_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_115 = transpose(perm = var_3906, x = var_3905)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 512, 128]> var_3935_cast_fp16 = pow(x = k_115, y = var_6_promoted_78_to_fp16)[name = string("op_3935_cast_fp16")];
            tensor<int32, [1]> var_157_axes_0 = const()[name = string("var_157_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_157_keep_dims_0 = const()[name = string("var_157_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_157_cast_fp16 = reduce_mean(axes = var_157_axes_0, keep_dims = var_157_keep_dims_0, x = var_3935_cast_fp16)[name = string("var_157_cast_fp16")];
            fp16 var_3938_to_fp16 = const()[name = string("op_3938_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_3939_cast_fp16 = add(x = var_157_cast_fp16, y = var_3938_to_fp16)[name = string("op_3939_cast_fp16")];
            fp32 var_3940_epsilon_0 = const()[name = string("op_3940_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_3940_cast_fp16 = rsqrt(epsilon = var_3940_epsilon_0, x = var_3939_cast_fp16)[name = string("op_3940_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_663_cast_fp16 = mul(x = k_115, y = var_3940_cast_fp16)[name = string("x_663_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_117 = mul(x = x_663_cast_fp16, y = encoder_layers_19_self_attn_k_norm_weight)[name = string("k_117")];
            tensor<fp16, [1, 16, 512, 128]> var_3944 = mul(x = q_117, y = cos)[name = string("op_3944")];
            tensor<int32, [2]> var_3945_split_sizes_0 = const()[name = string("op_3945_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3945_axis_0 = const()[name = string("op_3945_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_3945_0, tensor<fp16, [1, 16, 512, 64]> var_3945_1 = split(axis = var_3945_axis_0, split_sizes = var_3945_split_sizes_0, x = q_117)[name = string("op_3945")];
            fp16 const_60_promoted = const()[name = string("const_60_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_3947 = mul(x = var_3945_1, y = const_60_promoted)[name = string("op_3947")];
            bool var_3949_interleave_0 = const()[name = string("op_3949_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3949 = concat(axis = var_18, interleave = var_3949_interleave_0, values = (var_3947, var_3945_0))[name = string("op_3949")];
            tensor<fp16, [1, 16, 512, 128]> var_3950 = mul(x = var_3949, y = sin)[name = string("op_3950")];
            tensor<fp16, [1, 16, 512, 128]> query_39 = add(x = var_3944, y = var_3950)[name = string("query_39")];
            tensor<fp16, [1, 8, 512, 128]> var_3952 = mul(x = k_117, y = cos)[name = string("op_3952")];
            tensor<int32, [2]> var_3953_split_sizes_0 = const()[name = string("op_3953_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_3953_axis_0 = const()[name = string("op_3953_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_3953_0, tensor<fp16, [1, 8, 512, 64]> var_3953_1 = split(axis = var_3953_axis_0, split_sizes = var_3953_split_sizes_0, x = k_117)[name = string("op_3953")];
            fp16 const_61_promoted = const()[name = string("const_61_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_3955 = mul(x = var_3953_1, y = const_61_promoted)[name = string("op_3955")];
            bool var_3957_interleave_0 = const()[name = string("op_3957_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_3957 = concat(axis = var_18, interleave = var_3957_interleave_0, values = (var_3955, var_3953_0))[name = string("op_3957")];
            tensor<fp16, [1, 8, 512, 128]> var_3958 = mul(x = var_3957, y = sin)[name = string("op_3958")];
            tensor<fp16, [1, 8, 512, 128]> x_665 = add(x = var_3952, y = var_3958)[name = string("x_665")];
            tensor<int32, [1]> var_3960_axes_0 = const()[name = string("op_3960_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3960 = expand_dims(axes = var_3960_axes_0, x = x_665)[name = string("op_3960")];
            tensor<int32, [5]> x_667_reps_0 = const()[name = string("x_667_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_667 = tile(reps = x_667_reps_0, x = var_3960)[name = string("x_667")];
            tensor<int32, [4]> var_3963 = const()[name = string("op_3963"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_39 = reshape(shape = var_3963, x = x_667)[name = string("key_39")];
            tensor<int32, [1]> var_3965_axes_0 = const()[name = string("op_3965_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_669 = transpose(perm = var_3916, x = var_3915)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_3965 = expand_dims(axes = var_3965_axes_0, x = x_669)[name = string("op_3965")];
            tensor<int32, [5]> x_671_reps_0 = const()[name = string("x_671_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_671 = tile(reps = x_671_reps_0, x = var_3965)[name = string("x_671")];
            tensor<int32, [4]> var_3968 = const()[name = string("op_3968"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_39 = reshape(shape = var_3968, x = x_671)[name = string("value_39")];
            bool var_3973_transpose_x_1 = const()[name = string("op_3973_transpose_x_1"), val = bool(false)];
            bool var_3973_transpose_y_1 = const()[name = string("op_3973_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_3973_cast_fp16 = matmul(transpose_x = var_3973_transpose_x_1, transpose_y = var_3973_transpose_y_1, x = query_39, y = key_39)[name = string("op_3973_cast_fp16")];
            fp16 var_3974_to_fp16 = const()[name = string("op_3974_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_115_cast_fp16 = mul(x = var_3973_cast_fp16, y = var_3974_to_fp16)[name = string("attn_weights_115_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_117_cast_fp16 = add(x = attn_weights_115_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_117_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_3978_cast_fp16 = softmax(axis = var_18, x = attn_weights_117_cast_fp16)[name = string("op_3978_cast_fp16")];
            bool var_3982_transpose_x_0 = const()[name = string("op_3982_transpose_x_0"), val = bool(false)];
            bool var_3982_transpose_y_0 = const()[name = string("op_3982_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_3982_cast_fp16 = matmul(transpose_x = var_3982_transpose_x_0, transpose_y = var_3982_transpose_y_0, x = var_3978_cast_fp16, y = value_39)[name = string("op_3982_cast_fp16")];
            tensor<int32, [4]> var_3984 = const()[name = string("op_3984"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3987 = const()[name = string("op_3987"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_3985 = transpose(perm = var_3984, x = var_3982_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 512, 2048]> attn_out_117 = reshape(shape = var_3987, x = var_3985)[name = string("attn_out_117")];
            tensor<int32, [3]> var_3989 = const()[name = string("op_3989"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_19 = const()[name = string("squeeze_19"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1171109760)))];
            string var_3998_pad_type_0 = const()[name = string("op_3998_pad_type_0"), val = string("valid")];
            int32 var_3998_groups_0 = const()[name = string("op_3998_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_3998_strides_0 = const()[name = string("op_3998_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_3998_pad_0 = const()[name = string("op_3998_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_3998_dilations_0 = const()[name = string("op_3998_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_3990 = transpose(perm = var_3989, x = attn_out_117)[name = string("transpose_75")];
            tensor<fp16, [1, 1024, 512]> var_3998 = conv(dilations = var_3998_dilations_0, groups = var_3998_groups_0, pad = var_3998_pad_0, pad_type = var_3998_pad_type_0, strides = var_3998_strides_0, weight = squeeze_19, x = var_3990)[name = string("op_3998")];
            tensor<int32, [3]> var_3999 = const()[name = string("op_3999"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_119 = transpose(perm = var_3999, x = var_3998)[name = string("transpose_74")];
            tensor<fp16, [1, 512, 1024]> x_673_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = attn_out_119)[name = string("x_673_cast_fp16")];
            fp16 var_6_promoted_79_to_fp16 = const()[name = string("op_6_promoted_79_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4005_cast_fp16 = pow(x = x_673_cast_fp16, y = var_6_promoted_79_to_fp16)[name = string("op_4005_cast_fp16")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = string("var_159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_159_keep_dims_0 = const()[name = string("var_159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_159_cast_fp16 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_4005_cast_fp16)[name = string("var_159_cast_fp16")];
            fp16 var_4008_to_fp16 = const()[name = string("op_4008_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4009_cast_fp16 = add(x = var_159_cast_fp16, y = var_4008_to_fp16)[name = string("op_4009_cast_fp16")];
            fp32 var_4010_epsilon_0 = const()[name = string("op_4010_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4010_cast_fp16 = rsqrt(epsilon = var_4010_epsilon_0, x = var_4009_cast_fp16)[name = string("op_4010_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_677_cast_fp16 = mul(x = x_673_cast_fp16, y = var_4010_cast_fp16)[name = string("x_677_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175304128)))];
            tensor<fp16, [1, 512, 1024]> var_4013_cast_fp16 = mul(x = x_677_cast_fp16, y = encoder_layers_19_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4013_cast_fp16")];
            tensor<int32, [3]> var_4018 = const()[name = string("op_4018"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_195_axes_0 = const()[name = string("input_195_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4019 = transpose(perm = var_4018, x = var_4013_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 1024, 1, 512]> input_195 = expand_dims(axes = input_195_axes_0, x = var_4019)[name = string("input_195")];
            string input_197_pad_type_0 = const()[name = string("input_197_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_197_strides_0 = const()[name = string("input_197_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_197_pad_0 = const()[name = string("input_197_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_197_dilations_0 = const()[name = string("input_197_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_197_groups_0 = const()[name = string("input_197_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_197 = conv(dilations = input_197_dilations_0, groups = input_197_groups_0, pad = input_197_pad_0, pad_type = input_197_pad_type_0, strides = input_197_strides_0, weight = encoder_layers_19_mlp_gate_proj_weight, x = input_195)[name = string("input_197")];
            string up_39_pad_type_0 = const()[name = string("up_39_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_39_strides_0 = const()[name = string("up_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_39_pad_0 = const()[name = string("up_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_39_dilations_0 = const()[name = string("up_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_39_groups_0 = const()[name = string("up_39_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_39 = conv(dilations = up_39_dilations_0, groups = up_39_groups_0, pad = up_39_pad_0, pad_type = up_39_pad_type_0, strides = up_39_strides_0, weight = encoder_layers_19_mlp_up_proj_weight, x = input_195)[name = string("up_39")];
            tensor<fp16, [1, 3072, 1, 512]> var_4033 = silu(x = input_197)[name = string("op_4033")];
            tensor<fp16, [1, 3072, 1, 512]> input_199 = mul(x = var_4033, y = up_39)[name = string("input_199")];
            string var_4040_pad_type_0 = const()[name = string("op_4040_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4040_strides_0 = const()[name = string("op_4040_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4040_pad_0 = const()[name = string("op_4040_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4040_dilations_0 = const()[name = string("op_4040_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4040_groups_0 = const()[name = string("op_4040_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4040 = conv(dilations = var_4040_dilations_0, groups = var_4040_groups_0, pad = var_4040_pad_0, pad_type = var_4040_pad_type_0, strides = var_4040_strides_0, weight = encoder_layers_19_mlp_down_proj_weight, x = input_199)[name = string("op_4040")];
            tensor<int32, [1]> var_4041_axes_0 = const()[name = string("op_4041_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4041 = squeeze(axes = var_4041_axes_0, x = var_4040)[name = string("op_4041")];
            tensor<int32, [3]> var_4042 = const()[name = string("op_4042"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_39 = transpose(perm = var_4042, x = var_4041)[name = string("transpose_72")];
            tensor<fp16, [1, 512, 1024]> hidden_states_41_cast_fp16 = add(x = x_673_cast_fp16, y = mlp_out_39)[name = string("hidden_states_41_cast_fp16")];
            fp16 var_6_promoted_80_to_fp16 = const()[name = string("op_6_promoted_80_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4069_cast_fp16 = pow(x = hidden_states_41_cast_fp16, y = var_6_promoted_80_to_fp16)[name = string("op_4069_cast_fp16")];
            tensor<int32, [1]> var_161_axes_0 = const()[name = string("var_161_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_161_keep_dims_0 = const()[name = string("var_161_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_161_cast_fp16 = reduce_mean(axes = var_161_axes_0, keep_dims = var_161_keep_dims_0, x = var_4069_cast_fp16)[name = string("var_161_cast_fp16")];
            fp16 var_4072_to_fp16 = const()[name = string("op_4072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4073_cast_fp16 = add(x = var_161_cast_fp16, y = var_4072_to_fp16)[name = string("op_4073_cast_fp16")];
            fp32 var_4074_epsilon_0 = const()[name = string("op_4074_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4074_cast_fp16 = rsqrt(epsilon = var_4074_epsilon_0, x = var_4073_cast_fp16)[name = string("op_4074_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_683_cast_fp16 = mul(x = hidden_states_41_cast_fp16, y = var_4074_cast_fp16)[name = string("x_683_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_20_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_20_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175306240)))];
            tensor<fp16, [1, 512, 1024]> var_4077_cast_fp16 = mul(x = x_683_cast_fp16, y = encoder_layers_20_input_layernorm_weight_promoted_to_fp16)[name = string("op_4077_cast_fp16")];
            tensor<int32, [3]> var_4082 = const()[name = string("op_4082"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4083 = transpose(perm = var_4082, x = var_4077_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 1024, 1, 512]> input_201 = expand_dims(axes = input_201_axes_0, x = var_4083)[name = string("input_201")];
            string var_4090_pad_type_0 = const()[name = string("op_4090_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4090_strides_0 = const()[name = string("op_4090_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4090_pad_0 = const()[name = string("op_4090_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4090_dilations_0 = const()[name = string("op_4090_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4090_groups_0 = const()[name = string("op_4090_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_4090 = conv(dilations = var_4090_dilations_0, groups = var_4090_groups_0, pad = var_4090_pad_0, pad_type = var_4090_pad_type_0, strides = var_4090_strides_0, weight = encoder_layers_20_self_attn_q_proj_weight, x = input_201)[name = string("op_4090")];
            tensor<int32, [4]> var_4091 = const()[name = string("op_4091"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_4092 = reshape(shape = var_4091, x = var_4090)[name = string("op_4092")];
            tensor<int32, [4]> var_4093 = const()[name = string("op_4093"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4100_pad_type_0 = const()[name = string("op_4100_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4100_strides_0 = const()[name = string("op_4100_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4100_pad_0 = const()[name = string("op_4100_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4100_dilations_0 = const()[name = string("op_4100_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4100_groups_0 = const()[name = string("op_4100_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4100 = conv(dilations = var_4100_dilations_0, groups = var_4100_groups_0, pad = var_4100_pad_0, pad_type = var_4100_pad_type_0, strides = var_4100_strides_0, weight = encoder_layers_20_self_attn_k_proj_weight, x = input_201)[name = string("op_4100")];
            tensor<int32, [4]> var_4101 = const()[name = string("op_4101"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4102 = reshape(shape = var_4101, x = var_4100)[name = string("op_4102")];
            tensor<int32, [4]> var_4103 = const()[name = string("op_4103"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4110_pad_type_0 = const()[name = string("op_4110_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4110_strides_0 = const()[name = string("op_4110_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4110_pad_0 = const()[name = string("op_4110_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4110_dilations_0 = const()[name = string("op_4110_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4110_groups_0 = const()[name = string("op_4110_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4110 = conv(dilations = var_4110_dilations_0, groups = var_4110_groups_0, pad = var_4110_pad_0, pad_type = var_4110_pad_type_0, strides = var_4110_strides_0, weight = encoder_layers_20_self_attn_v_proj_weight, x = input_201)[name = string("op_4110")];
            tensor<int32, [4]> var_4111 = const()[name = string("op_4111"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4112 = reshape(shape = var_4111, x = var_4110)[name = string("op_4112")];
            tensor<int32, [4]> var_4113 = const()[name = string("op_4113"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_81_to_fp16 = const()[name = string("op_6_promoted_81_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_121 = transpose(perm = var_4093, x = var_4092)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 512, 128]> var_4119_cast_fp16 = pow(x = q_121, y = var_6_promoted_81_to_fp16)[name = string("op_4119_cast_fp16")];
            tensor<int32, [1]> var_163_axes_0 = const()[name = string("var_163_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_163_keep_dims_0 = const()[name = string("var_163_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_163_cast_fp16 = reduce_mean(axes = var_163_axes_0, keep_dims = var_163_keep_dims_0, x = var_4119_cast_fp16)[name = string("var_163_cast_fp16")];
            fp16 var_4122_to_fp16 = const()[name = string("op_4122_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_4123_cast_fp16 = add(x = var_163_cast_fp16, y = var_4122_to_fp16)[name = string("op_4123_cast_fp16")];
            fp32 var_4124_epsilon_0 = const()[name = string("op_4124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_4124_cast_fp16 = rsqrt(epsilon = var_4124_epsilon_0, x = var_4123_cast_fp16)[name = string("op_4124_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_691_cast_fp16 = mul(x = q_121, y = var_4124_cast_fp16)[name = string("x_691_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_123 = mul(x = x_691_cast_fp16, y = encoder_layers_20_self_attn_q_norm_weight)[name = string("q_123")];
            fp16 var_6_promoted_82_to_fp16 = const()[name = string("op_6_promoted_82_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_121 = transpose(perm = var_4103, x = var_4102)[name = string("transpose_69")];
            tensor<fp16, [1, 8, 512, 128]> var_4132_cast_fp16 = pow(x = k_121, y = var_6_promoted_82_to_fp16)[name = string("op_4132_cast_fp16")];
            tensor<int32, [1]> var_165_axes_0 = const()[name = string("var_165_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_165_keep_dims_0 = const()[name = string("var_165_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_165_cast_fp16 = reduce_mean(axes = var_165_axes_0, keep_dims = var_165_keep_dims_0, x = var_4132_cast_fp16)[name = string("var_165_cast_fp16")];
            fp16 var_4135_to_fp16 = const()[name = string("op_4135_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_4136_cast_fp16 = add(x = var_165_cast_fp16, y = var_4135_to_fp16)[name = string("op_4136_cast_fp16")];
            fp32 var_4137_epsilon_0 = const()[name = string("op_4137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_4137_cast_fp16 = rsqrt(epsilon = var_4137_epsilon_0, x = var_4136_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_697_cast_fp16 = mul(x = k_121, y = var_4137_cast_fp16)[name = string("x_697_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_123 = mul(x = x_697_cast_fp16, y = encoder_layers_20_self_attn_k_norm_weight)[name = string("k_123")];
            tensor<fp16, [1, 16, 512, 128]> var_4141 = mul(x = q_123, y = cos)[name = string("op_4141")];
            tensor<int32, [2]> var_4142_split_sizes_0 = const()[name = string("op_4142_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4142_axis_0 = const()[name = string("op_4142_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_4142_0, tensor<fp16, [1, 16, 512, 64]> var_4142_1 = split(axis = var_4142_axis_0, split_sizes = var_4142_split_sizes_0, x = q_123)[name = string("op_4142")];
            fp16 const_63_promoted = const()[name = string("const_63_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_4144 = mul(x = var_4142_1, y = const_63_promoted)[name = string("op_4144")];
            bool var_4146_interleave_0 = const()[name = string("op_4146_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4146 = concat(axis = var_18, interleave = var_4146_interleave_0, values = (var_4144, var_4142_0))[name = string("op_4146")];
            tensor<fp16, [1, 16, 512, 128]> var_4147 = mul(x = var_4146, y = sin)[name = string("op_4147")];
            tensor<fp16, [1, 16, 512, 128]> query_41 = add(x = var_4141, y = var_4147)[name = string("query_41")];
            tensor<fp16, [1, 8, 512, 128]> var_4149 = mul(x = k_123, y = cos)[name = string("op_4149")];
            tensor<int32, [2]> var_4150_split_sizes_0 = const()[name = string("op_4150_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4150_axis_0 = const()[name = string("op_4150_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_4150_0, tensor<fp16, [1, 8, 512, 64]> var_4150_1 = split(axis = var_4150_axis_0, split_sizes = var_4150_split_sizes_0, x = k_123)[name = string("op_4150")];
            fp16 const_64_promoted = const()[name = string("const_64_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_4152 = mul(x = var_4150_1, y = const_64_promoted)[name = string("op_4152")];
            bool var_4154_interleave_0 = const()[name = string("op_4154_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_4154 = concat(axis = var_18, interleave = var_4154_interleave_0, values = (var_4152, var_4150_0))[name = string("op_4154")];
            tensor<fp16, [1, 8, 512, 128]> var_4155 = mul(x = var_4154, y = sin)[name = string("op_4155")];
            tensor<fp16, [1, 8, 512, 128]> x_699 = add(x = var_4149, y = var_4155)[name = string("x_699")];
            tensor<int32, [1]> var_4157_axes_0 = const()[name = string("op_4157_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4157 = expand_dims(axes = var_4157_axes_0, x = x_699)[name = string("op_4157")];
            tensor<int32, [5]> x_701_reps_0 = const()[name = string("x_701_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_701 = tile(reps = x_701_reps_0, x = var_4157)[name = string("x_701")];
            tensor<int32, [4]> var_4160 = const()[name = string("op_4160"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_41 = reshape(shape = var_4160, x = x_701)[name = string("key_41")];
            tensor<int32, [1]> var_4162_axes_0 = const()[name = string("op_4162_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_703 = transpose(perm = var_4113, x = var_4112)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4162 = expand_dims(axes = var_4162_axes_0, x = x_703)[name = string("op_4162")];
            tensor<int32, [5]> x_705_reps_0 = const()[name = string("x_705_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_705 = tile(reps = x_705_reps_0, x = var_4162)[name = string("x_705")];
            tensor<int32, [4]> var_4165 = const()[name = string("op_4165"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_41 = reshape(shape = var_4165, x = x_705)[name = string("value_41")];
            bool var_4170_transpose_x_1 = const()[name = string("op_4170_transpose_x_1"), val = bool(false)];
            bool var_4170_transpose_y_1 = const()[name = string("op_4170_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_4170_cast_fp16 = matmul(transpose_x = var_4170_transpose_x_1, transpose_y = var_4170_transpose_y_1, x = query_41, y = key_41)[name = string("op_4170_cast_fp16")];
            fp16 var_4171_to_fp16 = const()[name = string("op_4171_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_121_cast_fp16 = mul(x = var_4170_cast_fp16, y = var_4171_to_fp16)[name = string("attn_weights_121_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_123_cast_fp16 = add(x = attn_weights_121_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_123_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_4175_cast_fp16 = softmax(axis = var_18, x = attn_weights_123_cast_fp16)[name = string("op_4175_cast_fp16")];
            bool var_4179_transpose_x_0 = const()[name = string("op_4179_transpose_x_0"), val = bool(false)];
            bool var_4179_transpose_y_0 = const()[name = string("op_4179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4179_cast_fp16 = matmul(transpose_x = var_4179_transpose_x_0, transpose_y = var_4179_transpose_y_0, x = var_4175_cast_fp16, y = value_41)[name = string("op_4179_cast_fp16")];
            tensor<int32, [4]> var_4181 = const()[name = string("op_4181"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4184 = const()[name = string("op_4184"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_4182 = transpose(perm = var_4181, x = var_4179_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 512, 2048]> attn_out_123 = reshape(shape = var_4184, x = var_4182)[name = string("attn_out_123")];
            tensor<int32, [3]> var_4186 = const()[name = string("op_4186"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_20 = const()[name = string("squeeze_20"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1175308352)))];
            string var_4195_pad_type_0 = const()[name = string("op_4195_pad_type_0"), val = string("valid")];
            int32 var_4195_groups_0 = const()[name = string("op_4195_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4195_strides_0 = const()[name = string("op_4195_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4195_pad_0 = const()[name = string("op_4195_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4195_dilations_0 = const()[name = string("op_4195_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_4187 = transpose(perm = var_4186, x = attn_out_123)[name = string("transpose_66")];
            tensor<fp16, [1, 1024, 512]> var_4195 = conv(dilations = var_4195_dilations_0, groups = var_4195_groups_0, pad = var_4195_pad_0, pad_type = var_4195_pad_type_0, strides = var_4195_strides_0, weight = squeeze_20, x = var_4187)[name = string("op_4195")];
            tensor<int32, [3]> var_4196 = const()[name = string("op_4196"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_125 = transpose(perm = var_4196, x = var_4195)[name = string("transpose_65")];
            tensor<fp16, [1, 512, 1024]> x_707_cast_fp16 = add(x = hidden_states_41_cast_fp16, y = attn_out_125)[name = string("x_707_cast_fp16")];
            fp16 var_6_promoted_83_to_fp16 = const()[name = string("op_6_promoted_83_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4202_cast_fp16 = pow(x = x_707_cast_fp16, y = var_6_promoted_83_to_fp16)[name = string("op_4202_cast_fp16")];
            tensor<int32, [1]> var_167_axes_0 = const()[name = string("var_167_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_167_keep_dims_0 = const()[name = string("var_167_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_167_cast_fp16 = reduce_mean(axes = var_167_axes_0, keep_dims = var_167_keep_dims_0, x = var_4202_cast_fp16)[name = string("var_167_cast_fp16")];
            fp16 var_4205_to_fp16 = const()[name = string("op_4205_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4206_cast_fp16 = add(x = var_167_cast_fp16, y = var_4205_to_fp16)[name = string("op_4206_cast_fp16")];
            fp32 var_4207_epsilon_0 = const()[name = string("op_4207_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4207_cast_fp16 = rsqrt(epsilon = var_4207_epsilon_0, x = var_4206_cast_fp16)[name = string("op_4207_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_711_cast_fp16 = mul(x = x_707_cast_fp16, y = var_4207_cast_fp16)[name = string("x_711_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179502720)))];
            tensor<fp16, [1, 512, 1024]> var_4210_cast_fp16 = mul(x = x_711_cast_fp16, y = encoder_layers_20_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4210_cast_fp16")];
            tensor<int32, [3]> var_4215 = const()[name = string("op_4215"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_205_axes_0 = const()[name = string("input_205_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4216 = transpose(perm = var_4215, x = var_4210_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 1024, 1, 512]> input_205 = expand_dims(axes = input_205_axes_0, x = var_4216)[name = string("input_205")];
            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_207 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_20_mlp_gate_proj_weight, x = input_205)[name = string("input_207")];
            string up_41_pad_type_0 = const()[name = string("up_41_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_41_strides_0 = const()[name = string("up_41_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_41_pad_0 = const()[name = string("up_41_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_41_dilations_0 = const()[name = string("up_41_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_41_groups_0 = const()[name = string("up_41_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_41 = conv(dilations = up_41_dilations_0, groups = up_41_groups_0, pad = up_41_pad_0, pad_type = up_41_pad_type_0, strides = up_41_strides_0, weight = encoder_layers_20_mlp_up_proj_weight, x = input_205)[name = string("up_41")];
            tensor<fp16, [1, 3072, 1, 512]> var_4230 = silu(x = input_207)[name = string("op_4230")];
            tensor<fp16, [1, 3072, 1, 512]> input_209 = mul(x = var_4230, y = up_41)[name = string("input_209")];
            string var_4237_pad_type_0 = const()[name = string("op_4237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4237_strides_0 = const()[name = string("op_4237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4237_pad_0 = const()[name = string("op_4237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4237_dilations_0 = const()[name = string("op_4237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4237_groups_0 = const()[name = string("op_4237_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4237 = conv(dilations = var_4237_dilations_0, groups = var_4237_groups_0, pad = var_4237_pad_0, pad_type = var_4237_pad_type_0, strides = var_4237_strides_0, weight = encoder_layers_20_mlp_down_proj_weight, x = input_209)[name = string("op_4237")];
            tensor<int32, [1]> var_4238_axes_0 = const()[name = string("op_4238_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4238 = squeeze(axes = var_4238_axes_0, x = var_4237)[name = string("op_4238")];
            tensor<int32, [3]> var_4239 = const()[name = string("op_4239"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_41 = transpose(perm = var_4239, x = var_4238)[name = string("transpose_63")];
            tensor<fp16, [1, 512, 1024]> hidden_states_43_cast_fp16 = add(x = x_707_cast_fp16, y = mlp_out_41)[name = string("hidden_states_43_cast_fp16")];
            fp16 var_6_promoted_84_to_fp16 = const()[name = string("op_6_promoted_84_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4266_cast_fp16 = pow(x = hidden_states_43_cast_fp16, y = var_6_promoted_84_to_fp16)[name = string("op_4266_cast_fp16")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = string("var_169_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_169_keep_dims_0 = const()[name = string("var_169_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_169_cast_fp16 = reduce_mean(axes = var_169_axes_0, keep_dims = var_169_keep_dims_0, x = var_4266_cast_fp16)[name = string("var_169_cast_fp16")];
            fp16 var_4269_to_fp16 = const()[name = string("op_4269_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4270_cast_fp16 = add(x = var_169_cast_fp16, y = var_4269_to_fp16)[name = string("op_4270_cast_fp16")];
            fp32 var_4271_epsilon_0 = const()[name = string("op_4271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4271_cast_fp16 = rsqrt(epsilon = var_4271_epsilon_0, x = var_4270_cast_fp16)[name = string("op_4271_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_717_cast_fp16 = mul(x = hidden_states_43_cast_fp16, y = var_4271_cast_fp16)[name = string("x_717_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_21_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_21_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179504832)))];
            tensor<fp16, [1, 512, 1024]> var_4274_cast_fp16 = mul(x = x_717_cast_fp16, y = encoder_layers_21_input_layernorm_weight_promoted_to_fp16)[name = string("op_4274_cast_fp16")];
            tensor<int32, [3]> var_4279 = const()[name = string("op_4279"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4280 = transpose(perm = var_4279, x = var_4274_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 1024, 1, 512]> input_211 = expand_dims(axes = input_211_axes_0, x = var_4280)[name = string("input_211")];
            string var_4287_pad_type_0 = const()[name = string("op_4287_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4287_strides_0 = const()[name = string("op_4287_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4287_pad_0 = const()[name = string("op_4287_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4287_dilations_0 = const()[name = string("op_4287_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4287_groups_0 = const()[name = string("op_4287_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_4287 = conv(dilations = var_4287_dilations_0, groups = var_4287_groups_0, pad = var_4287_pad_0, pad_type = var_4287_pad_type_0, strides = var_4287_strides_0, weight = encoder_layers_21_self_attn_q_proj_weight, x = input_211)[name = string("op_4287")];
            tensor<int32, [4]> var_4288 = const()[name = string("op_4288"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_4289 = reshape(shape = var_4288, x = var_4287)[name = string("op_4289")];
            tensor<int32, [4]> var_4290 = const()[name = string("op_4290"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4297_pad_type_0 = const()[name = string("op_4297_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4297_strides_0 = const()[name = string("op_4297_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4297_pad_0 = const()[name = string("op_4297_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4297_dilations_0 = const()[name = string("op_4297_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4297_groups_0 = const()[name = string("op_4297_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4297 = conv(dilations = var_4297_dilations_0, groups = var_4297_groups_0, pad = var_4297_pad_0, pad_type = var_4297_pad_type_0, strides = var_4297_strides_0, weight = encoder_layers_21_self_attn_k_proj_weight, x = input_211)[name = string("op_4297")];
            tensor<int32, [4]> var_4298 = const()[name = string("op_4298"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4299 = reshape(shape = var_4298, x = var_4297)[name = string("op_4299")];
            tensor<int32, [4]> var_4300 = const()[name = string("op_4300"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4307_pad_type_0 = const()[name = string("op_4307_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4307_strides_0 = const()[name = string("op_4307_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4307_pad_0 = const()[name = string("op_4307_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4307_dilations_0 = const()[name = string("op_4307_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4307_groups_0 = const()[name = string("op_4307_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4307 = conv(dilations = var_4307_dilations_0, groups = var_4307_groups_0, pad = var_4307_pad_0, pad_type = var_4307_pad_type_0, strides = var_4307_strides_0, weight = encoder_layers_21_self_attn_v_proj_weight, x = input_211)[name = string("op_4307")];
            tensor<int32, [4]> var_4308 = const()[name = string("op_4308"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4309 = reshape(shape = var_4308, x = var_4307)[name = string("op_4309")];
            tensor<int32, [4]> var_4310 = const()[name = string("op_4310"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_85_to_fp16 = const()[name = string("op_6_promoted_85_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_127 = transpose(perm = var_4290, x = var_4289)[name = string("transpose_61")];
            tensor<fp16, [1, 16, 512, 128]> var_4316_cast_fp16 = pow(x = q_127, y = var_6_promoted_85_to_fp16)[name = string("op_4316_cast_fp16")];
            tensor<int32, [1]> var_171_axes_0 = const()[name = string("var_171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_171_keep_dims_0 = const()[name = string("var_171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_171_cast_fp16 = reduce_mean(axes = var_171_axes_0, keep_dims = var_171_keep_dims_0, x = var_4316_cast_fp16)[name = string("var_171_cast_fp16")];
            fp16 var_4319_to_fp16 = const()[name = string("op_4319_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_4320_cast_fp16 = add(x = var_171_cast_fp16, y = var_4319_to_fp16)[name = string("op_4320_cast_fp16")];
            fp32 var_4321_epsilon_0 = const()[name = string("op_4321_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_4321_cast_fp16 = rsqrt(epsilon = var_4321_epsilon_0, x = var_4320_cast_fp16)[name = string("op_4321_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_725_cast_fp16 = mul(x = q_127, y = var_4321_cast_fp16)[name = string("x_725_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_129 = mul(x = x_725_cast_fp16, y = encoder_layers_21_self_attn_q_norm_weight)[name = string("q_129")];
            fp16 var_6_promoted_86_to_fp16 = const()[name = string("op_6_promoted_86_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_127 = transpose(perm = var_4300, x = var_4299)[name = string("transpose_60")];
            tensor<fp16, [1, 8, 512, 128]> var_4329_cast_fp16 = pow(x = k_127, y = var_6_promoted_86_to_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [1]> var_173_axes_0 = const()[name = string("var_173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_173_keep_dims_0 = const()[name = string("var_173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_173_cast_fp16 = reduce_mean(axes = var_173_axes_0, keep_dims = var_173_keep_dims_0, x = var_4329_cast_fp16)[name = string("var_173_cast_fp16")];
            fp16 var_4332_to_fp16 = const()[name = string("op_4332_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_4333_cast_fp16 = add(x = var_173_cast_fp16, y = var_4332_to_fp16)[name = string("op_4333_cast_fp16")];
            fp32 var_4334_epsilon_0 = const()[name = string("op_4334_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_4334_cast_fp16 = rsqrt(epsilon = var_4334_epsilon_0, x = var_4333_cast_fp16)[name = string("op_4334_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_731_cast_fp16 = mul(x = k_127, y = var_4334_cast_fp16)[name = string("x_731_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_129 = mul(x = x_731_cast_fp16, y = encoder_layers_21_self_attn_k_norm_weight)[name = string("k_129")];
            tensor<fp16, [1, 16, 512, 128]> var_4338 = mul(x = q_129, y = cos)[name = string("op_4338")];
            tensor<int32, [2]> var_4339_split_sizes_0 = const()[name = string("op_4339_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4339_axis_0 = const()[name = string("op_4339_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_4339_0, tensor<fp16, [1, 16, 512, 64]> var_4339_1 = split(axis = var_4339_axis_0, split_sizes = var_4339_split_sizes_0, x = q_129)[name = string("op_4339")];
            fp16 const_66_promoted = const()[name = string("const_66_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_4341 = mul(x = var_4339_1, y = const_66_promoted)[name = string("op_4341")];
            bool var_4343_interleave_0 = const()[name = string("op_4343_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4343 = concat(axis = var_18, interleave = var_4343_interleave_0, values = (var_4341, var_4339_0))[name = string("op_4343")];
            tensor<fp16, [1, 16, 512, 128]> var_4344 = mul(x = var_4343, y = sin)[name = string("op_4344")];
            tensor<fp16, [1, 16, 512, 128]> query_43 = add(x = var_4338, y = var_4344)[name = string("query_43")];
            tensor<fp16, [1, 8, 512, 128]> var_4346 = mul(x = k_129, y = cos)[name = string("op_4346")];
            tensor<int32, [2]> var_4347_split_sizes_0 = const()[name = string("op_4347_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4347_axis_0 = const()[name = string("op_4347_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_4347_0, tensor<fp16, [1, 8, 512, 64]> var_4347_1 = split(axis = var_4347_axis_0, split_sizes = var_4347_split_sizes_0, x = k_129)[name = string("op_4347")];
            fp16 const_67_promoted = const()[name = string("const_67_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_4349 = mul(x = var_4347_1, y = const_67_promoted)[name = string("op_4349")];
            bool var_4351_interleave_0 = const()[name = string("op_4351_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_4351 = concat(axis = var_18, interleave = var_4351_interleave_0, values = (var_4349, var_4347_0))[name = string("op_4351")];
            tensor<fp16, [1, 8, 512, 128]> var_4352 = mul(x = var_4351, y = sin)[name = string("op_4352")];
            tensor<fp16, [1, 8, 512, 128]> x_733 = add(x = var_4346, y = var_4352)[name = string("x_733")];
            tensor<int32, [1]> var_4354_axes_0 = const()[name = string("op_4354_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4354 = expand_dims(axes = var_4354_axes_0, x = x_733)[name = string("op_4354")];
            tensor<int32, [5]> x_735_reps_0 = const()[name = string("x_735_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_735 = tile(reps = x_735_reps_0, x = var_4354)[name = string("x_735")];
            tensor<int32, [4]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_43 = reshape(shape = var_4357, x = x_735)[name = string("key_43")];
            tensor<int32, [1]> var_4359_axes_0 = const()[name = string("op_4359_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_737 = transpose(perm = var_4310, x = var_4309)[name = string("transpose_59")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4359 = expand_dims(axes = var_4359_axes_0, x = x_737)[name = string("op_4359")];
            tensor<int32, [5]> x_739_reps_0 = const()[name = string("x_739_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_739 = tile(reps = x_739_reps_0, x = var_4359)[name = string("x_739")];
            tensor<int32, [4]> var_4362 = const()[name = string("op_4362"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_43 = reshape(shape = var_4362, x = x_739)[name = string("value_43")];
            bool var_4367_transpose_x_1 = const()[name = string("op_4367_transpose_x_1"), val = bool(false)];
            bool var_4367_transpose_y_1 = const()[name = string("op_4367_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_4367_cast_fp16 = matmul(transpose_x = var_4367_transpose_x_1, transpose_y = var_4367_transpose_y_1, x = query_43, y = key_43)[name = string("op_4367_cast_fp16")];
            fp16 var_4368_to_fp16 = const()[name = string("op_4368_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_127_cast_fp16 = mul(x = var_4367_cast_fp16, y = var_4368_to_fp16)[name = string("attn_weights_127_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_129_cast_fp16 = add(x = attn_weights_127_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_129_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_4372_cast_fp16 = softmax(axis = var_18, x = attn_weights_129_cast_fp16)[name = string("op_4372_cast_fp16")];
            bool var_4376_transpose_x_0 = const()[name = string("op_4376_transpose_x_0"), val = bool(false)];
            bool var_4376_transpose_y_0 = const()[name = string("op_4376_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4376_cast_fp16 = matmul(transpose_x = var_4376_transpose_x_0, transpose_y = var_4376_transpose_y_0, x = var_4372_cast_fp16, y = value_43)[name = string("op_4376_cast_fp16")];
            tensor<int32, [4]> var_4378 = const()[name = string("op_4378"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4381 = const()[name = string("op_4381"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_4379 = transpose(perm = var_4378, x = var_4376_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 512, 2048]> attn_out_129 = reshape(shape = var_4381, x = var_4379)[name = string("attn_out_129")];
            tensor<int32, [3]> var_4383 = const()[name = string("op_4383"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_21 = const()[name = string("squeeze_21"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1179506944)))];
            string var_4392_pad_type_0 = const()[name = string("op_4392_pad_type_0"), val = string("valid")];
            int32 var_4392_groups_0 = const()[name = string("op_4392_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4392_strides_0 = const()[name = string("op_4392_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4392_pad_0 = const()[name = string("op_4392_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4392_dilations_0 = const()[name = string("op_4392_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_4384 = transpose(perm = var_4383, x = attn_out_129)[name = string("transpose_57")];
            tensor<fp16, [1, 1024, 512]> var_4392 = conv(dilations = var_4392_dilations_0, groups = var_4392_groups_0, pad = var_4392_pad_0, pad_type = var_4392_pad_type_0, strides = var_4392_strides_0, weight = squeeze_21, x = var_4384)[name = string("op_4392")];
            tensor<int32, [3]> var_4393 = const()[name = string("op_4393"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_131 = transpose(perm = var_4393, x = var_4392)[name = string("transpose_56")];
            tensor<fp16, [1, 512, 1024]> x_741_cast_fp16 = add(x = hidden_states_43_cast_fp16, y = attn_out_131)[name = string("x_741_cast_fp16")];
            fp16 var_6_promoted_87_to_fp16 = const()[name = string("op_6_promoted_87_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4399_cast_fp16 = pow(x = x_741_cast_fp16, y = var_6_promoted_87_to_fp16)[name = string("op_4399_cast_fp16")];
            tensor<int32, [1]> var_175_axes_0 = const()[name = string("var_175_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_175_keep_dims_0 = const()[name = string("var_175_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_175_cast_fp16 = reduce_mean(axes = var_175_axes_0, keep_dims = var_175_keep_dims_0, x = var_4399_cast_fp16)[name = string("var_175_cast_fp16")];
            fp16 var_4402_to_fp16 = const()[name = string("op_4402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4403_cast_fp16 = add(x = var_175_cast_fp16, y = var_4402_to_fp16)[name = string("op_4403_cast_fp16")];
            fp32 var_4404_epsilon_0 = const()[name = string("op_4404_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4404_cast_fp16 = rsqrt(epsilon = var_4404_epsilon_0, x = var_4403_cast_fp16)[name = string("op_4404_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_745_cast_fp16 = mul(x = x_741_cast_fp16, y = var_4404_cast_fp16)[name = string("x_745_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183701312)))];
            tensor<fp16, [1, 512, 1024]> var_4407_cast_fp16 = mul(x = x_745_cast_fp16, y = encoder_layers_21_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4407_cast_fp16")];
            tensor<int32, [3]> var_4412 = const()[name = string("op_4412"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_215_axes_0 = const()[name = string("input_215_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4413 = transpose(perm = var_4412, x = var_4407_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 1024, 1, 512]> input_215 = expand_dims(axes = input_215_axes_0, x = var_4413)[name = string("input_215")];
            string input_217_pad_type_0 = const()[name = string("input_217_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_217_strides_0 = const()[name = string("input_217_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_217_pad_0 = const()[name = string("input_217_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_217_dilations_0 = const()[name = string("input_217_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_217_groups_0 = const()[name = string("input_217_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_217 = conv(dilations = input_217_dilations_0, groups = input_217_groups_0, pad = input_217_pad_0, pad_type = input_217_pad_type_0, strides = input_217_strides_0, weight = encoder_layers_21_mlp_gate_proj_weight, x = input_215)[name = string("input_217")];
            string up_43_pad_type_0 = const()[name = string("up_43_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_43_strides_0 = const()[name = string("up_43_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_43_pad_0 = const()[name = string("up_43_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_43_dilations_0 = const()[name = string("up_43_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_43_groups_0 = const()[name = string("up_43_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_43 = conv(dilations = up_43_dilations_0, groups = up_43_groups_0, pad = up_43_pad_0, pad_type = up_43_pad_type_0, strides = up_43_strides_0, weight = encoder_layers_21_mlp_up_proj_weight, x = input_215)[name = string("up_43")];
            tensor<fp16, [1, 3072, 1, 512]> var_4427 = silu(x = input_217)[name = string("op_4427")];
            tensor<fp16, [1, 3072, 1, 512]> input_219 = mul(x = var_4427, y = up_43)[name = string("input_219")];
            string var_4434_pad_type_0 = const()[name = string("op_4434_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4434_strides_0 = const()[name = string("op_4434_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4434_pad_0 = const()[name = string("op_4434_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4434_dilations_0 = const()[name = string("op_4434_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4434_groups_0 = const()[name = string("op_4434_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4434 = conv(dilations = var_4434_dilations_0, groups = var_4434_groups_0, pad = var_4434_pad_0, pad_type = var_4434_pad_type_0, strides = var_4434_strides_0, weight = encoder_layers_21_mlp_down_proj_weight, x = input_219)[name = string("op_4434")];
            tensor<int32, [1]> var_4435_axes_0 = const()[name = string("op_4435_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4435 = squeeze(axes = var_4435_axes_0, x = var_4434)[name = string("op_4435")];
            tensor<int32, [3]> var_4436 = const()[name = string("op_4436"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_43 = transpose(perm = var_4436, x = var_4435)[name = string("transpose_54")];
            tensor<fp16, [1, 512, 1024]> hidden_states_45_cast_fp16 = add(x = x_741_cast_fp16, y = mlp_out_43)[name = string("hidden_states_45_cast_fp16")];
            fp16 var_6_promoted_88_to_fp16 = const()[name = string("op_6_promoted_88_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4463_cast_fp16 = pow(x = hidden_states_45_cast_fp16, y = var_6_promoted_88_to_fp16)[name = string("op_4463_cast_fp16")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = string("var_177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_177_keep_dims_0 = const()[name = string("var_177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_177_cast_fp16 = reduce_mean(axes = var_177_axes_0, keep_dims = var_177_keep_dims_0, x = var_4463_cast_fp16)[name = string("var_177_cast_fp16")];
            fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4467_cast_fp16 = add(x = var_177_cast_fp16, y = var_4466_to_fp16)[name = string("op_4467_cast_fp16")];
            fp32 var_4468_epsilon_0 = const()[name = string("op_4468_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4468_cast_fp16 = rsqrt(epsilon = var_4468_epsilon_0, x = var_4467_cast_fp16)[name = string("op_4468_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_751_cast_fp16 = mul(x = hidden_states_45_cast_fp16, y = var_4468_cast_fp16)[name = string("x_751_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_22_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_22_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183703424)))];
            tensor<fp16, [1, 512, 1024]> var_4471_cast_fp16 = mul(x = x_751_cast_fp16, y = encoder_layers_22_input_layernorm_weight_promoted_to_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [3]> var_4476 = const()[name = string("op_4476"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4477 = transpose(perm = var_4476, x = var_4471_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 1024, 1, 512]> input_221 = expand_dims(axes = input_221_axes_0, x = var_4477)[name = string("input_221")];
            string var_4484_pad_type_0 = const()[name = string("op_4484_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4484_strides_0 = const()[name = string("op_4484_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4484_pad_0 = const()[name = string("op_4484_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4484_dilations_0 = const()[name = string("op_4484_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4484_groups_0 = const()[name = string("op_4484_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_4484 = conv(dilations = var_4484_dilations_0, groups = var_4484_groups_0, pad = var_4484_pad_0, pad_type = var_4484_pad_type_0, strides = var_4484_strides_0, weight = encoder_layers_22_self_attn_q_proj_weight, x = input_221)[name = string("op_4484")];
            tensor<int32, [4]> var_4485 = const()[name = string("op_4485"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_4486 = reshape(shape = var_4485, x = var_4484)[name = string("op_4486")];
            tensor<int32, [4]> var_4487 = const()[name = string("op_4487"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4494_pad_type_0 = const()[name = string("op_4494_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4494_strides_0 = const()[name = string("op_4494_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4494_pad_0 = const()[name = string("op_4494_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4494_dilations_0 = const()[name = string("op_4494_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4494_groups_0 = const()[name = string("op_4494_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4494 = conv(dilations = var_4494_dilations_0, groups = var_4494_groups_0, pad = var_4494_pad_0, pad_type = var_4494_pad_type_0, strides = var_4494_strides_0, weight = encoder_layers_22_self_attn_k_proj_weight, x = input_221)[name = string("op_4494")];
            tensor<int32, [4]> var_4495 = const()[name = string("op_4495"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4496 = reshape(shape = var_4495, x = var_4494)[name = string("op_4496")];
            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4504_pad_type_0 = const()[name = string("op_4504_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4504_strides_0 = const()[name = string("op_4504_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4504_pad_0 = const()[name = string("op_4504_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4504_dilations_0 = const()[name = string("op_4504_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4504_groups_0 = const()[name = string("op_4504_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4504 = conv(dilations = var_4504_dilations_0, groups = var_4504_groups_0, pad = var_4504_pad_0, pad_type = var_4504_pad_type_0, strides = var_4504_strides_0, weight = encoder_layers_22_self_attn_v_proj_weight, x = input_221)[name = string("op_4504")];
            tensor<int32, [4]> var_4505 = const()[name = string("op_4505"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4506 = reshape(shape = var_4505, x = var_4504)[name = string("op_4506")];
            tensor<int32, [4]> var_4507 = const()[name = string("op_4507"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_89_to_fp16 = const()[name = string("op_6_promoted_89_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_133 = transpose(perm = var_4487, x = var_4486)[name = string("transpose_52")];
            tensor<fp16, [1, 16, 512, 128]> var_4513_cast_fp16 = pow(x = q_133, y = var_6_promoted_89_to_fp16)[name = string("op_4513_cast_fp16")];
            tensor<int32, [1]> var_179_axes_0 = const()[name = string("var_179_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_179_keep_dims_0 = const()[name = string("var_179_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_179_cast_fp16_0 = reduce_mean(axes = var_179_axes_0, keep_dims = var_179_keep_dims_0, x = var_4513_cast_fp16)[name = string("var_179_cast_fp16")];
            fp16 var_4516_to_fp16 = const()[name = string("op_4516_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_4517_cast_fp16 = add(x = var_179_cast_fp16_0, y = var_4516_to_fp16)[name = string("op_4517_cast_fp16")];
            fp32 var_4518_epsilon_0 = const()[name = string("op_4518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_4518_cast_fp16 = rsqrt(epsilon = var_4518_epsilon_0, x = var_4517_cast_fp16)[name = string("op_4518_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_759_cast_fp16 = mul(x = q_133, y = var_4518_cast_fp16)[name = string("x_759_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_135 = mul(x = x_759_cast_fp16, y = encoder_layers_22_self_attn_q_norm_weight)[name = string("q_135")];
            fp16 var_6_promoted_90_to_fp16 = const()[name = string("op_6_promoted_90_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_133 = transpose(perm = var_4497, x = var_4496)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 512, 128]> var_4526_cast_fp16 = pow(x = k_133, y = var_6_promoted_90_to_fp16)[name = string("op_4526_cast_fp16")];
            tensor<int32, [1]> var_181_axes_0 = const()[name = string("var_181_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_181_keep_dims_0 = const()[name = string("var_181_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_181_cast_fp16 = reduce_mean(axes = var_181_axes_0, keep_dims = var_181_keep_dims_0, x = var_4526_cast_fp16)[name = string("var_181_cast_fp16")];
            fp16 var_4529_to_fp16 = const()[name = string("op_4529_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_4530_cast_fp16 = add(x = var_181_cast_fp16, y = var_4529_to_fp16)[name = string("op_4530_cast_fp16")];
            fp32 var_4531_epsilon_0 = const()[name = string("op_4531_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_4531_cast_fp16 = rsqrt(epsilon = var_4531_epsilon_0, x = var_4530_cast_fp16)[name = string("op_4531_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_765_cast_fp16 = mul(x = k_133, y = var_4531_cast_fp16)[name = string("x_765_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_135 = mul(x = x_765_cast_fp16, y = encoder_layers_22_self_attn_k_norm_weight)[name = string("k_135")];
            tensor<fp16, [1, 16, 512, 128]> var_4535 = mul(x = q_135, y = cos)[name = string("op_4535")];
            tensor<int32, [2]> var_4536_split_sizes_0 = const()[name = string("op_4536_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4536_axis_0 = const()[name = string("op_4536_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_4536_0, tensor<fp16, [1, 16, 512, 64]> var_4536_1 = split(axis = var_4536_axis_0, split_sizes = var_4536_split_sizes_0, x = q_135)[name = string("op_4536")];
            fp16 const_69_promoted = const()[name = string("const_69_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_4538 = mul(x = var_4536_1, y = const_69_promoted)[name = string("op_4538")];
            bool var_4540_interleave_0 = const()[name = string("op_4540_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4540 = concat(axis = var_18, interleave = var_4540_interleave_0, values = (var_4538, var_4536_0))[name = string("op_4540")];
            tensor<fp16, [1, 16, 512, 128]> var_4541 = mul(x = var_4540, y = sin)[name = string("op_4541")];
            tensor<fp16, [1, 16, 512, 128]> query_45 = add(x = var_4535, y = var_4541)[name = string("query_45")];
            tensor<fp16, [1, 8, 512, 128]> var_4543 = mul(x = k_135, y = cos)[name = string("op_4543")];
            tensor<int32, [2]> var_4544_split_sizes_0 = const()[name = string("op_4544_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4544_axis_0 = const()[name = string("op_4544_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_4544_0, tensor<fp16, [1, 8, 512, 64]> var_4544_1 = split(axis = var_4544_axis_0, split_sizes = var_4544_split_sizes_0, x = k_135)[name = string("op_4544")];
            fp16 const_70_promoted = const()[name = string("const_70_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_4546 = mul(x = var_4544_1, y = const_70_promoted)[name = string("op_4546")];
            bool var_4548_interleave_0 = const()[name = string("op_4548_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_4548 = concat(axis = var_18, interleave = var_4548_interleave_0, values = (var_4546, var_4544_0))[name = string("op_4548")];
            tensor<fp16, [1, 8, 512, 128]> var_4549 = mul(x = var_4548, y = sin)[name = string("op_4549")];
            tensor<fp16, [1, 8, 512, 128]> x_767 = add(x = var_4543, y = var_4549)[name = string("x_767")];
            tensor<int32, [1]> var_4551_axes_0 = const()[name = string("op_4551_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4551 = expand_dims(axes = var_4551_axes_0, x = x_767)[name = string("op_4551")];
            tensor<int32, [5]> x_769_reps_0 = const()[name = string("x_769_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_769 = tile(reps = x_769_reps_0, x = var_4551)[name = string("x_769")];
            tensor<int32, [4]> var_4554 = const()[name = string("op_4554"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_45 = reshape(shape = var_4554, x = x_769)[name = string("key_45")];
            tensor<int32, [1]> var_4556_axes_0 = const()[name = string("op_4556_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_771 = transpose(perm = var_4507, x = var_4506)[name = string("transpose_50")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4556 = expand_dims(axes = var_4556_axes_0, x = x_771)[name = string("op_4556")];
            tensor<int32, [5]> x_773_reps_0 = const()[name = string("x_773_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_773 = tile(reps = x_773_reps_0, x = var_4556)[name = string("x_773")];
            tensor<int32, [4]> var_4559 = const()[name = string("op_4559"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_45 = reshape(shape = var_4559, x = x_773)[name = string("value_45")];
            bool var_4564_transpose_x_1 = const()[name = string("op_4564_transpose_x_1"), val = bool(false)];
            bool var_4564_transpose_y_1 = const()[name = string("op_4564_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_4564_cast_fp16 = matmul(transpose_x = var_4564_transpose_x_1, transpose_y = var_4564_transpose_y_1, x = query_45, y = key_45)[name = string("op_4564_cast_fp16")];
            fp16 var_4565_to_fp16 = const()[name = string("op_4565_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_133_cast_fp16 = mul(x = var_4564_cast_fp16, y = var_4565_to_fp16)[name = string("attn_weights_133_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_135_cast_fp16 = add(x = attn_weights_133_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_135_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_4569_cast_fp16 = softmax(axis = var_18, x = attn_weights_135_cast_fp16)[name = string("op_4569_cast_fp16")];
            bool var_4573_transpose_x_0 = const()[name = string("op_4573_transpose_x_0"), val = bool(false)];
            bool var_4573_transpose_y_0 = const()[name = string("op_4573_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4573_cast_fp16 = matmul(transpose_x = var_4573_transpose_x_0, transpose_y = var_4573_transpose_y_0, x = var_4569_cast_fp16, y = value_45)[name = string("op_4573_cast_fp16")];
            tensor<int32, [4]> var_4575 = const()[name = string("op_4575"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4578 = const()[name = string("op_4578"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_4576 = transpose(perm = var_4575, x = var_4573_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 512, 2048]> attn_out_135 = reshape(shape = var_4578, x = var_4576)[name = string("attn_out_135")];
            tensor<int32, [3]> var_4580 = const()[name = string("op_4580"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_22 = const()[name = string("squeeze_22"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1183705536)))];
            string var_4589_pad_type_0 = const()[name = string("op_4589_pad_type_0"), val = string("valid")];
            int32 var_4589_groups_0 = const()[name = string("op_4589_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4589_strides_0 = const()[name = string("op_4589_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4589_pad_0 = const()[name = string("op_4589_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4589_dilations_0 = const()[name = string("op_4589_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_4581 = transpose(perm = var_4580, x = attn_out_135)[name = string("transpose_48")];
            tensor<fp16, [1, 1024, 512]> var_4589 = conv(dilations = var_4589_dilations_0, groups = var_4589_groups_0, pad = var_4589_pad_0, pad_type = var_4589_pad_type_0, strides = var_4589_strides_0, weight = squeeze_22, x = var_4581)[name = string("op_4589")];
            tensor<int32, [3]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_137 = transpose(perm = var_4590, x = var_4589)[name = string("transpose_47")];
            tensor<fp16, [1, 512, 1024]> x_775_cast_fp16 = add(x = hidden_states_45_cast_fp16, y = attn_out_137)[name = string("x_775_cast_fp16")];
            fp16 var_6_promoted_91_to_fp16 = const()[name = string("op_6_promoted_91_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4596_cast_fp16 = pow(x = x_775_cast_fp16, y = var_6_promoted_91_to_fp16)[name = string("op_4596_cast_fp16")];
            tensor<int32, [1]> var_183_axes_0 = const()[name = string("var_183_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_183_keep_dims_0 = const()[name = string("var_183_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_183_cast_fp16_0 = reduce_mean(axes = var_183_axes_0, keep_dims = var_183_keep_dims_0, x = var_4596_cast_fp16)[name = string("var_183_cast_fp16")];
            fp16 var_4599_to_fp16 = const()[name = string("op_4599_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4600_cast_fp16 = add(x = var_183_cast_fp16_0, y = var_4599_to_fp16)[name = string("op_4600_cast_fp16")];
            fp32 var_4601_epsilon_0 = const()[name = string("op_4601_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4601_cast_fp16 = rsqrt(epsilon = var_4601_epsilon_0, x = var_4600_cast_fp16)[name = string("op_4601_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_779_cast_fp16 = mul(x = x_775_cast_fp16, y = var_4601_cast_fp16)[name = string("x_779_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187899904)))];
            tensor<fp16, [1, 512, 1024]> var_4604_cast_fp16 = mul(x = x_779_cast_fp16, y = encoder_layers_22_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4604_cast_fp16")];
            tensor<int32, [3]> var_4609 = const()[name = string("op_4609"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_225_axes_0 = const()[name = string("input_225_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4610 = transpose(perm = var_4609, x = var_4604_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 1024, 1, 512]> input_225 = expand_dims(axes = input_225_axes_0, x = var_4610)[name = string("input_225")];
            string input_227_pad_type_0 = const()[name = string("input_227_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_227_strides_0 = const()[name = string("input_227_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_227_pad_0 = const()[name = string("input_227_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_227_dilations_0 = const()[name = string("input_227_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_227_groups_0 = const()[name = string("input_227_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_227 = conv(dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = encoder_layers_22_mlp_gate_proj_weight, x = input_225)[name = string("input_227")];
            string up_45_pad_type_0 = const()[name = string("up_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_45_strides_0 = const()[name = string("up_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_45_pad_0 = const()[name = string("up_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_45_dilations_0 = const()[name = string("up_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_45_groups_0 = const()[name = string("up_45_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_45 = conv(dilations = up_45_dilations_0, groups = up_45_groups_0, pad = up_45_pad_0, pad_type = up_45_pad_type_0, strides = up_45_strides_0, weight = encoder_layers_22_mlp_up_proj_weight, x = input_225)[name = string("up_45")];
            tensor<fp16, [1, 3072, 1, 512]> var_4624 = silu(x = input_227)[name = string("op_4624")];
            tensor<fp16, [1, 3072, 1, 512]> input_229 = mul(x = var_4624, y = up_45)[name = string("input_229")];
            string var_4631_pad_type_0 = const()[name = string("op_4631_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4631_strides_0 = const()[name = string("op_4631_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4631_pad_0 = const()[name = string("op_4631_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4631_dilations_0 = const()[name = string("op_4631_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4631_groups_0 = const()[name = string("op_4631_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4631 = conv(dilations = var_4631_dilations_0, groups = var_4631_groups_0, pad = var_4631_pad_0, pad_type = var_4631_pad_type_0, strides = var_4631_strides_0, weight = encoder_layers_22_mlp_down_proj_weight, x = input_229)[name = string("op_4631")];
            tensor<int32, [1]> var_4632_axes_0 = const()[name = string("op_4632_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4632 = squeeze(axes = var_4632_axes_0, x = var_4631)[name = string("op_4632")];
            tensor<int32, [3]> var_4633 = const()[name = string("op_4633"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_45 = transpose(perm = var_4633, x = var_4632)[name = string("transpose_45")];
            tensor<fp16, [1, 512, 1024]> hidden_states_47_cast_fp16 = add(x = x_775_cast_fp16, y = mlp_out_45)[name = string("hidden_states_47_cast_fp16")];
            fp16 var_6_promoted_92_to_fp16 = const()[name = string("op_6_promoted_92_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4660_cast_fp16 = pow(x = hidden_states_47_cast_fp16, y = var_6_promoted_92_to_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [1]> var_185_axes_0 = const()[name = string("var_185_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_185_keep_dims_0 = const()[name = string("var_185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_185_cast_fp16 = reduce_mean(axes = var_185_axes_0, keep_dims = var_185_keep_dims_0, x = var_4660_cast_fp16)[name = string("var_185_cast_fp16")];
            fp16 var_4663_to_fp16 = const()[name = string("op_4663_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4664_cast_fp16 = add(x = var_185_cast_fp16, y = var_4663_to_fp16)[name = string("op_4664_cast_fp16")];
            fp32 var_4665_epsilon_0 = const()[name = string("op_4665_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4665_cast_fp16 = rsqrt(epsilon = var_4665_epsilon_0, x = var_4664_cast_fp16)[name = string("op_4665_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_785_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = var_4665_cast_fp16)[name = string("x_785_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_23_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_23_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187902016)))];
            tensor<fp16, [1, 512, 1024]> var_4668_cast_fp16 = mul(x = x_785_cast_fp16, y = encoder_layers_23_input_layernorm_weight_promoted_to_fp16)[name = string("op_4668_cast_fp16")];
            tensor<int32, [3]> var_4673 = const()[name = string("op_4673"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_231_axes_0 = const()[name = string("input_231_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4674 = transpose(perm = var_4673, x = var_4668_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 1024, 1, 512]> input_231 = expand_dims(axes = input_231_axes_0, x = var_4674)[name = string("input_231")];
            string var_4681_pad_type_0 = const()[name = string("op_4681_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4681_strides_0 = const()[name = string("op_4681_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4681_pad_0 = const()[name = string("op_4681_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4681_dilations_0 = const()[name = string("op_4681_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4681_groups_0 = const()[name = string("op_4681_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_4681 = conv(dilations = var_4681_dilations_0, groups = var_4681_groups_0, pad = var_4681_pad_0, pad_type = var_4681_pad_type_0, strides = var_4681_strides_0, weight = encoder_layers_23_self_attn_q_proj_weight, x = input_231)[name = string("op_4681")];
            tensor<int32, [4]> var_4682 = const()[name = string("op_4682"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_4683 = reshape(shape = var_4682, x = var_4681)[name = string("op_4683")];
            tensor<int32, [4]> var_4684 = const()[name = string("op_4684"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4691_pad_type_0 = const()[name = string("op_4691_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4691_strides_0 = const()[name = string("op_4691_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4691_pad_0 = const()[name = string("op_4691_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4691_dilations_0 = const()[name = string("op_4691_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4691_groups_0 = const()[name = string("op_4691_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4691 = conv(dilations = var_4691_dilations_0, groups = var_4691_groups_0, pad = var_4691_pad_0, pad_type = var_4691_pad_type_0, strides = var_4691_strides_0, weight = encoder_layers_23_self_attn_k_proj_weight, x = input_231)[name = string("op_4691")];
            tensor<int32, [4]> var_4692 = const()[name = string("op_4692"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4693 = reshape(shape = var_4692, x = var_4691)[name = string("op_4693")];
            tensor<int32, [4]> var_4694 = const()[name = string("op_4694"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4701_pad_type_0 = const()[name = string("op_4701_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4701_strides_0 = const()[name = string("op_4701_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4701_pad_0 = const()[name = string("op_4701_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4701_dilations_0 = const()[name = string("op_4701_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4701_groups_0 = const()[name = string("op_4701_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4701 = conv(dilations = var_4701_dilations_0, groups = var_4701_groups_0, pad = var_4701_pad_0, pad_type = var_4701_pad_type_0, strides = var_4701_strides_0, weight = encoder_layers_23_self_attn_v_proj_weight, x = input_231)[name = string("op_4701")];
            tensor<int32, [4]> var_4702 = const()[name = string("op_4702"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4703 = reshape(shape = var_4702, x = var_4701)[name = string("op_4703")];
            tensor<int32, [4]> var_4704 = const()[name = string("op_4704"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_93_to_fp16 = const()[name = string("op_6_promoted_93_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_139 = transpose(perm = var_4684, x = var_4683)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 512, 128]> var_4710_cast_fp16 = pow(x = q_139, y = var_6_promoted_93_to_fp16)[name = string("op_4710_cast_fp16")];
            tensor<int32, [1]> var_187_axes_0 = const()[name = string("var_187_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_187_keep_dims_0 = const()[name = string("var_187_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_187_cast_fp16 = reduce_mean(axes = var_187_axes_0, keep_dims = var_187_keep_dims_0, x = var_4710_cast_fp16)[name = string("var_187_cast_fp16")];
            fp16 var_4713_to_fp16 = const()[name = string("op_4713_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_4714_cast_fp16 = add(x = var_187_cast_fp16, y = var_4713_to_fp16)[name = string("op_4714_cast_fp16")];
            fp32 var_4715_epsilon_0 = const()[name = string("op_4715_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_4715_cast_fp16 = rsqrt(epsilon = var_4715_epsilon_0, x = var_4714_cast_fp16)[name = string("op_4715_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_793_cast_fp16 = mul(x = q_139, y = var_4715_cast_fp16)[name = string("x_793_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_141 = mul(x = x_793_cast_fp16, y = encoder_layers_23_self_attn_q_norm_weight)[name = string("q_141")];
            fp16 var_6_promoted_94_to_fp16 = const()[name = string("op_6_promoted_94_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_139 = transpose(perm = var_4694, x = var_4693)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 512, 128]> var_4723_cast_fp16 = pow(x = k_139, y = var_6_promoted_94_to_fp16)[name = string("op_4723_cast_fp16")];
            tensor<int32, [1]> var_189_axes_0 = const()[name = string("var_189_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_189_keep_dims_0 = const()[name = string("var_189_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_189_cast_fp16 = reduce_mean(axes = var_189_axes_0, keep_dims = var_189_keep_dims_0, x = var_4723_cast_fp16)[name = string("var_189_cast_fp16")];
            fp16 var_4726_to_fp16 = const()[name = string("op_4726_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_4727_cast_fp16 = add(x = var_189_cast_fp16, y = var_4726_to_fp16)[name = string("op_4727_cast_fp16")];
            fp32 var_4728_epsilon_0 = const()[name = string("op_4728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_4728_cast_fp16 = rsqrt(epsilon = var_4728_epsilon_0, x = var_4727_cast_fp16)[name = string("op_4728_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_799_cast_fp16 = mul(x = k_139, y = var_4728_cast_fp16)[name = string("x_799_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_141 = mul(x = x_799_cast_fp16, y = encoder_layers_23_self_attn_k_norm_weight)[name = string("k_141")];
            tensor<fp16, [1, 16, 512, 128]> var_4732 = mul(x = q_141, y = cos)[name = string("op_4732")];
            tensor<int32, [2]> var_4733_split_sizes_0 = const()[name = string("op_4733_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4733_axis_0 = const()[name = string("op_4733_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_4733_0, tensor<fp16, [1, 16, 512, 64]> var_4733_1 = split(axis = var_4733_axis_0, split_sizes = var_4733_split_sizes_0, x = q_141)[name = string("op_4733")];
            fp16 const_72_promoted = const()[name = string("const_72_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_4735 = mul(x = var_4733_1, y = const_72_promoted)[name = string("op_4735")];
            bool var_4737_interleave_0 = const()[name = string("op_4737_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4737 = concat(axis = var_18, interleave = var_4737_interleave_0, values = (var_4735, var_4733_0))[name = string("op_4737")];
            tensor<fp16, [1, 16, 512, 128]> var_4738 = mul(x = var_4737, y = sin)[name = string("op_4738")];
            tensor<fp16, [1, 16, 512, 128]> query_47 = add(x = var_4732, y = var_4738)[name = string("query_47")];
            tensor<fp16, [1, 8, 512, 128]> var_4740 = mul(x = k_141, y = cos)[name = string("op_4740")];
            tensor<int32, [2]> var_4741_split_sizes_0 = const()[name = string("op_4741_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4741_axis_0 = const()[name = string("op_4741_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_4741_0, tensor<fp16, [1, 8, 512, 64]> var_4741_1 = split(axis = var_4741_axis_0, split_sizes = var_4741_split_sizes_0, x = k_141)[name = string("op_4741")];
            fp16 const_73_promoted = const()[name = string("const_73_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_4743 = mul(x = var_4741_1, y = const_73_promoted)[name = string("op_4743")];
            bool var_4745_interleave_0 = const()[name = string("op_4745_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_4745 = concat(axis = var_18, interleave = var_4745_interleave_0, values = (var_4743, var_4741_0))[name = string("op_4745")];
            tensor<fp16, [1, 8, 512, 128]> var_4746 = mul(x = var_4745, y = sin)[name = string("op_4746")];
            tensor<fp16, [1, 8, 512, 128]> x_801 = add(x = var_4740, y = var_4746)[name = string("x_801")];
            tensor<int32, [1]> var_4748_axes_0 = const()[name = string("op_4748_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4748 = expand_dims(axes = var_4748_axes_0, x = x_801)[name = string("op_4748")];
            tensor<int32, [5]> x_803_reps_0 = const()[name = string("x_803_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_803 = tile(reps = x_803_reps_0, x = var_4748)[name = string("x_803")];
            tensor<int32, [4]> var_4751 = const()[name = string("op_4751"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_47 = reshape(shape = var_4751, x = x_803)[name = string("key_47")];
            tensor<int32, [1]> var_4753_axes_0 = const()[name = string("op_4753_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_805 = transpose(perm = var_4704, x = var_4703)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4753 = expand_dims(axes = var_4753_axes_0, x = x_805)[name = string("op_4753")];
            tensor<int32, [5]> x_807_reps_0 = const()[name = string("x_807_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_807 = tile(reps = x_807_reps_0, x = var_4753)[name = string("x_807")];
            tensor<int32, [4]> var_4756 = const()[name = string("op_4756"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_47 = reshape(shape = var_4756, x = x_807)[name = string("value_47")];
            bool var_4761_transpose_x_1 = const()[name = string("op_4761_transpose_x_1"), val = bool(false)];
            bool var_4761_transpose_y_1 = const()[name = string("op_4761_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_4761_cast_fp16 = matmul(transpose_x = var_4761_transpose_x_1, transpose_y = var_4761_transpose_y_1, x = query_47, y = key_47)[name = string("op_4761_cast_fp16")];
            fp16 var_4762_to_fp16 = const()[name = string("op_4762_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_139_cast_fp16 = mul(x = var_4761_cast_fp16, y = var_4762_to_fp16)[name = string("attn_weights_139_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_141_cast_fp16 = add(x = attn_weights_139_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_141_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_4766_cast_fp16 = softmax(axis = var_18, x = attn_weights_141_cast_fp16)[name = string("op_4766_cast_fp16")];
            bool var_4770_transpose_x_0 = const()[name = string("op_4770_transpose_x_0"), val = bool(false)];
            bool var_4770_transpose_y_0 = const()[name = string("op_4770_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4770_cast_fp16 = matmul(transpose_x = var_4770_transpose_x_0, transpose_y = var_4770_transpose_y_0, x = var_4766_cast_fp16, y = value_47)[name = string("op_4770_cast_fp16")];
            tensor<int32, [4]> var_4772 = const()[name = string("op_4772"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4775 = const()[name = string("op_4775"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_4773 = transpose(perm = var_4772, x = var_4770_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 512, 2048]> attn_out_141 = reshape(shape = var_4775, x = var_4773)[name = string("attn_out_141")];
            tensor<int32, [3]> var_4777 = const()[name = string("op_4777"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_23 = const()[name = string("squeeze_23"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1187904128)))];
            string var_4786_pad_type_0 = const()[name = string("op_4786_pad_type_0"), val = string("valid")];
            int32 var_4786_groups_0 = const()[name = string("op_4786_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4786_strides_0 = const()[name = string("op_4786_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4786_pad_0 = const()[name = string("op_4786_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4786_dilations_0 = const()[name = string("op_4786_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_4778 = transpose(perm = var_4777, x = attn_out_141)[name = string("transpose_39")];
            tensor<fp16, [1, 1024, 512]> var_4786 = conv(dilations = var_4786_dilations_0, groups = var_4786_groups_0, pad = var_4786_pad_0, pad_type = var_4786_pad_type_0, strides = var_4786_strides_0, weight = squeeze_23, x = var_4778)[name = string("op_4786")];
            tensor<int32, [3]> var_4787 = const()[name = string("op_4787"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_143 = transpose(perm = var_4787, x = var_4786)[name = string("transpose_38")];
            tensor<fp16, [1, 512, 1024]> x_809_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = attn_out_143)[name = string("x_809_cast_fp16")];
            fp16 var_6_promoted_95_to_fp16 = const()[name = string("op_6_promoted_95_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4793_cast_fp16 = pow(x = x_809_cast_fp16, y = var_6_promoted_95_to_fp16)[name = string("op_4793_cast_fp16")];
            tensor<int32, [1]> var_191_axes_0 = const()[name = string("var_191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_191_keep_dims_0 = const()[name = string("var_191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_191_cast_fp16 = reduce_mean(axes = var_191_axes_0, keep_dims = var_191_keep_dims_0, x = var_4793_cast_fp16)[name = string("var_191_cast_fp16")];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4797_cast_fp16 = add(x = var_191_cast_fp16, y = var_4796_to_fp16)[name = string("op_4797_cast_fp16")];
            fp32 var_4798_epsilon_0 = const()[name = string("op_4798_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4798_cast_fp16 = rsqrt(epsilon = var_4798_epsilon_0, x = var_4797_cast_fp16)[name = string("op_4798_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_813_cast_fp16 = mul(x = x_809_cast_fp16, y = var_4798_cast_fp16)[name = string("x_813_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192098496)))];
            tensor<fp16, [1, 512, 1024]> var_4801_cast_fp16 = mul(x = x_813_cast_fp16, y = encoder_layers_23_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4801_cast_fp16")];
            tensor<int32, [3]> var_4806 = const()[name = string("op_4806"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4807 = transpose(perm = var_4806, x = var_4801_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1024, 1, 512]> input_235 = expand_dims(axes = input_235_axes_0, x = var_4807)[name = string("input_235")];
            string input_237_pad_type_0 = const()[name = string("input_237_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_237_strides_0 = const()[name = string("input_237_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_237_pad_0 = const()[name = string("input_237_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_237_dilations_0 = const()[name = string("input_237_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_237_groups_0 = const()[name = string("input_237_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_237 = conv(dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = encoder_layers_23_mlp_gate_proj_weight, x = input_235)[name = string("input_237")];
            string up_47_pad_type_0 = const()[name = string("up_47_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_47_strides_0 = const()[name = string("up_47_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_47_pad_0 = const()[name = string("up_47_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_47_dilations_0 = const()[name = string("up_47_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_47_groups_0 = const()[name = string("up_47_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_47 = conv(dilations = up_47_dilations_0, groups = up_47_groups_0, pad = up_47_pad_0, pad_type = up_47_pad_type_0, strides = up_47_strides_0, weight = encoder_layers_23_mlp_up_proj_weight, x = input_235)[name = string("up_47")];
            tensor<fp16, [1, 3072, 1, 512]> var_4821 = silu(x = input_237)[name = string("op_4821")];
            tensor<fp16, [1, 3072, 1, 512]> input_239 = mul(x = var_4821, y = up_47)[name = string("input_239")];
            string var_4828_pad_type_0 = const()[name = string("op_4828_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4828_strides_0 = const()[name = string("op_4828_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4828_pad_0 = const()[name = string("op_4828_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4828_dilations_0 = const()[name = string("op_4828_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4828_groups_0 = const()[name = string("op_4828_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4828 = conv(dilations = var_4828_dilations_0, groups = var_4828_groups_0, pad = var_4828_pad_0, pad_type = var_4828_pad_type_0, strides = var_4828_strides_0, weight = encoder_layers_23_mlp_down_proj_weight, x = input_239)[name = string("op_4828")];
            tensor<int32, [1]> var_4829_axes_0 = const()[name = string("op_4829_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4829 = squeeze(axes = var_4829_axes_0, x = var_4828)[name = string("op_4829")];
            tensor<int32, [3]> var_4830 = const()[name = string("op_4830"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_47 = transpose(perm = var_4830, x = var_4829)[name = string("transpose_36")];
            tensor<fp16, [1, 512, 1024]> hidden_states_49_cast_fp16 = add(x = x_809_cast_fp16, y = mlp_out_47)[name = string("hidden_states_49_cast_fp16")];
            fp16 var_6_promoted_96_to_fp16 = const()[name = string("op_6_promoted_96_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4857_cast_fp16 = pow(x = hidden_states_49_cast_fp16, y = var_6_promoted_96_to_fp16)[name = string("op_4857_cast_fp16")];
            tensor<int32, [1]> var_193_axes_0 = const()[name = string("var_193_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_193_keep_dims_0 = const()[name = string("var_193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_193_cast_fp16 = reduce_mean(axes = var_193_axes_0, keep_dims = var_193_keep_dims_0, x = var_4857_cast_fp16)[name = string("var_193_cast_fp16")];
            fp16 var_4860_to_fp16 = const()[name = string("op_4860_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4861_cast_fp16 = add(x = var_193_cast_fp16, y = var_4860_to_fp16)[name = string("op_4861_cast_fp16")];
            fp32 var_4862_epsilon_0 = const()[name = string("op_4862_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4862_cast_fp16 = rsqrt(epsilon = var_4862_epsilon_0, x = var_4861_cast_fp16)[name = string("op_4862_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_819_cast_fp16 = mul(x = hidden_states_49_cast_fp16, y = var_4862_cast_fp16)[name = string("x_819_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_24_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_24_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192100608)))];
            tensor<fp16, [1, 512, 1024]> var_4865_cast_fp16 = mul(x = x_819_cast_fp16, y = encoder_layers_24_input_layernorm_weight_promoted_to_fp16)[name = string("op_4865_cast_fp16")];
            tensor<int32, [3]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_241_axes_0 = const()[name = string("input_241_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_4871 = transpose(perm = var_4870, x = var_4865_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 1024, 1, 512]> input_241 = expand_dims(axes = input_241_axes_0, x = var_4871)[name = string("input_241")];
            string var_4878_pad_type_0 = const()[name = string("op_4878_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4878_strides_0 = const()[name = string("op_4878_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4878_pad_0 = const()[name = string("op_4878_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4878_dilations_0 = const()[name = string("op_4878_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4878_groups_0 = const()[name = string("op_4878_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_4878 = conv(dilations = var_4878_dilations_0, groups = var_4878_groups_0, pad = var_4878_pad_0, pad_type = var_4878_pad_type_0, strides = var_4878_strides_0, weight = encoder_layers_24_self_attn_q_proj_weight, x = input_241)[name = string("op_4878")];
            tensor<int32, [4]> var_4879 = const()[name = string("op_4879"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_4880 = reshape(shape = var_4879, x = var_4878)[name = string("op_4880")];
            tensor<int32, [4]> var_4881 = const()[name = string("op_4881"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4888_pad_type_0 = const()[name = string("op_4888_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4888_strides_0 = const()[name = string("op_4888_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4888_pad_0 = const()[name = string("op_4888_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4888_dilations_0 = const()[name = string("op_4888_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4888_groups_0 = const()[name = string("op_4888_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4888 = conv(dilations = var_4888_dilations_0, groups = var_4888_groups_0, pad = var_4888_pad_0, pad_type = var_4888_pad_type_0, strides = var_4888_strides_0, weight = encoder_layers_24_self_attn_k_proj_weight, x = input_241)[name = string("op_4888")];
            tensor<int32, [4]> var_4889 = const()[name = string("op_4889"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4890 = reshape(shape = var_4889, x = var_4888)[name = string("op_4890")];
            tensor<int32, [4]> var_4891 = const()[name = string("op_4891"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_4898_pad_type_0 = const()[name = string("op_4898_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_4898_strides_0 = const()[name = string("op_4898_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_4898_pad_0 = const()[name = string("op_4898_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_4898_dilations_0 = const()[name = string("op_4898_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_4898_groups_0 = const()[name = string("op_4898_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_4898 = conv(dilations = var_4898_dilations_0, groups = var_4898_groups_0, pad = var_4898_pad_0, pad_type = var_4898_pad_type_0, strides = var_4898_strides_0, weight = encoder_layers_24_self_attn_v_proj_weight, x = input_241)[name = string("op_4898")];
            tensor<int32, [4]> var_4899 = const()[name = string("op_4899"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_4900 = reshape(shape = var_4899, x = var_4898)[name = string("op_4900")];
            tensor<int32, [4]> var_4901 = const()[name = string("op_4901"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_97_to_fp16 = const()[name = string("op_6_promoted_97_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_145 = transpose(perm = var_4881, x = var_4880)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 512, 128]> var_4907_cast_fp16 = pow(x = q_145, y = var_6_promoted_97_to_fp16)[name = string("op_4907_cast_fp16")];
            tensor<int32, [1]> var_195_axes_0 = const()[name = string("var_195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_195_keep_dims_0 = const()[name = string("var_195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_195_cast_fp16 = reduce_mean(axes = var_195_axes_0, keep_dims = var_195_keep_dims_0, x = var_4907_cast_fp16)[name = string("var_195_cast_fp16")];
            fp16 var_4910_to_fp16 = const()[name = string("op_4910_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_4911_cast_fp16 = add(x = var_195_cast_fp16, y = var_4910_to_fp16)[name = string("op_4911_cast_fp16")];
            fp32 var_4912_epsilon_0 = const()[name = string("op_4912_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_4912_cast_fp16 = rsqrt(epsilon = var_4912_epsilon_0, x = var_4911_cast_fp16)[name = string("op_4912_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_827_cast_fp16 = mul(x = q_145, y = var_4912_cast_fp16)[name = string("x_827_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_147 = mul(x = x_827_cast_fp16, y = encoder_layers_24_self_attn_q_norm_weight)[name = string("q_147")];
            fp16 var_6_promoted_98_to_fp16 = const()[name = string("op_6_promoted_98_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_145 = transpose(perm = var_4891, x = var_4890)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 512, 128]> var_4920_cast_fp16 = pow(x = k_145, y = var_6_promoted_98_to_fp16)[name = string("op_4920_cast_fp16")];
            tensor<int32, [1]> var_197_axes_0 = const()[name = string("var_197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_197_keep_dims_0 = const()[name = string("var_197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_197_cast_fp16_0 = reduce_mean(axes = var_197_axes_0, keep_dims = var_197_keep_dims_0, x = var_4920_cast_fp16)[name = string("var_197_cast_fp16")];
            fp16 var_4923_to_fp16 = const()[name = string("op_4923_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_4924_cast_fp16 = add(x = var_197_cast_fp16_0, y = var_4923_to_fp16)[name = string("op_4924_cast_fp16")];
            fp32 var_4925_epsilon_0 = const()[name = string("op_4925_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_4925_cast_fp16 = rsqrt(epsilon = var_4925_epsilon_0, x = var_4924_cast_fp16)[name = string("op_4925_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_833_cast_fp16 = mul(x = k_145, y = var_4925_cast_fp16)[name = string("x_833_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_147 = mul(x = x_833_cast_fp16, y = encoder_layers_24_self_attn_k_norm_weight)[name = string("k_147")];
            tensor<fp16, [1, 16, 512, 128]> var_4929 = mul(x = q_147, y = cos)[name = string("op_4929")];
            tensor<int32, [2]> var_4930_split_sizes_0 = const()[name = string("op_4930_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4930_axis_0 = const()[name = string("op_4930_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_4930_0, tensor<fp16, [1, 16, 512, 64]> var_4930_1 = split(axis = var_4930_axis_0, split_sizes = var_4930_split_sizes_0, x = q_147)[name = string("op_4930")];
            fp16 const_75_promoted = const()[name = string("const_75_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_4932 = mul(x = var_4930_1, y = const_75_promoted)[name = string("op_4932")];
            bool var_4934_interleave_0 = const()[name = string("op_4934_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4934 = concat(axis = var_18, interleave = var_4934_interleave_0, values = (var_4932, var_4930_0))[name = string("op_4934")];
            tensor<fp16, [1, 16, 512, 128]> var_4935 = mul(x = var_4934, y = sin)[name = string("op_4935")];
            tensor<fp16, [1, 16, 512, 128]> query_49 = add(x = var_4929, y = var_4935)[name = string("query_49")];
            tensor<fp16, [1, 8, 512, 128]> var_4937 = mul(x = k_147, y = cos)[name = string("op_4937")];
            tensor<int32, [2]> var_4938_split_sizes_0 = const()[name = string("op_4938_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_4938_axis_0 = const()[name = string("op_4938_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_4938_0, tensor<fp16, [1, 8, 512, 64]> var_4938_1 = split(axis = var_4938_axis_0, split_sizes = var_4938_split_sizes_0, x = k_147)[name = string("op_4938")];
            fp16 const_76_promoted = const()[name = string("const_76_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_4940 = mul(x = var_4938_1, y = const_76_promoted)[name = string("op_4940")];
            bool var_4942_interleave_0 = const()[name = string("op_4942_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_4942 = concat(axis = var_18, interleave = var_4942_interleave_0, values = (var_4940, var_4938_0))[name = string("op_4942")];
            tensor<fp16, [1, 8, 512, 128]> var_4943 = mul(x = var_4942, y = sin)[name = string("op_4943")];
            tensor<fp16, [1, 8, 512, 128]> x_835 = add(x = var_4937, y = var_4943)[name = string("x_835")];
            tensor<int32, [1]> var_4945_axes_0 = const()[name = string("op_4945_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4945 = expand_dims(axes = var_4945_axes_0, x = x_835)[name = string("op_4945")];
            tensor<int32, [5]> x_837_reps_0 = const()[name = string("x_837_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_837 = tile(reps = x_837_reps_0, x = var_4945)[name = string("x_837")];
            tensor<int32, [4]> var_4948 = const()[name = string("op_4948"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_49 = reshape(shape = var_4948, x = x_837)[name = string("key_49")];
            tensor<int32, [1]> var_4950_axes_0 = const()[name = string("op_4950_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_839 = transpose(perm = var_4901, x = var_4900)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_4950 = expand_dims(axes = var_4950_axes_0, x = x_839)[name = string("op_4950")];
            tensor<int32, [5]> x_841_reps_0 = const()[name = string("x_841_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_841 = tile(reps = x_841_reps_0, x = var_4950)[name = string("x_841")];
            tensor<int32, [4]> var_4953 = const()[name = string("op_4953"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_49 = reshape(shape = var_4953, x = x_841)[name = string("value_49")];
            bool var_4958_transpose_x_1 = const()[name = string("op_4958_transpose_x_1"), val = bool(false)];
            bool var_4958_transpose_y_1 = const()[name = string("op_4958_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_4958_cast_fp16 = matmul(transpose_x = var_4958_transpose_x_1, transpose_y = var_4958_transpose_y_1, x = query_49, y = key_49)[name = string("op_4958_cast_fp16")];
            fp16 var_4959_to_fp16 = const()[name = string("op_4959_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_145_cast_fp16 = mul(x = var_4958_cast_fp16, y = var_4959_to_fp16)[name = string("attn_weights_145_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_147_cast_fp16 = add(x = attn_weights_145_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_147_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_4963_cast_fp16 = softmax(axis = var_18, x = attn_weights_147_cast_fp16)[name = string("op_4963_cast_fp16")];
            bool var_4967_transpose_x_0 = const()[name = string("op_4967_transpose_x_0"), val = bool(false)];
            bool var_4967_transpose_y_0 = const()[name = string("op_4967_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_4967_cast_fp16 = matmul(transpose_x = var_4967_transpose_x_0, transpose_y = var_4967_transpose_y_0, x = var_4963_cast_fp16, y = value_49)[name = string("op_4967_cast_fp16")];
            tensor<int32, [4]> var_4969 = const()[name = string("op_4969"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4972 = const()[name = string("op_4972"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_4970 = transpose(perm = var_4969, x = var_4967_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 512, 2048]> attn_out_147 = reshape(shape = var_4972, x = var_4970)[name = string("attn_out_147")];
            tensor<int32, [3]> var_4974 = const()[name = string("op_4974"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_24 = const()[name = string("squeeze_24"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1192102720)))];
            string var_4983_pad_type_0 = const()[name = string("op_4983_pad_type_0"), val = string("valid")];
            int32 var_4983_groups_0 = const()[name = string("op_4983_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_4983_strides_0 = const()[name = string("op_4983_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_4983_pad_0 = const()[name = string("op_4983_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_4983_dilations_0 = const()[name = string("op_4983_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_4975 = transpose(perm = var_4974, x = attn_out_147)[name = string("transpose_30")];
            tensor<fp16, [1, 1024, 512]> var_4983 = conv(dilations = var_4983_dilations_0, groups = var_4983_groups_0, pad = var_4983_pad_0, pad_type = var_4983_pad_type_0, strides = var_4983_strides_0, weight = squeeze_24, x = var_4975)[name = string("op_4983")];
            tensor<int32, [3]> var_4984 = const()[name = string("op_4984"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_149 = transpose(perm = var_4984, x = var_4983)[name = string("transpose_29")];
            tensor<fp16, [1, 512, 1024]> x_843_cast_fp16 = add(x = hidden_states_49_cast_fp16, y = attn_out_149)[name = string("x_843_cast_fp16")];
            fp16 var_6_promoted_99_to_fp16 = const()[name = string("op_6_promoted_99_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_4990_cast_fp16 = pow(x = x_843_cast_fp16, y = var_6_promoted_99_to_fp16)[name = string("op_4990_cast_fp16")];
            tensor<int32, [1]> var_199_axes_0 = const()[name = string("var_199_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_199_keep_dims_0 = const()[name = string("var_199_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_199_cast_fp16 = reduce_mean(axes = var_199_axes_0, keep_dims = var_199_keep_dims_0, x = var_4990_cast_fp16)[name = string("var_199_cast_fp16")];
            fp16 var_4993_to_fp16 = const()[name = string("op_4993_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_4994_cast_fp16 = add(x = var_199_cast_fp16, y = var_4993_to_fp16)[name = string("op_4994_cast_fp16")];
            fp32 var_4995_epsilon_0 = const()[name = string("op_4995_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_4995_cast_fp16 = rsqrt(epsilon = var_4995_epsilon_0, x = var_4994_cast_fp16)[name = string("op_4995_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_847_cast_fp16 = mul(x = x_843_cast_fp16, y = var_4995_cast_fp16)[name = string("x_847_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196297088)))];
            tensor<fp16, [1, 512, 1024]> var_4998_cast_fp16 = mul(x = x_847_cast_fp16, y = encoder_layers_24_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_4998_cast_fp16")];
            tensor<int32, [3]> var_5003 = const()[name = string("op_5003"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5004 = transpose(perm = var_5003, x = var_4998_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 1024, 1, 512]> input_245 = expand_dims(axes = input_245_axes_0, x = var_5004)[name = string("input_245")];
            string input_247_pad_type_0 = const()[name = string("input_247_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_247_strides_0 = const()[name = string("input_247_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_247_pad_0 = const()[name = string("input_247_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_247_dilations_0 = const()[name = string("input_247_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_247_groups_0 = const()[name = string("input_247_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_247 = conv(dilations = input_247_dilations_0, groups = input_247_groups_0, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = input_247_strides_0, weight = encoder_layers_24_mlp_gate_proj_weight, x = input_245)[name = string("input_247")];
            string up_49_pad_type_0 = const()[name = string("up_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_49_strides_0 = const()[name = string("up_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_49_pad_0 = const()[name = string("up_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_49_dilations_0 = const()[name = string("up_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_49_groups_0 = const()[name = string("up_49_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_49 = conv(dilations = up_49_dilations_0, groups = up_49_groups_0, pad = up_49_pad_0, pad_type = up_49_pad_type_0, strides = up_49_strides_0, weight = encoder_layers_24_mlp_up_proj_weight, x = input_245)[name = string("up_49")];
            tensor<fp16, [1, 3072, 1, 512]> var_5018 = silu(x = input_247)[name = string("op_5018")];
            tensor<fp16, [1, 3072, 1, 512]> input_249 = mul(x = var_5018, y = up_49)[name = string("input_249")];
            string var_5025_pad_type_0 = const()[name = string("op_5025_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5025_strides_0 = const()[name = string("op_5025_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5025_pad_0 = const()[name = string("op_5025_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5025_dilations_0 = const()[name = string("op_5025_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5025_groups_0 = const()[name = string("op_5025_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5025 = conv(dilations = var_5025_dilations_0, groups = var_5025_groups_0, pad = var_5025_pad_0, pad_type = var_5025_pad_type_0, strides = var_5025_strides_0, weight = encoder_layers_24_mlp_down_proj_weight, x = input_249)[name = string("op_5025")];
            tensor<int32, [1]> var_5026_axes_0 = const()[name = string("op_5026_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5026 = squeeze(axes = var_5026_axes_0, x = var_5025)[name = string("op_5026")];
            tensor<int32, [3]> var_5027 = const()[name = string("op_5027"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_49 = transpose(perm = var_5027, x = var_5026)[name = string("transpose_27")];
            tensor<fp16, [1, 512, 1024]> hidden_states_51_cast_fp16 = add(x = x_843_cast_fp16, y = mlp_out_49)[name = string("hidden_states_51_cast_fp16")];
            fp16 var_6_promoted_100_to_fp16 = const()[name = string("op_6_promoted_100_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5054_cast_fp16 = pow(x = hidden_states_51_cast_fp16, y = var_6_promoted_100_to_fp16)[name = string("op_5054_cast_fp16")];
            tensor<int32, [1]> var_201_axes_0 = const()[name = string("var_201_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_201_keep_dims_0 = const()[name = string("var_201_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_201_cast_fp16 = reduce_mean(axes = var_201_axes_0, keep_dims = var_201_keep_dims_0, x = var_5054_cast_fp16)[name = string("var_201_cast_fp16")];
            fp16 var_5057_to_fp16 = const()[name = string("op_5057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5058_cast_fp16 = add(x = var_201_cast_fp16, y = var_5057_to_fp16)[name = string("op_5058_cast_fp16")];
            fp32 var_5059_epsilon_0 = const()[name = string("op_5059_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5059_cast_fp16 = rsqrt(epsilon = var_5059_epsilon_0, x = var_5058_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_853_cast_fp16 = mul(x = hidden_states_51_cast_fp16, y = var_5059_cast_fp16)[name = string("x_853_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_25_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_25_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196299200)))];
            tensor<fp16, [1, 512, 1024]> var_5062_cast_fp16 = mul(x = x_853_cast_fp16, y = encoder_layers_25_input_layernorm_weight_promoted_to_fp16)[name = string("op_5062_cast_fp16")];
            tensor<int32, [3]> var_5067 = const()[name = string("op_5067"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_251_axes_0 = const()[name = string("input_251_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5068 = transpose(perm = var_5067, x = var_5062_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 1024, 1, 512]> input_251 = expand_dims(axes = input_251_axes_0, x = var_5068)[name = string("input_251")];
            string var_5075_pad_type_0 = const()[name = string("op_5075_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5075_strides_0 = const()[name = string("op_5075_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5075_pad_0 = const()[name = string("op_5075_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5075_dilations_0 = const()[name = string("op_5075_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5075_groups_0 = const()[name = string("op_5075_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_5075 = conv(dilations = var_5075_dilations_0, groups = var_5075_groups_0, pad = var_5075_pad_0, pad_type = var_5075_pad_type_0, strides = var_5075_strides_0, weight = encoder_layers_25_self_attn_q_proj_weight, x = input_251)[name = string("op_5075")];
            tensor<int32, [4]> var_5076 = const()[name = string("op_5076"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_5077 = reshape(shape = var_5076, x = var_5075)[name = string("op_5077")];
            tensor<int32, [4]> var_5078 = const()[name = string("op_5078"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5085_pad_type_0 = const()[name = string("op_5085_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5085_strides_0 = const()[name = string("op_5085_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5085_pad_0 = const()[name = string("op_5085_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5085_dilations_0 = const()[name = string("op_5085_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5085_groups_0 = const()[name = string("op_5085_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5085 = conv(dilations = var_5085_dilations_0, groups = var_5085_groups_0, pad = var_5085_pad_0, pad_type = var_5085_pad_type_0, strides = var_5085_strides_0, weight = encoder_layers_25_self_attn_k_proj_weight, x = input_251)[name = string("op_5085")];
            tensor<int32, [4]> var_5086 = const()[name = string("op_5086"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5087 = reshape(shape = var_5086, x = var_5085)[name = string("op_5087")];
            tensor<int32, [4]> var_5088 = const()[name = string("op_5088"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5095_pad_type_0 = const()[name = string("op_5095_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5095_strides_0 = const()[name = string("op_5095_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5095_pad_0 = const()[name = string("op_5095_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5095_dilations_0 = const()[name = string("op_5095_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5095_groups_0 = const()[name = string("op_5095_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5095 = conv(dilations = var_5095_dilations_0, groups = var_5095_groups_0, pad = var_5095_pad_0, pad_type = var_5095_pad_type_0, strides = var_5095_strides_0, weight = encoder_layers_25_self_attn_v_proj_weight, x = input_251)[name = string("op_5095")];
            tensor<int32, [4]> var_5096 = const()[name = string("op_5096"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5097 = reshape(shape = var_5096, x = var_5095)[name = string("op_5097")];
            tensor<int32, [4]> var_5098 = const()[name = string("op_5098"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_101_to_fp16 = const()[name = string("op_6_promoted_101_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_151 = transpose(perm = var_5078, x = var_5077)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 512, 128]> var_5104_cast_fp16 = pow(x = q_151, y = var_6_promoted_101_to_fp16)[name = string("op_5104_cast_fp16")];
            tensor<int32, [1]> var_203_axes_0 = const()[name = string("var_203_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_203_keep_dims_0 = const()[name = string("var_203_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_203_cast_fp16 = reduce_mean(axes = var_203_axes_0, keep_dims = var_203_keep_dims_0, x = var_5104_cast_fp16)[name = string("var_203_cast_fp16")];
            fp16 var_5107_to_fp16 = const()[name = string("op_5107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_5108_cast_fp16 = add(x = var_203_cast_fp16, y = var_5107_to_fp16)[name = string("op_5108_cast_fp16")];
            fp32 var_5109_epsilon_0 = const()[name = string("op_5109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_5109_cast_fp16 = rsqrt(epsilon = var_5109_epsilon_0, x = var_5108_cast_fp16)[name = string("op_5109_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_861_cast_fp16 = mul(x = q_151, y = var_5109_cast_fp16)[name = string("x_861_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_153 = mul(x = x_861_cast_fp16, y = encoder_layers_25_self_attn_q_norm_weight)[name = string("q_153")];
            fp16 var_6_promoted_102_to_fp16 = const()[name = string("op_6_promoted_102_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_151 = transpose(perm = var_5088, x = var_5087)[name = string("transpose_24")];
            tensor<fp16, [1, 8, 512, 128]> var_5117_cast_fp16 = pow(x = k_151, y = var_6_promoted_102_to_fp16)[name = string("op_5117_cast_fp16")];
            tensor<int32, [1]> var_205_axes_0 = const()[name = string("var_205_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_205_keep_dims_0 = const()[name = string("var_205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_205_cast_fp16 = reduce_mean(axes = var_205_axes_0, keep_dims = var_205_keep_dims_0, x = var_5117_cast_fp16)[name = string("var_205_cast_fp16")];
            fp16 var_5120_to_fp16 = const()[name = string("op_5120_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_5121_cast_fp16 = add(x = var_205_cast_fp16, y = var_5120_to_fp16)[name = string("op_5121_cast_fp16")];
            fp32 var_5122_epsilon_0 = const()[name = string("op_5122_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_5122_cast_fp16 = rsqrt(epsilon = var_5122_epsilon_0, x = var_5121_cast_fp16)[name = string("op_5122_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_867_cast_fp16 = mul(x = k_151, y = var_5122_cast_fp16)[name = string("x_867_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_153 = mul(x = x_867_cast_fp16, y = encoder_layers_25_self_attn_k_norm_weight)[name = string("k_153")];
            tensor<fp16, [1, 16, 512, 128]> var_5126 = mul(x = q_153, y = cos)[name = string("op_5126")];
            tensor<int32, [2]> var_5127_split_sizes_0 = const()[name = string("op_5127_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5127_axis_0 = const()[name = string("op_5127_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_5127_0, tensor<fp16, [1, 16, 512, 64]> var_5127_1 = split(axis = var_5127_axis_0, split_sizes = var_5127_split_sizes_0, x = q_153)[name = string("op_5127")];
            fp16 const_78_promoted = const()[name = string("const_78_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_5129 = mul(x = var_5127_1, y = const_78_promoted)[name = string("op_5129")];
            bool var_5131_interleave_0 = const()[name = string("op_5131_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5131 = concat(axis = var_18, interleave = var_5131_interleave_0, values = (var_5129, var_5127_0))[name = string("op_5131")];
            tensor<fp16, [1, 16, 512, 128]> var_5132 = mul(x = var_5131, y = sin)[name = string("op_5132")];
            tensor<fp16, [1, 16, 512, 128]> query_51 = add(x = var_5126, y = var_5132)[name = string("query_51")];
            tensor<fp16, [1, 8, 512, 128]> var_5134 = mul(x = k_153, y = cos)[name = string("op_5134")];
            tensor<int32, [2]> var_5135_split_sizes_0 = const()[name = string("op_5135_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5135_axis_0 = const()[name = string("op_5135_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_5135_0, tensor<fp16, [1, 8, 512, 64]> var_5135_1 = split(axis = var_5135_axis_0, split_sizes = var_5135_split_sizes_0, x = k_153)[name = string("op_5135")];
            fp16 const_79_promoted = const()[name = string("const_79_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_5137 = mul(x = var_5135_1, y = const_79_promoted)[name = string("op_5137")];
            bool var_5139_interleave_0 = const()[name = string("op_5139_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_5139 = concat(axis = var_18, interleave = var_5139_interleave_0, values = (var_5137, var_5135_0))[name = string("op_5139")];
            tensor<fp16, [1, 8, 512, 128]> var_5140 = mul(x = var_5139, y = sin)[name = string("op_5140")];
            tensor<fp16, [1, 8, 512, 128]> x_869 = add(x = var_5134, y = var_5140)[name = string("x_869")];
            tensor<int32, [1]> var_5142_axes_0 = const()[name = string("op_5142_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5142 = expand_dims(axes = var_5142_axes_0, x = x_869)[name = string("op_5142")];
            tensor<int32, [5]> x_871_reps_0 = const()[name = string("x_871_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_871 = tile(reps = x_871_reps_0, x = var_5142)[name = string("x_871")];
            tensor<int32, [4]> var_5145 = const()[name = string("op_5145"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_51 = reshape(shape = var_5145, x = x_871)[name = string("key_51")];
            tensor<int32, [1]> var_5147_axes_0 = const()[name = string("op_5147_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_873 = transpose(perm = var_5098, x = var_5097)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5147 = expand_dims(axes = var_5147_axes_0, x = x_873)[name = string("op_5147")];
            tensor<int32, [5]> x_875_reps_0 = const()[name = string("x_875_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_875 = tile(reps = x_875_reps_0, x = var_5147)[name = string("x_875")];
            tensor<int32, [4]> var_5150 = const()[name = string("op_5150"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_51 = reshape(shape = var_5150, x = x_875)[name = string("value_51")];
            bool var_5155_transpose_x_1 = const()[name = string("op_5155_transpose_x_1"), val = bool(false)];
            bool var_5155_transpose_y_1 = const()[name = string("op_5155_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_5155_cast_fp16 = matmul(transpose_x = var_5155_transpose_x_1, transpose_y = var_5155_transpose_y_1, x = query_51, y = key_51)[name = string("op_5155_cast_fp16")];
            fp16 var_5156_to_fp16 = const()[name = string("op_5156_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_151_cast_fp16 = mul(x = var_5155_cast_fp16, y = var_5156_to_fp16)[name = string("attn_weights_151_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_153_cast_fp16 = add(x = attn_weights_151_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_153_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_5160_cast_fp16 = softmax(axis = var_18, x = attn_weights_153_cast_fp16)[name = string("op_5160_cast_fp16")];
            bool var_5164_transpose_x_0 = const()[name = string("op_5164_transpose_x_0"), val = bool(false)];
            bool var_5164_transpose_y_0 = const()[name = string("op_5164_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5164_cast_fp16 = matmul(transpose_x = var_5164_transpose_x_0, transpose_y = var_5164_transpose_y_0, x = var_5160_cast_fp16, y = value_51)[name = string("op_5164_cast_fp16")];
            tensor<int32, [4]> var_5166 = const()[name = string("op_5166"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5169 = const()[name = string("op_5169"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_5167 = transpose(perm = var_5166, x = var_5164_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 512, 2048]> attn_out_153 = reshape(shape = var_5169, x = var_5167)[name = string("attn_out_153")];
            tensor<int32, [3]> var_5171 = const()[name = string("op_5171"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_25 = const()[name = string("squeeze_25"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1196301312)))];
            string var_5180_pad_type_0 = const()[name = string("op_5180_pad_type_0"), val = string("valid")];
            int32 var_5180_groups_0 = const()[name = string("op_5180_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5180_strides_0 = const()[name = string("op_5180_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5180_pad_0 = const()[name = string("op_5180_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5180_dilations_0 = const()[name = string("op_5180_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_5172 = transpose(perm = var_5171, x = attn_out_153)[name = string("transpose_21")];
            tensor<fp16, [1, 1024, 512]> var_5180 = conv(dilations = var_5180_dilations_0, groups = var_5180_groups_0, pad = var_5180_pad_0, pad_type = var_5180_pad_type_0, strides = var_5180_strides_0, weight = squeeze_25, x = var_5172)[name = string("op_5180")];
            tensor<int32, [3]> var_5181 = const()[name = string("op_5181"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_155 = transpose(perm = var_5181, x = var_5180)[name = string("transpose_20")];
            tensor<fp16, [1, 512, 1024]> x_877_cast_fp16 = add(x = hidden_states_51_cast_fp16, y = attn_out_155)[name = string("x_877_cast_fp16")];
            fp16 var_6_promoted_103_to_fp16 = const()[name = string("op_6_promoted_103_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5187_cast_fp16 = pow(x = x_877_cast_fp16, y = var_6_promoted_103_to_fp16)[name = string("op_5187_cast_fp16")];
            tensor<int32, [1]> var_207_axes_0 = const()[name = string("var_207_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_207_keep_dims_0 = const()[name = string("var_207_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_207_cast_fp16 = reduce_mean(axes = var_207_axes_0, keep_dims = var_207_keep_dims_0, x = var_5187_cast_fp16)[name = string("var_207_cast_fp16")];
            fp16 var_5190_to_fp16 = const()[name = string("op_5190_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5191_cast_fp16 = add(x = var_207_cast_fp16, y = var_5190_to_fp16)[name = string("op_5191_cast_fp16")];
            fp32 var_5192_epsilon_0 = const()[name = string("op_5192_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5192_cast_fp16 = rsqrt(epsilon = var_5192_epsilon_0, x = var_5191_cast_fp16)[name = string("op_5192_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_881_cast_fp16 = mul(x = x_877_cast_fp16, y = var_5192_cast_fp16)[name = string("x_881_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200495680)))];
            tensor<fp16, [1, 512, 1024]> var_5195_cast_fp16 = mul(x = x_881_cast_fp16, y = encoder_layers_25_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5195_cast_fp16")];
            tensor<int32, [3]> var_5200 = const()[name = string("op_5200"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_255_axes_0 = const()[name = string("input_255_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5201 = transpose(perm = var_5200, x = var_5195_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 1024, 1, 512]> input_255 = expand_dims(axes = input_255_axes_0, x = var_5201)[name = string("input_255")];
            string input_257_pad_type_0 = const()[name = string("input_257_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_257_strides_0 = const()[name = string("input_257_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_257_pad_0 = const()[name = string("input_257_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_257_dilations_0 = const()[name = string("input_257_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_257_groups_0 = const()[name = string("input_257_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_257 = conv(dilations = input_257_dilations_0, groups = input_257_groups_0, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = input_257_strides_0, weight = encoder_layers_25_mlp_gate_proj_weight, x = input_255)[name = string("input_257")];
            string up_51_pad_type_0 = const()[name = string("up_51_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_51_strides_0 = const()[name = string("up_51_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_51_pad_0 = const()[name = string("up_51_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_51_dilations_0 = const()[name = string("up_51_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_51_groups_0 = const()[name = string("up_51_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_51 = conv(dilations = up_51_dilations_0, groups = up_51_groups_0, pad = up_51_pad_0, pad_type = up_51_pad_type_0, strides = up_51_strides_0, weight = encoder_layers_25_mlp_up_proj_weight, x = input_255)[name = string("up_51")];
            tensor<fp16, [1, 3072, 1, 512]> var_5215 = silu(x = input_257)[name = string("op_5215")];
            tensor<fp16, [1, 3072, 1, 512]> input_259 = mul(x = var_5215, y = up_51)[name = string("input_259")];
            string var_5222_pad_type_0 = const()[name = string("op_5222_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5222_strides_0 = const()[name = string("op_5222_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5222_pad_0 = const()[name = string("op_5222_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5222_dilations_0 = const()[name = string("op_5222_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5222_groups_0 = const()[name = string("op_5222_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5222 = conv(dilations = var_5222_dilations_0, groups = var_5222_groups_0, pad = var_5222_pad_0, pad_type = var_5222_pad_type_0, strides = var_5222_strides_0, weight = encoder_layers_25_mlp_down_proj_weight, x = input_259)[name = string("op_5222")];
            tensor<int32, [1]> var_5223_axes_0 = const()[name = string("op_5223_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5223 = squeeze(axes = var_5223_axes_0, x = var_5222)[name = string("op_5223")];
            tensor<int32, [3]> var_5224 = const()[name = string("op_5224"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_51 = transpose(perm = var_5224, x = var_5223)[name = string("transpose_18")];
            tensor<fp16, [1, 512, 1024]> hidden_states_53_cast_fp16 = add(x = x_877_cast_fp16, y = mlp_out_51)[name = string("hidden_states_53_cast_fp16")];
            fp16 var_6_promoted_104_to_fp16 = const()[name = string("op_6_promoted_104_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5251_cast_fp16 = pow(x = hidden_states_53_cast_fp16, y = var_6_promoted_104_to_fp16)[name = string("op_5251_cast_fp16")];
            tensor<int32, [1]> var_209_axes_0 = const()[name = string("var_209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_209_keep_dims_0 = const()[name = string("var_209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_209_cast_fp16 = reduce_mean(axes = var_209_axes_0, keep_dims = var_209_keep_dims_0, x = var_5251_cast_fp16)[name = string("var_209_cast_fp16")];
            fp16 var_5254_to_fp16 = const()[name = string("op_5254_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5255_cast_fp16 = add(x = var_209_cast_fp16, y = var_5254_to_fp16)[name = string("op_5255_cast_fp16")];
            fp32 var_5256_epsilon_0 = const()[name = string("op_5256_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5256_cast_fp16 = rsqrt(epsilon = var_5256_epsilon_0, x = var_5255_cast_fp16)[name = string("op_5256_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_887_cast_fp16 = mul(x = hidden_states_53_cast_fp16, y = var_5256_cast_fp16)[name = string("x_887_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_26_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_26_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200497792)))];
            tensor<fp16, [1, 512, 1024]> var_5259_cast_fp16 = mul(x = x_887_cast_fp16, y = encoder_layers_26_input_layernorm_weight_promoted_to_fp16)[name = string("op_5259_cast_fp16")];
            tensor<int32, [3]> var_5264 = const()[name = string("op_5264"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_261_axes_0 = const()[name = string("input_261_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5265 = transpose(perm = var_5264, x = var_5259_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 1024, 1, 512]> input_261 = expand_dims(axes = input_261_axes_0, x = var_5265)[name = string("input_261")];
            string var_5272_pad_type_0 = const()[name = string("op_5272_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5272_strides_0 = const()[name = string("op_5272_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5272_pad_0 = const()[name = string("op_5272_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5272_dilations_0 = const()[name = string("op_5272_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5272_groups_0 = const()[name = string("op_5272_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_5272 = conv(dilations = var_5272_dilations_0, groups = var_5272_groups_0, pad = var_5272_pad_0, pad_type = var_5272_pad_type_0, strides = var_5272_strides_0, weight = encoder_layers_26_self_attn_q_proj_weight, x = input_261)[name = string("op_5272")];
            tensor<int32, [4]> var_5273 = const()[name = string("op_5273"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_5274 = reshape(shape = var_5273, x = var_5272)[name = string("op_5274")];
            tensor<int32, [4]> var_5275 = const()[name = string("op_5275"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5282_pad_type_0 = const()[name = string("op_5282_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5282_strides_0 = const()[name = string("op_5282_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5282_pad_0 = const()[name = string("op_5282_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5282_dilations_0 = const()[name = string("op_5282_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5282_groups_0 = const()[name = string("op_5282_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5282 = conv(dilations = var_5282_dilations_0, groups = var_5282_groups_0, pad = var_5282_pad_0, pad_type = var_5282_pad_type_0, strides = var_5282_strides_0, weight = encoder_layers_26_self_attn_k_proj_weight, x = input_261)[name = string("op_5282")];
            tensor<int32, [4]> var_5283 = const()[name = string("op_5283"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5284 = reshape(shape = var_5283, x = var_5282)[name = string("op_5284")];
            tensor<int32, [4]> var_5285 = const()[name = string("op_5285"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5292_pad_type_0 = const()[name = string("op_5292_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5292_strides_0 = const()[name = string("op_5292_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5292_pad_0 = const()[name = string("op_5292_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5292_dilations_0 = const()[name = string("op_5292_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5292_groups_0 = const()[name = string("op_5292_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5292 = conv(dilations = var_5292_dilations_0, groups = var_5292_groups_0, pad = var_5292_pad_0, pad_type = var_5292_pad_type_0, strides = var_5292_strides_0, weight = encoder_layers_26_self_attn_v_proj_weight, x = input_261)[name = string("op_5292")];
            tensor<int32, [4]> var_5293 = const()[name = string("op_5293"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5294 = reshape(shape = var_5293, x = var_5292)[name = string("op_5294")];
            tensor<int32, [4]> var_5295 = const()[name = string("op_5295"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_105_to_fp16 = const()[name = string("op_6_promoted_105_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_157 = transpose(perm = var_5275, x = var_5274)[name = string("transpose_16")];
            tensor<fp16, [1, 16, 512, 128]> var_5301_cast_fp16 = pow(x = q_157, y = var_6_promoted_105_to_fp16)[name = string("op_5301_cast_fp16")];
            tensor<int32, [1]> var_211_axes_0 = const()[name = string("var_211_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_211_keep_dims_0 = const()[name = string("var_211_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_211_cast_fp16 = reduce_mean(axes = var_211_axes_0, keep_dims = var_211_keep_dims_0, x = var_5301_cast_fp16)[name = string("var_211_cast_fp16")];
            fp16 var_5304_to_fp16 = const()[name = string("op_5304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_5305_cast_fp16 = add(x = var_211_cast_fp16, y = var_5304_to_fp16)[name = string("op_5305_cast_fp16")];
            fp32 var_5306_epsilon_0 = const()[name = string("op_5306_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_5306_cast_fp16 = rsqrt(epsilon = var_5306_epsilon_0, x = var_5305_cast_fp16)[name = string("op_5306_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_895_cast_fp16 = mul(x = q_157, y = var_5306_cast_fp16)[name = string("x_895_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_159 = mul(x = x_895_cast_fp16, y = encoder_layers_26_self_attn_q_norm_weight)[name = string("q_159")];
            fp16 var_6_promoted_106_to_fp16 = const()[name = string("op_6_promoted_106_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_157 = transpose(perm = var_5285, x = var_5284)[name = string("transpose_15")];
            tensor<fp16, [1, 8, 512, 128]> var_5314_cast_fp16 = pow(x = k_157, y = var_6_promoted_106_to_fp16)[name = string("op_5314_cast_fp16")];
            tensor<int32, [1]> var_213_axes_0 = const()[name = string("var_213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_213_keep_dims_0 = const()[name = string("var_213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_213_cast_fp16 = reduce_mean(axes = var_213_axes_0, keep_dims = var_213_keep_dims_0, x = var_5314_cast_fp16)[name = string("var_213_cast_fp16")];
            fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_5318_cast_fp16 = add(x = var_213_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")];
            fp32 var_5319_epsilon_0 = const()[name = string("op_5319_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_5319_cast_fp16 = rsqrt(epsilon = var_5319_epsilon_0, x = var_5318_cast_fp16)[name = string("op_5319_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_901_cast_fp16 = mul(x = k_157, y = var_5319_cast_fp16)[name = string("x_901_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_159 = mul(x = x_901_cast_fp16, y = encoder_layers_26_self_attn_k_norm_weight)[name = string("k_159")];
            tensor<fp16, [1, 16, 512, 128]> var_5323 = mul(x = q_159, y = cos)[name = string("op_5323")];
            tensor<int32, [2]> var_5324_split_sizes_0 = const()[name = string("op_5324_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5324_axis_0 = const()[name = string("op_5324_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_5324_0, tensor<fp16, [1, 16, 512, 64]> var_5324_1 = split(axis = var_5324_axis_0, split_sizes = var_5324_split_sizes_0, x = q_159)[name = string("op_5324")];
            fp16 const_81_promoted = const()[name = string("const_81_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_5326 = mul(x = var_5324_1, y = const_81_promoted)[name = string("op_5326")];
            bool var_5328_interleave_0 = const()[name = string("op_5328_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5328 = concat(axis = var_18, interleave = var_5328_interleave_0, values = (var_5326, var_5324_0))[name = string("op_5328")];
            tensor<fp16, [1, 16, 512, 128]> var_5329 = mul(x = var_5328, y = sin)[name = string("op_5329")];
            tensor<fp16, [1, 16, 512, 128]> query_53 = add(x = var_5323, y = var_5329)[name = string("query_53")];
            tensor<fp16, [1, 8, 512, 128]> var_5331 = mul(x = k_159, y = cos)[name = string("op_5331")];
            tensor<int32, [2]> var_5332_split_sizes_0 = const()[name = string("op_5332_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5332_axis_0 = const()[name = string("op_5332_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_5332_0, tensor<fp16, [1, 8, 512, 64]> var_5332_1 = split(axis = var_5332_axis_0, split_sizes = var_5332_split_sizes_0, x = k_159)[name = string("op_5332")];
            fp16 const_82_promoted = const()[name = string("const_82_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_5334 = mul(x = var_5332_1, y = const_82_promoted)[name = string("op_5334")];
            bool var_5336_interleave_0 = const()[name = string("op_5336_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_5336 = concat(axis = var_18, interleave = var_5336_interleave_0, values = (var_5334, var_5332_0))[name = string("op_5336")];
            tensor<fp16, [1, 8, 512, 128]> var_5337 = mul(x = var_5336, y = sin)[name = string("op_5337")];
            tensor<fp16, [1, 8, 512, 128]> x_903 = add(x = var_5331, y = var_5337)[name = string("x_903")];
            tensor<int32, [1]> var_5339_axes_0 = const()[name = string("op_5339_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5339 = expand_dims(axes = var_5339_axes_0, x = x_903)[name = string("op_5339")];
            tensor<int32, [5]> x_905_reps_0 = const()[name = string("x_905_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_905 = tile(reps = x_905_reps_0, x = var_5339)[name = string("x_905")];
            tensor<int32, [4]> var_5342 = const()[name = string("op_5342"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key_53 = reshape(shape = var_5342, x = x_905)[name = string("key_53")];
            tensor<int32, [1]> var_5344_axes_0 = const()[name = string("op_5344_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_907 = transpose(perm = var_5295, x = var_5294)[name = string("transpose_14")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5344 = expand_dims(axes = var_5344_axes_0, x = x_907)[name = string("op_5344")];
            tensor<int32, [5]> x_909_reps_0 = const()[name = string("x_909_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_909 = tile(reps = x_909_reps_0, x = var_5344)[name = string("x_909")];
            tensor<int32, [4]> var_5347 = const()[name = string("op_5347"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value_53 = reshape(shape = var_5347, x = x_909)[name = string("value_53")];
            bool var_5352_transpose_x_1 = const()[name = string("op_5352_transpose_x_1"), val = bool(false)];
            bool var_5352_transpose_y_1 = const()[name = string("op_5352_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_5352_cast_fp16 = matmul(transpose_x = var_5352_transpose_x_1, transpose_y = var_5352_transpose_y_1, x = query_53, y = key_53)[name = string("op_5352_cast_fp16")];
            fp16 var_5353_to_fp16 = const()[name = string("op_5353_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_157_cast_fp16 = mul(x = var_5352_cast_fp16, y = var_5353_to_fp16)[name = string("attn_weights_157_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_159_cast_fp16 = add(x = attn_weights_157_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_159_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_5357_cast_fp16 = softmax(axis = var_18, x = attn_weights_159_cast_fp16)[name = string("op_5357_cast_fp16")];
            bool var_5361_transpose_x_0 = const()[name = string("op_5361_transpose_x_0"), val = bool(false)];
            bool var_5361_transpose_y_0 = const()[name = string("op_5361_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5361_cast_fp16 = matmul(transpose_x = var_5361_transpose_x_0, transpose_y = var_5361_transpose_y_0, x = var_5357_cast_fp16, y = value_53)[name = string("op_5361_cast_fp16")];
            tensor<int32, [4]> var_5363 = const()[name = string("op_5363"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5366 = const()[name = string("op_5366"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_5364 = transpose(perm = var_5363, x = var_5361_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 512, 2048]> attn_out_159 = reshape(shape = var_5366, x = var_5364)[name = string("attn_out_159")];
            tensor<int32, [3]> var_5368 = const()[name = string("op_5368"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_26 = const()[name = string("squeeze_26"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1200499904)))];
            string var_5377_pad_type_0 = const()[name = string("op_5377_pad_type_0"), val = string("valid")];
            int32 var_5377_groups_0 = const()[name = string("op_5377_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5377_strides_0 = const()[name = string("op_5377_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5377_pad_0 = const()[name = string("op_5377_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5377_dilations_0 = const()[name = string("op_5377_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_5369 = transpose(perm = var_5368, x = attn_out_159)[name = string("transpose_12")];
            tensor<fp16, [1, 1024, 512]> var_5377 = conv(dilations = var_5377_dilations_0, groups = var_5377_groups_0, pad = var_5377_pad_0, pad_type = var_5377_pad_type_0, strides = var_5377_strides_0, weight = squeeze_26, x = var_5369)[name = string("op_5377")];
            tensor<int32, [3]> var_5378 = const()[name = string("op_5378"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out_161 = transpose(perm = var_5378, x = var_5377)[name = string("transpose_11")];
            tensor<fp16, [1, 512, 1024]> x_911_cast_fp16 = add(x = hidden_states_53_cast_fp16, y = attn_out_161)[name = string("x_911_cast_fp16")];
            fp16 var_6_promoted_107_to_fp16 = const()[name = string("op_6_promoted_107_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5384_cast_fp16 = pow(x = x_911_cast_fp16, y = var_6_promoted_107_to_fp16)[name = string("op_5384_cast_fp16")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = string("var_215_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_215_keep_dims_0 = const()[name = string("var_215_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_215_cast_fp16 = reduce_mean(axes = var_215_axes_0, keep_dims = var_215_keep_dims_0, x = var_5384_cast_fp16)[name = string("var_215_cast_fp16")];
            fp16 var_5387_to_fp16 = const()[name = string("op_5387_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5388_cast_fp16 = add(x = var_215_cast_fp16, y = var_5387_to_fp16)[name = string("op_5388_cast_fp16")];
            fp32 var_5389_epsilon_0 = const()[name = string("op_5389_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5389_cast_fp16 = rsqrt(epsilon = var_5389_epsilon_0, x = var_5388_cast_fp16)[name = string("op_5389_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_915_cast_fp16 = mul(x = x_911_cast_fp16, y = var_5389_cast_fp16)[name = string("x_915_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204694272)))];
            tensor<fp16, [1, 512, 1024]> var_5392_cast_fp16 = mul(x = x_915_cast_fp16, y = encoder_layers_26_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5392_cast_fp16")];
            tensor<int32, [3]> var_5397 = const()[name = string("op_5397"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_265_axes_0 = const()[name = string("input_265_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5398 = transpose(perm = var_5397, x = var_5392_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 1024, 1, 512]> input_265 = expand_dims(axes = input_265_axes_0, x = var_5398)[name = string("input_265")];
            string input_267_pad_type_0 = const()[name = string("input_267_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_267_strides_0 = const()[name = string("input_267_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_267_pad_0 = const()[name = string("input_267_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_267_dilations_0 = const()[name = string("input_267_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_267_groups_0 = const()[name = string("input_267_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_267 = conv(dilations = input_267_dilations_0, groups = input_267_groups_0, pad = input_267_pad_0, pad_type = input_267_pad_type_0, strides = input_267_strides_0, weight = encoder_layers_26_mlp_gate_proj_weight, x = input_265)[name = string("input_267")];
            string up_53_pad_type_0 = const()[name = string("up_53_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_53_strides_0 = const()[name = string("up_53_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_53_pad_0 = const()[name = string("up_53_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_53_dilations_0 = const()[name = string("up_53_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_53_groups_0 = const()[name = string("up_53_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up_53 = conv(dilations = up_53_dilations_0, groups = up_53_groups_0, pad = up_53_pad_0, pad_type = up_53_pad_type_0, strides = up_53_strides_0, weight = encoder_layers_26_mlp_up_proj_weight, x = input_265)[name = string("up_53")];
            tensor<fp16, [1, 3072, 1, 512]> var_5412 = silu(x = input_267)[name = string("op_5412")];
            tensor<fp16, [1, 3072, 1, 512]> input_269 = mul(x = var_5412, y = up_53)[name = string("input_269")];
            string var_5419_pad_type_0 = const()[name = string("op_5419_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5419_strides_0 = const()[name = string("op_5419_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5419_pad_0 = const()[name = string("op_5419_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5419_dilations_0 = const()[name = string("op_5419_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5419_groups_0 = const()[name = string("op_5419_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5419 = conv(dilations = var_5419_dilations_0, groups = var_5419_groups_0, pad = var_5419_pad_0, pad_type = var_5419_pad_type_0, strides = var_5419_strides_0, weight = encoder_layers_26_mlp_down_proj_weight, x = input_269)[name = string("op_5419")];
            tensor<int32, [1]> var_5420_axes_0 = const()[name = string("op_5420_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5420 = squeeze(axes = var_5420_axes_0, x = var_5419)[name = string("op_5420")];
            tensor<int32, [3]> var_5421 = const()[name = string("op_5421"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out_53 = transpose(perm = var_5421, x = var_5420)[name = string("transpose_9")];
            tensor<fp16, [1, 512, 1024]> hidden_states_cast_fp16 = add(x = x_911_cast_fp16, y = mlp_out_53)[name = string("hidden_states_cast_fp16")];
            fp16 var_6_promoted_108_to_fp16 = const()[name = string("op_6_promoted_108_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5448_cast_fp16 = pow(x = hidden_states_cast_fp16, y = var_6_promoted_108_to_fp16)[name = string("op_5448_cast_fp16")];
            tensor<int32, [1]> var_217_axes_0_0 = const()[name = string("var_217_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_217_keep_dims_0 = const()[name = string("var_217_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_217_cast_fp16 = reduce_mean(axes = var_217_axes_0_0, keep_dims = var_217_keep_dims_0, x = var_5448_cast_fp16)[name = string("var_217_cast_fp16")];
            fp16 var_5451_to_fp16 = const()[name = string("op_5451_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5452_cast_fp16 = add(x = var_217_cast_fp16, y = var_5451_to_fp16)[name = string("op_5452_cast_fp16")];
            fp32 var_5453_epsilon_0 = const()[name = string("op_5453_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5453_cast_fp16 = rsqrt(epsilon = var_5453_epsilon_0, x = var_5452_cast_fp16)[name = string("op_5453_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_921_cast_fp16 = mul(x = hidden_states_cast_fp16, y = var_5453_cast_fp16)[name = string("x_921_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_27_input_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_27_input_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204696384)))];
            tensor<fp16, [1, 512, 1024]> var_5456_cast_fp16 = mul(x = x_921_cast_fp16, y = encoder_layers_27_input_layernorm_weight_promoted_to_fp16)[name = string("op_5456_cast_fp16")];
            tensor<int32, [3]> var_5461 = const()[name = string("op_5461"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_271_axes_0 = const()[name = string("input_271_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5462 = transpose(perm = var_5461, x = var_5456_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 1024, 1, 512]> input_271 = expand_dims(axes = input_271_axes_0, x = var_5462)[name = string("input_271")];
            string var_5469_pad_type_0 = const()[name = string("op_5469_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5469_strides_0 = const()[name = string("op_5469_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5469_pad_0 = const()[name = string("op_5469_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5469_dilations_0 = const()[name = string("op_5469_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5469_groups_0 = const()[name = string("op_5469_groups_0"), val = int32(1)];
            tensor<fp16, [1, 2048, 1, 512]> var_5469 = conv(dilations = var_5469_dilations_0, groups = var_5469_groups_0, pad = var_5469_pad_0, pad_type = var_5469_pad_type_0, strides = var_5469_strides_0, weight = encoder_layers_27_self_attn_q_proj_weight, x = input_271)[name = string("op_5469")];
            tensor<int32, [4]> var_5470 = const()[name = string("op_5470"), val = tensor<int32, [4]>([1, 16, 128, 512])];
            tensor<fp16, [1, 16, 128, 512]> var_5471 = reshape(shape = var_5470, x = var_5469)[name = string("op_5471")];
            tensor<int32, [4]> var_5472 = const()[name = string("op_5472"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5479_pad_type_0 = const()[name = string("op_5479_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5479_strides_0 = const()[name = string("op_5479_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5479_pad_0 = const()[name = string("op_5479_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5479_dilations_0 = const()[name = string("op_5479_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5479_groups_0 = const()[name = string("op_5479_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5479 = conv(dilations = var_5479_dilations_0, groups = var_5479_groups_0, pad = var_5479_pad_0, pad_type = var_5479_pad_type_0, strides = var_5479_strides_0, weight = encoder_layers_27_self_attn_k_proj_weight, x = input_271)[name = string("op_5479")];
            tensor<int32, [4]> var_5480 = const()[name = string("op_5480"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5481 = reshape(shape = var_5480, x = var_5479)[name = string("op_5481")];
            tensor<int32, [4]> var_5482 = const()[name = string("op_5482"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            string var_5489_pad_type_0 = const()[name = string("op_5489_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5489_strides_0 = const()[name = string("op_5489_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5489_pad_0 = const()[name = string("op_5489_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5489_dilations_0 = const()[name = string("op_5489_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5489_groups_0 = const()[name = string("op_5489_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5489 = conv(dilations = var_5489_dilations_0, groups = var_5489_groups_0, pad = var_5489_pad_0, pad_type = var_5489_pad_type_0, strides = var_5489_strides_0, weight = encoder_layers_27_self_attn_v_proj_weight, x = input_271)[name = string("op_5489")];
            tensor<int32, [4]> var_5490 = const()[name = string("op_5490"), val = tensor<int32, [4]>([1, 8, 128, 512])];
            tensor<fp16, [1, 8, 128, 512]> var_5491 = reshape(shape = var_5490, x = var_5489)[name = string("op_5491")];
            tensor<int32, [4]> var_5492 = const()[name = string("op_5492"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            fp16 var_6_promoted_109_to_fp16 = const()[name = string("op_6_promoted_109_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 512, 128]> q_163 = transpose(perm = var_5472, x = var_5471)[name = string("transpose_7")];
            tensor<fp16, [1, 16, 512, 128]> var_5498_cast_fp16 = pow(x = q_163, y = var_6_promoted_109_to_fp16)[name = string("op_5498_cast_fp16")];
            tensor<int32, [1]> var_219_axes_0 = const()[name = string("var_219_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_219_keep_dims_0 = const()[name = string("var_219_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 1]> var_219_cast_fp16 = reduce_mean(axes = var_219_axes_0, keep_dims = var_219_keep_dims_0, x = var_5498_cast_fp16)[name = string("var_219_cast_fp16")];
            fp16 var_5501_to_fp16 = const()[name = string("op_5501_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 512, 1]> var_5502_cast_fp16 = add(x = var_219_cast_fp16, y = var_5501_to_fp16)[name = string("op_5502_cast_fp16")];
            fp32 var_5503_epsilon_0 = const()[name = string("op_5503_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 512, 1]> var_5503_cast_fp16 = rsqrt(epsilon = var_5503_epsilon_0, x = var_5502_cast_fp16)[name = string("op_5503_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> x_929_cast_fp16 = mul(x = q_163, y = var_5503_cast_fp16)[name = string("x_929_cast_fp16")];
            tensor<fp16, [1, 16, 512, 128]> q_165 = mul(x = x_929_cast_fp16, y = encoder_layers_27_self_attn_q_norm_weight)[name = string("q_165")];
            fp16 var_6_promoted_110_to_fp16 = const()[name = string("op_6_promoted_110_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 512, 128]> k_163 = transpose(perm = var_5482, x = var_5481)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 512, 128]> var_5511_cast_fp16 = pow(x = k_163, y = var_6_promoted_110_to_fp16)[name = string("op_5511_cast_fp16")];
            tensor<int32, [1]> var_221_axes_0 = const()[name = string("var_221_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_221_keep_dims_0 = const()[name = string("var_221_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 1]> var_221_cast_fp16 = reduce_mean(axes = var_221_axes_0, keep_dims = var_221_keep_dims_0, x = var_5511_cast_fp16)[name = string("var_221_cast_fp16")];
            fp16 var_5514_to_fp16 = const()[name = string("op_5514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 512, 1]> var_5515_cast_fp16 = add(x = var_221_cast_fp16, y = var_5514_to_fp16)[name = string("op_5515_cast_fp16")];
            fp32 var_5516_epsilon_0 = const()[name = string("op_5516_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 512, 1]> var_5516_cast_fp16 = rsqrt(epsilon = var_5516_epsilon_0, x = var_5515_cast_fp16)[name = string("op_5516_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> x_935_cast_fp16 = mul(x = k_163, y = var_5516_cast_fp16)[name = string("x_935_cast_fp16")];
            tensor<fp16, [1, 8, 512, 128]> k_165 = mul(x = x_935_cast_fp16, y = encoder_layers_27_self_attn_k_norm_weight)[name = string("k_165")];
            tensor<fp16, [1, 16, 512, 128]> var_5520 = mul(x = q_165, y = cos)[name = string("op_5520")];
            tensor<int32, [2]> var_5521_split_sizes_0 = const()[name = string("op_5521_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5521_axis_0 = const()[name = string("op_5521_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 16, 512, 64]> var_5521_0, tensor<fp16, [1, 16, 512, 64]> var_5521_1 = split(axis = var_5521_axis_0, split_sizes = var_5521_split_sizes_0, x = q_165)[name = string("op_5521")];
            fp16 const_84_promoted = const()[name = string("const_84_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 512, 64]> var_5523 = mul(x = var_5521_1, y = const_84_promoted)[name = string("op_5523")];
            bool var_5525_interleave_0 = const()[name = string("op_5525_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5525 = concat(axis = var_18, interleave = var_5525_interleave_0, values = (var_5523, var_5521_0))[name = string("op_5525")];
            tensor<fp16, [1, 16, 512, 128]> var_5526 = mul(x = var_5525, y = sin)[name = string("op_5526")];
            tensor<fp16, [1, 16, 512, 128]> query = add(x = var_5520, y = var_5526)[name = string("query")];
            tensor<fp16, [1, 8, 512, 128]> var_5528 = mul(x = k_165, y = cos)[name = string("op_5528")];
            tensor<int32, [2]> var_5529_split_sizes_0 = const()[name = string("op_5529_split_sizes_0"), val = tensor<int32, [2]>([64, 64])];
            int32 var_5529_axis_0 = const()[name = string("op_5529_axis_0"), val = int32(-1)];
            tensor<fp16, [1, 8, 512, 64]> var_5529_0, tensor<fp16, [1, 8, 512, 64]> var_5529_1 = split(axis = var_5529_axis_0, split_sizes = var_5529_split_sizes_0, x = k_165)[name = string("op_5529")];
            fp16 const_85_promoted = const()[name = string("const_85_promoted"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 8, 512, 64]> var_5531 = mul(x = var_5529_1, y = const_85_promoted)[name = string("op_5531")];
            bool var_5533_interleave_0 = const()[name = string("op_5533_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 512, 128]> var_5533 = concat(axis = var_18, interleave = var_5533_interleave_0, values = (var_5531, var_5529_0))[name = string("op_5533")];
            tensor<fp16, [1, 8, 512, 128]> var_5534 = mul(x = var_5533, y = sin)[name = string("op_5534")];
            tensor<fp16, [1, 8, 512, 128]> x_937 = add(x = var_5528, y = var_5534)[name = string("x_937")];
            tensor<int32, [1]> var_5536_axes_0 = const()[name = string("op_5536_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5536 = expand_dims(axes = var_5536_axes_0, x = x_937)[name = string("op_5536")];
            tensor<int32, [5]> x_939_reps_0 = const()[name = string("x_939_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_939 = tile(reps = x_939_reps_0, x = var_5536)[name = string("x_939")];
            tensor<int32, [4]> var_5539 = const()[name = string("op_5539"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> key = reshape(shape = var_5539, x = x_939)[name = string("key")];
            tensor<int32, [1]> var_5541_axes_0 = const()[name = string("op_5541_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 8, 512, 128]> x_941 = transpose(perm = var_5492, x = var_5491)[name = string("transpose_5")];
            tensor<fp16, [1, 8, 1, 512, 128]> var_5541 = expand_dims(axes = var_5541_axes_0, x = x_941)[name = string("op_5541")];
            tensor<int32, [5]> x_943_reps_0 = const()[name = string("x_943_reps_0"), val = tensor<int32, [5]>([1, 1, 2, 1, 1])];
            tensor<fp16, [1, 8, 2, 512, 128]> x_943 = tile(reps = x_943_reps_0, x = var_5541)[name = string("x_943")];
            tensor<int32, [4]> var_5544 = const()[name = string("op_5544"), val = tensor<int32, [4]>([1, 16, 512, 128])];
            tensor<fp16, [1, 16, 512, 128]> value = reshape(shape = var_5544, x = x_943)[name = string("value")];
            bool var_5549_transpose_x_1 = const()[name = string("op_5549_transpose_x_1"), val = bool(false)];
            bool var_5549_transpose_y_1 = const()[name = string("op_5549_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 512, 512]> var_5549_cast_fp16 = matmul(transpose_x = var_5549_transpose_x_1, transpose_y = var_5549_transpose_y_1, x = query, y = key)[name = string("op_5549_cast_fp16")];
            fp16 var_5550_to_fp16 = const()[name = string("op_5550_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_163_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = string("attn_weights_163_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> attn_weights_165_cast_fp16 = add(x = attn_weights_163_cast_fp16, y = causal_mask_cast_fp16)[name = string("attn_weights_165_cast_fp16")];
            tensor<fp16, [1, 16, 512, 512]> var_5554_cast_fp16 = softmax(axis = var_18, x = attn_weights_165_cast_fp16)[name = string("op_5554_cast_fp16")];
            bool var_5558_transpose_x_0 = const()[name = string("op_5558_transpose_x_0"), val = bool(false)];
            bool var_5558_transpose_y_0 = const()[name = string("op_5558_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 512, 128]> var_5558_cast_fp16 = matmul(transpose_x = var_5558_transpose_x_0, transpose_y = var_5558_transpose_y_0, x = var_5554_cast_fp16, y = value)[name = string("op_5558_cast_fp16")];
            tensor<int32, [4]> var_5560 = const()[name = string("op_5560"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5563 = const()[name = string("op_5563"), val = tensor<int32, [3]>([1, 512, 2048])];
            tensor<fp16, [1, 512, 16, 128]> var_5561 = transpose(perm = var_5560, x = var_5558_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 512, 2048]> attn_out_165 = reshape(shape = var_5563, x = var_5561)[name = string("attn_out_165")];
            tensor<int32, [3]> var_5565 = const()[name = string("op_5565"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1024, 2048, 1]> squeeze_27 = const()[name = string("squeeze_27"), val = tensor<fp16, [1024, 2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1204698496)))];
            string var_5574_pad_type_0 = const()[name = string("op_5574_pad_type_0"), val = string("valid")];
            int32 var_5574_groups_0 = const()[name = string("op_5574_groups_0"), val = int32(1)];
            tensor<int32, [1]> var_5574_strides_0 = const()[name = string("op_5574_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> var_5574_pad_0 = const()[name = string("op_5574_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> var_5574_dilations_0 = const()[name = string("op_5574_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 2048, 512]> var_5566 = transpose(perm = var_5565, x = attn_out_165)[name = string("transpose_3")];
            tensor<fp16, [1, 1024, 512]> var_5574 = conv(dilations = var_5574_dilations_0, groups = var_5574_groups_0, pad = var_5574_pad_0, pad_type = var_5574_pad_type_0, strides = var_5574_strides_0, weight = squeeze_27, x = var_5566)[name = string("op_5574")];
            tensor<int32, [3]> var_5575 = const()[name = string("op_5575"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> attn_out = transpose(perm = var_5575, x = var_5574)[name = string("transpose_2")];
            tensor<fp16, [1, 512, 1024]> x_945_cast_fp16 = add(x = hidden_states_cast_fp16, y = attn_out)[name = string("x_945_cast_fp16")];
            fp16 var_6_promoted_111_to_fp16 = const()[name = string("op_6_promoted_111_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5581_cast_fp16 = pow(x = x_945_cast_fp16, y = var_6_promoted_111_to_fp16)[name = string("op_5581_cast_fp16")];
            tensor<int32, [1]> var_223_axes_0 = const()[name = string("var_223_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_223_keep_dims_0 = const()[name = string("var_223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_223_cast_fp16 = reduce_mean(axes = var_223_axes_0, keep_dims = var_223_keep_dims_0, x = var_5581_cast_fp16)[name = string("var_223_cast_fp16")];
            fp16 var_5584_to_fp16 = const()[name = string("op_5584_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5585_cast_fp16 = add(x = var_223_cast_fp16, y = var_5584_to_fp16)[name = string("op_5585_cast_fp16")];
            fp32 var_5586_epsilon_0 = const()[name = string("op_5586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5586_cast_fp16 = rsqrt(epsilon = var_5586_epsilon_0, x = var_5585_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_949_cast_fp16 = mul(x = x_945_cast_fp16, y = var_5586_cast_fp16)[name = string("x_949_cast_fp16")];
            tensor<fp16, [1024]> encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16 = const()[name = string("encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208892864)))];
            tensor<fp16, [1, 512, 1024]> var_5589_cast_fp16 = mul(x = x_949_cast_fp16, y = encoder_layers_27_post_attention_layernorm_weight_promoted_to_fp16)[name = string("op_5589_cast_fp16")];
            tensor<int32, [3]> var_5594 = const()[name = string("op_5594"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<int32, [1]> input_275_axes_0 = const()[name = string("input_275_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5595 = transpose(perm = var_5594, x = var_5589_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 1024, 1, 512]> input_275 = expand_dims(axes = input_275_axes_0, x = var_5595)[name = string("input_275")];
            string input_277_pad_type_0 = const()[name = string("input_277_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_277_strides_0 = const()[name = string("input_277_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_277_pad_0 = const()[name = string("input_277_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_277_dilations_0 = const()[name = string("input_277_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_277_groups_0 = const()[name = string("input_277_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> input_277 = conv(dilations = input_277_dilations_0, groups = input_277_groups_0, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = input_277_strides_0, weight = encoder_layers_27_mlp_gate_proj_weight, x = input_275)[name = string("input_277")];
            string up_pad_type_0 = const()[name = string("up_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> up_strides_0 = const()[name = string("up_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> up_pad_0 = const()[name = string("up_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> up_dilations_0 = const()[name = string("up_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 up_groups_0 = const()[name = string("up_groups_0"), val = int32(1)];
            tensor<fp16, [1, 3072, 1, 512]> up = conv(dilations = up_dilations_0, groups = up_groups_0, pad = up_pad_0, pad_type = up_pad_type_0, strides = up_strides_0, weight = encoder_layers_27_mlp_up_proj_weight, x = input_275)[name = string("up")];
            tensor<fp16, [1, 3072, 1, 512]> var_5609 = silu(x = input_277)[name = string("op_5609")];
            tensor<fp16, [1, 3072, 1, 512]> input = mul(x = var_5609, y = up)[name = string("input")];
            string var_5616_pad_type_0 = const()[name = string("op_5616_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_5616_strides_0 = const()[name = string("op_5616_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_5616_pad_0 = const()[name = string("op_5616_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_5616_dilations_0 = const()[name = string("op_5616_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_5616_groups_0 = const()[name = string("op_5616_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1024, 1, 512]> var_5616 = conv(dilations = var_5616_dilations_0, groups = var_5616_groups_0, pad = var_5616_pad_0, pad_type = var_5616_pad_type_0, strides = var_5616_strides_0, weight = encoder_layers_27_mlp_down_proj_weight, x = input)[name = string("op_5616")];
            tensor<int32, [1]> var_5617_axes_0 = const()[name = string("op_5617_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 512]> var_5617 = squeeze(axes = var_5617_axes_0, x = var_5616)[name = string("op_5617")];
            tensor<int32, [3]> var_5618 = const()[name = string("op_5618"), val = tensor<int32, [3]>([0, 2, 1])];
            tensor<fp16, [1, 512, 1024]> mlp_out = transpose(perm = var_5618, x = var_5617)[name = string("transpose_0")];
            tensor<fp16, [1, 512, 1024]> x_953_cast_fp16 = add(x = x_945_cast_fp16, y = mlp_out)[name = string("x_953_cast_fp16")];
            fp16 var_6_promoted_112_to_fp16 = const()[name = string("op_6_promoted_112_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 512, 1024]> var_5624_cast_fp16 = pow(x = x_953_cast_fp16, y = var_6_promoted_112_to_fp16)[name = string("op_5624_cast_fp16")];
            tensor<int32, [1]> var_axes_0 = const()[name = string("var_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_keep_dims_0 = const()[name = string("var_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 512, 1]> var_cast_fp16 = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_5624_cast_fp16)[name = string("var_cast_fp16")];
            fp16 var_5627_to_fp16 = const()[name = string("op_5627_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 512, 1]> var_5628_cast_fp16 = add(x = var_cast_fp16, y = var_5627_to_fp16)[name = string("op_5628_cast_fp16")];
            fp32 var_5629_epsilon_0 = const()[name = string("op_5629_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 512, 1]> var_5629_cast_fp16 = rsqrt(epsilon = var_5629_epsilon_0, x = var_5628_cast_fp16)[name = string("op_5629_cast_fp16")];
            tensor<fp16, [1, 512, 1024]> x_957_cast_fp16 = mul(x = x_953_cast_fp16, y = var_5629_cast_fp16)[name = string("x_957_cast_fp16")];
            tensor<fp16, [1024]> encoder_norm_weight_promoted_to_fp16 = const()[name = string("encoder_norm_weight_promoted_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1208894976)))];
            tensor<fp16, [1, 512, 1024]> var_5632_cast_fp16 = mul(x = x_957_cast_fp16, y = encoder_norm_weight_promoted_to_fp16)[name = string("op_5632_cast_fp16")];
            tensor<int32, [1]> var_5635_axes_0 = const()[name = string("op_5635_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [512, 1024]> var_5635 = squeeze(axes = var_5635_axes_0, x = var_5632_cast_fp16)[name = string("op_5635")];
            bool pooled_transpose_x_0 = const()[name = string("pooled_transpose_x_0"), val = bool(false)];
            bool pooled_transpose_y_0 = const()[name = string("pooled_transpose_y_0"), val = bool(false)];
            tensor<fp16, [32, 1024]> pooled_cast_fp16 = matmul(transpose_x = pooled_transpose_x_0, transpose_y = pooled_transpose_y_0, x = pool_matrix, y = var_5635)[name = string("pooled_cast_fp16")];
            tensor<fp16, [32, 1024]> var_5647_cast_fp16 = tanh(x = pooled_cast_fp16)[name = string("op_5647_cast_fp16")];
            fp16 var_5648_to_fp16 = const()[name = string("op_5648_to_fp16"), val = fp16(0x1.fcp+6)];
            tensor<fp16, [32, 1024]> var_5649_cast_fp16 = mul(x = var_5647_cast_fp16, y = var_5648_to_fp16)[name = string("op_5649_cast_fp16")];
            tensor<fp16, [32, 1024]> var_5650_cast_fp16 = round(x = var_5649_cast_fp16)[name = string("op_5650_cast_fp16")];
            fp16 var_5651_promoted_to_fp16 = const()[name = string("op_5651_promoted_to_fp16"), val = fp16(-0x1p+7)];
            fp16 var_5652_promoted_to_fp16 = const()[name = string("op_5652_promoted_to_fp16"), val = fp16(0x1.fcp+6)];
            tensor<fp16, [32, 1024]> clip_0_cast_fp16 = clip(alpha = var_5651_promoted_to_fp16, beta = var_5652_promoted_to_fp16, x = var_5650_cast_fp16)[name = string("clip_0_cast_fp16")];
            string var_5658_dtype_0 = const()[name = string("op_5658_dtype_0"), val = string("int8")];
            tensor<int8, [32, 1024]> embedding = cast(dtype = var_5658_dtype_0, x = clip_0_cast_fp16)[name = string("cast_230")];
        } -> (embedding);
}