diff --git "a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil" "b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil"
--- "a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil"
+++ "b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil"
@@ -1,234 +1,248 @@
 program(1.0)
 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
 {
-    func main<ios17>(tensor<fp32, [1, 256, 18]> cnn_window, tensor<fp32, [2, 6, 4, 64, 64]> dec_kv, tensor<fp32, [2, 1]> dec_scale, tensor<fp32, [4, 1, 16, 256]> enc_conv_cache, tensor<fp32, [4, 1, 4, 64, 64]> enc_kv, tensor<fp32, [4, 1]> enc_scale, tensor<fp32, [1, 2, 345]> features, tensor<fp32, [2]> valid_mask) {
-            tensor<fp32, [256]> encoder_cnn_bias = const()[name = tensor<string, []>("encoder_cnn_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
-            tensor<fp32, [256, 256, 19]> encoder_cnn_weight = const()[name = tensor<string, []>("encoder_cnn_weight"), val = tensor<fp32, [256, 256, 19]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1152)))];
-            tensor<fp32, [2, 2]> encoder__causal_mask = const()[name = tensor<string, []>("encoder__causal_mask"), val = tensor<fp32, [2, 2]>([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])];
-            tensor<fp32, [2]> encoder__t_index = const()[name = tensor<string, []>("encoder__t_index"), val = tensor<fp32, [2]>([0x1p+0, 0x1p+1])];
-            tensor<fp32, [256]> encoder_input_projection_linear_bias = const()[name = tensor<string, []>("encoder_input_projection_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4981952)))];
-            tensor<fp32, [256, 345]> encoder_input_projection_linear_weight = const()[name = tensor<string, []>("encoder_input_projection_linear_weight"), val = tensor<fp32, [256, 345]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4983040)))];
-            tensor<fp32, [256]> encoder_pre_layer_norm_bias = const()[name = tensor<string, []>("encoder_pre_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5336384)))];
-            tensor<fp32, [256]> encoder_pre_layer_norm_weight = const()[name = tensor<string, []>("encoder_pre_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5337472)))];
-            tensor<fp32, [256]> encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5338560)))];
-            tensor<fp32, [256]> encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5339648)))];
-            tensor<fp32, [1024]> encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5340736)))];
-            tensor<fp32, [1024, 256]> encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5344896)))];
-            tensor<fp32, [256]> encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6393536)))];
-            tensor<fp32, [256, 1024]> encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6394624)))];
-            tensor<fp32, [256]> encoder_ret_lns_0_bias = const()[name = tensor<string, []>("encoder_ret_lns_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7443264)))];
-            tensor<fp32, [256]> encoder_ret_lns_0_weight = const()[name = tensor<string, []>("encoder_ret_lns_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7444352)))];
-            tensor<fp32, [256]> encoder_q_proj_0_bias = const()[name = tensor<string, []>("encoder_q_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7445440)))];
-            tensor<fp32, [256, 256]> encoder_q_proj_0_weight = const()[name = tensor<string, []>("encoder_q_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7446528)))];
-            tensor<fp32, [256]> encoder_k_proj_0_bias = const()[name = tensor<string, []>("encoder_k_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7708736)))];
-            tensor<fp32, [256, 256]> encoder_k_proj_0_weight = const()[name = tensor<string, []>("encoder_k_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7709824)))];
-            tensor<fp32, [256]> encoder_v_proj_0_bias = const()[name = tensor<string, []>("encoder_v_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7972032)))];
-            tensor<fp32, [256, 256]> encoder_v_proj_0_weight = const()[name = tensor<string, []>("encoder_v_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7973120)))];
-            tensor<fp32, [256]> encoder_g_proj_0_bias = const()[name = tensor<string, []>("encoder_g_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8235328)))];
-            tensor<fp32, [256, 256]> encoder_g_proj_0_weight = const()[name = tensor<string, []>("encoder_g_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8236416)))];
-            tensor<fp32, [256]> encoder_out_proj_0_bias = const()[name = tensor<string, []>("encoder_out_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8498624)))];
-            tensor<fp32, [256, 256]> encoder_out_proj_0_weight = const()[name = tensor<string, []>("encoder_out_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8499712)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_0_bias = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8761920)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_0_weight = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8763008)))];
-            tensor<fp32, [512]> encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8764096)))];
-            tensor<fp32, [512, 256, 1]> encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8766208)))];
-            tensor<fp32, [256, 1, 16]> encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9290560)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_5_running_var = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9307008)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9308096)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_5_bias = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9309184)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_5_weight = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9310272)))];
-            tensor<fp32, [256]> encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9311360)))];
-            tensor<fp32, [256, 256, 1]> encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9312448)))];
-            tensor<fp32, [256]> encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9574656)))];
-            tensor<fp32, [256]> encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9575744)))];
-            tensor<fp32, [1024]> encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9576832)))];
-            tensor<fp32, [1024, 256]> encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9580992)))];
-            tensor<fp32, [256]> encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10629632)))];
-            tensor<fp32, [256, 1024]> encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10630720)))];
-            tensor<fp32, [256]> encoder_layer_norm_0_bias = const()[name = tensor<string, []>("encoder_layer_norm_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11679360)))];
-            tensor<fp32, [256]> encoder_layer_norm_0_weight = const()[name = tensor<string, []>("encoder_layer_norm_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11680448)))];
-            tensor<fp32, [256]> encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11681536)))];
-            tensor<fp32, [256]> encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11682624)))];
-            tensor<fp32, [1024]> encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11683712)))];
-            tensor<fp32, [1024, 256]> encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11687872)))];
-            tensor<fp32, [256]> encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12736512)))];
-            tensor<fp32, [256, 1024]> encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12737600)))];
-            tensor<fp32, [256]> encoder_ret_lns_1_bias = const()[name = tensor<string, []>("encoder_ret_lns_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13786240)))];
-            tensor<fp32, [256]> encoder_ret_lns_1_weight = const()[name = tensor<string, []>("encoder_ret_lns_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13787328)))];
-            tensor<fp32, [256]> encoder_q_proj_1_bias = const()[name = tensor<string, []>("encoder_q_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13788416)))];
-            tensor<fp32, [256, 256]> encoder_q_proj_1_weight = const()[name = tensor<string, []>("encoder_q_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13789504)))];
-            tensor<fp32, [256]> encoder_k_proj_1_bias = const()[name = tensor<string, []>("encoder_k_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14051712)))];
-            tensor<fp32, [256, 256]> encoder_k_proj_1_weight = const()[name = tensor<string, []>("encoder_k_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14052800)))];
-            tensor<fp32, [256]> encoder_v_proj_1_bias = const()[name = tensor<string, []>("encoder_v_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14315008)))];
-            tensor<fp32, [256, 256]> encoder_v_proj_1_weight = const()[name = tensor<string, []>("encoder_v_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14316096)))];
-            tensor<fp32, [256]> encoder_g_proj_1_bias = const()[name = tensor<string, []>("encoder_g_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14578304)))];
-            tensor<fp32, [256, 256]> encoder_g_proj_1_weight = const()[name = tensor<string, []>("encoder_g_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14579392)))];
-            tensor<fp32, [256]> encoder_out_proj_1_bias = const()[name = tensor<string, []>("encoder_out_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14841600)))];
-            tensor<fp32, [256, 256]> encoder_out_proj_1_weight = const()[name = tensor<string, []>("encoder_out_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14842688)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_0_bias = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15104896)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_0_weight = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15105984)))];
-            tensor<fp32, [512]> encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15107072)))];
-            tensor<fp32, [512, 256, 1]> encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15109184)))];
-            tensor<fp32, [256, 1, 16]> encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15633536)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_5_running_var = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15649984)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15651072)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_5_bias = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15652160)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_5_weight = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15653248)))];
-            tensor<fp32, [256]> encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15654336)))];
-            tensor<fp32, [256, 256, 1]> encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15655424)))];
-            tensor<fp32, [256]> encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15917632)))];
-            tensor<fp32, [256]> encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15918720)))];
-            tensor<fp32, [1024]> encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15919808)))];
-            tensor<fp32, [1024, 256]> encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15923968)))];
-            tensor<fp32, [256]> encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16972608)))];
-            tensor<fp32, [256, 1024]> encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16973696)))];
-            tensor<fp32, [256]> encoder_layer_norm_1_bias = const()[name = tensor<string, []>("encoder_layer_norm_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18022336)))];
-            tensor<fp32, [256]> encoder_layer_norm_1_weight = const()[name = tensor<string, []>("encoder_layer_norm_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18023424)))];
-            tensor<fp32, [256]> encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18024512)))];
-            tensor<fp32, [256]> encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18025600)))];
-            tensor<fp32, [1024]> encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18026688)))];
-            tensor<fp32, [1024, 256]> encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18030848)))];
-            tensor<fp32, [256]> encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19079488)))];
-            tensor<fp32, [256, 1024]> encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19080576)))];
-            tensor<fp32, [256]> encoder_ret_lns_2_bias = const()[name = tensor<string, []>("encoder_ret_lns_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20129216)))];
-            tensor<fp32, [256]> encoder_ret_lns_2_weight = const()[name = tensor<string, []>("encoder_ret_lns_2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20130304)))];
-            tensor<fp32, [256]> encoder_q_proj_2_bias = const()[name = tensor<string, []>("encoder_q_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20131392)))];
-            tensor<fp32, [256, 256]> encoder_q_proj_2_weight = const()[name = tensor<string, []>("encoder_q_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20132480)))];
-            tensor<fp32, [256]> encoder_k_proj_2_bias = const()[name = tensor<string, []>("encoder_k_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20394688)))];
-            tensor<fp32, [256, 256]> encoder_k_proj_2_weight = const()[name = tensor<string, []>("encoder_k_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20395776)))];
-            tensor<fp32, [256]> encoder_v_proj_2_bias = const()[name = tensor<string, []>("encoder_v_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20657984)))];
-            tensor<fp32, [256, 256]> encoder_v_proj_2_weight = const()[name = tensor<string, []>("encoder_v_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20659072)))];
-            tensor<fp32, [256]> encoder_g_proj_2_bias = const()[name = tensor<string, []>("encoder_g_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20921280)))];
-            tensor<fp32, [256, 256]> encoder_g_proj_2_weight = const()[name = tensor<string, []>("encoder_g_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20922368)))];
-            tensor<fp32, [256]> encoder_out_proj_2_bias = const()[name = tensor<string, []>("encoder_out_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21184576)))];
-            tensor<fp32, [256, 256]> encoder_out_proj_2_weight = const()[name = tensor<string, []>("encoder_out_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21185664)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_0_bias = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21447872)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_0_weight = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21448960)))];
-            tensor<fp32, [512]> encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21450048)))];
-            tensor<fp32, [512, 256, 1]> encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21452160)))];
-            tensor<fp32, [256, 1, 16]> encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21976512)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_5_running_var = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21992960)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21994048)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_5_bias = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21995136)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_5_weight = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21996224)))];
-            tensor<fp32, [256]> encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21997312)))];
-            tensor<fp32, [256, 256, 1]> encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21998400)))];
-            tensor<fp32, [256]> encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22260608)))];
-            tensor<fp32, [256]> encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22261696)))];
-            tensor<fp32, [1024]> encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22262784)))];
-            tensor<fp32, [1024, 256]> encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22266944)))];
-            tensor<fp32, [256]> encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23315584)))];
-            tensor<fp32, [256, 1024]> encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23316672)))];
-            tensor<fp32, [256]> encoder_layer_norm_2_bias = const()[name = tensor<string, []>("encoder_layer_norm_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24365312)))];
-            tensor<fp32, [256]> encoder_layer_norm_2_weight = const()[name = tensor<string, []>("encoder_layer_norm_2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24366400)))];
-            tensor<fp32, [256]> encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24367488)))];
-            tensor<fp32, [256]> encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24368576)))];
-            tensor<fp32, [1024]> encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24369664)))];
-            tensor<fp32, [1024, 256]> encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24373824)))];
-            tensor<fp32, [256]> encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25422464)))];
-            tensor<fp32, [256, 1024]> encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25423552)))];
-            tensor<fp32, [256]> encoder_ret_lns_3_bias = const()[name = tensor<string, []>("encoder_ret_lns_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26472192)))];
-            tensor<fp32, [256]> encoder_ret_lns_3_weight = const()[name = tensor<string, []>("encoder_ret_lns_3_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26473280)))];
-            tensor<fp32, [256]> encoder_q_proj_3_bias = const()[name = tensor<string, []>("encoder_q_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26474368)))];
-            tensor<fp32, [256, 256]> encoder_q_proj_3_weight = const()[name = tensor<string, []>("encoder_q_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26475456)))];
-            tensor<fp32, [256]> encoder_k_proj_3_bias = const()[name = tensor<string, []>("encoder_k_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26737664)))];
-            tensor<fp32, [256, 256]> encoder_k_proj_3_weight = const()[name = tensor<string, []>("encoder_k_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26738752)))];
-            tensor<fp32, [256]> encoder_v_proj_3_bias = const()[name = tensor<string, []>("encoder_v_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27000960)))];
-            tensor<fp32, [256, 256]> encoder_v_proj_3_weight = const()[name = tensor<string, []>("encoder_v_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27002048)))];
-            tensor<fp32, [256]> encoder_g_proj_3_bias = const()[name = tensor<string, []>("encoder_g_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27264256)))];
-            tensor<fp32, [256, 256]> encoder_g_proj_3_weight = const()[name = tensor<string, []>("encoder_g_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27265344)))];
-            tensor<fp32, [256]> encoder_out_proj_3_bias = const()[name = tensor<string, []>("encoder_out_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27527552)))];
-            tensor<fp32, [256, 256]> encoder_out_proj_3_weight = const()[name = tensor<string, []>("encoder_out_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27528640)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_0_bias = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27790848)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_0_weight = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27791936)))];
-            tensor<fp32, [512]> encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27793024)))];
-            tensor<fp32, [512, 256, 1]> encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27795136)))];
-            tensor<fp32, [256, 1, 16]> encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28319488)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_5_running_var = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28335936)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28337024)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_5_bias = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28338112)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_5_weight = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28339200)))];
-            tensor<fp32, [256]> encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28340288)))];
-            tensor<fp32, [256, 256, 1]> encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor<string, []>("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28341376)))];
-            tensor<fp32, [256]> encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28603584)))];
-            tensor<fp32, [256]> encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28604672)))];
-            tensor<fp32, [1024]> encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28605760)))];
-            tensor<fp32, [1024, 256]> encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28609920)))];
-            tensor<fp32, [256]> encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29658560)))];
-            tensor<fp32, [256, 1024]> encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor<string, []>("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29659648)))];
-            tensor<fp32, [256]> encoder_layer_norm_3_bias = const()[name = tensor<string, []>("encoder_layer_norm_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30708288)))];
-            tensor<fp32, [256]> encoder_layer_norm_3_weight = const()[name = tensor<string, []>("encoder_layer_norm_3_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30709376)))];
-            tensor<fp32, [2, 2]> decoder__causal_mask = const()[name = tensor<string, []>("decoder__causal_mask"), val = tensor<fp32, [2, 2]>([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])];
-            tensor<fp32, [256]> decoder_convert_bias = const()[name = tensor<string, []>("decoder_convert_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30710464)))];
-            tensor<fp32, [256, 512]> decoder_convert_weight = const()[name = tensor<string, []>("decoder_convert_weight"), val = tensor<fp32, [256, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30711552)))];
-            tensor<fp32, [256]> decoder_q_proj_0_bias = const()[name = tensor<string, []>("decoder_q_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31235904)))];
-            tensor<fp32, [256, 256]> decoder_q_proj_0_weight = const()[name = tensor<string, []>("decoder_q_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31236992)))];
-            tensor<fp32, [256]> decoder_k_proj_0_bias = const()[name = tensor<string, []>("decoder_k_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31499200)))];
-            tensor<fp32, [256, 256]> decoder_k_proj_0_weight = const()[name = tensor<string, []>("decoder_k_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31500288)))];
-            tensor<fp32, [256]> decoder_v_proj_0_bias = const()[name = tensor<string, []>("decoder_v_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31762496)))];
-            tensor<fp32, [256, 256]> decoder_v_proj_0_weight = const()[name = tensor<string, []>("decoder_v_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31763584)))];
-            tensor<fp32, [256]> decoder_g_proj_0_bias = const()[name = tensor<string, []>("decoder_g_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32025792)))];
-            tensor<fp32, [256, 256]> decoder_g_proj_0_weight = const()[name = tensor<string, []>("decoder_g_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32026880)))];
-            tensor<fp32, [256]> decoder_out_proj_0_bias = const()[name = tensor<string, []>("decoder_out_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32289088)))];
-            tensor<fp32, [256, 256]> decoder_out_proj_0_weight = const()[name = tensor<string, []>("decoder_out_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32290176)))];
-            tensor<fp32, [256]> decoder_norm11_0_bias = const()[name = tensor<string, []>("decoder_norm11_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32552384)))];
-            tensor<fp32, [256]> decoder_norm11_0_weight = const()[name = tensor<string, []>("decoder_norm11_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32553472)))];
-            tensor<fp32, [256]> decoder_self_attn2_0_out_proj_bias = const()[name = tensor<string, []>("decoder_self_attn2_0_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32554560)))];
-            tensor<fp32, [256, 256]> decoder_self_attn2_0_out_proj_weight = const()[name = tensor<string, []>("decoder_self_attn2_0_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32555648)))];
-            tensor<fp32, [768]> decoder_self_attn2_0_in_proj_bias = const()[name = tensor<string, []>("decoder_self_attn2_0_in_proj_bias"), val = tensor<fp32, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32817856)))];
-            tensor<fp32, [768, 256]> decoder_self_attn2_0_in_proj_weight = const()[name = tensor<string, []>("decoder_self_attn2_0_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32820992)))];
-            tensor<fp32, [256]> decoder_norm21_0_bias = const()[name = tensor<string, []>("decoder_norm21_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33607488)))];
-            tensor<fp32, [256]> decoder_norm21_0_weight = const()[name = tensor<string, []>("decoder_norm21_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33608576)))];
-            tensor<fp32, [2048]> decoder_linear1_0_bias = const()[name = tensor<string, []>("decoder_linear1_0_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33609664)))];
-            tensor<fp32, [2048, 256]> decoder_linear1_0_weight = const()[name = tensor<string, []>("decoder_linear1_0_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33617920)))];
-            tensor<fp32, [256]> decoder_linear2_0_bias = const()[name = tensor<string, []>("decoder_linear2_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35715136)))];
-            tensor<fp32, [256, 2048]> decoder_linear2_0_weight = const()[name = tensor<string, []>("decoder_linear2_0_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35716224)))];
-            tensor<fp32, [256]> decoder_norm22_0_bias = const()[name = tensor<string, []>("decoder_norm22_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37813440)))];
-            tensor<fp32, [256]> decoder_norm22_0_weight = const()[name = tensor<string, []>("decoder_norm22_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37814528)))];
-            tensor<fp32, [256]> decoder_q_proj_1_bias = const()[name = tensor<string, []>("decoder_q_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37815616)))];
-            tensor<fp32, [256, 256]> decoder_q_proj_1_weight = const()[name = tensor<string, []>("decoder_q_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37816704)))];
-            tensor<fp32, [256]> decoder_k_proj_1_bias = const()[name = tensor<string, []>("decoder_k_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38078912)))];
-            tensor<fp32, [256, 256]> decoder_k_proj_1_weight = const()[name = tensor<string, []>("decoder_k_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38080000)))];
-            tensor<fp32, [256]> decoder_v_proj_1_bias = const()[name = tensor<string, []>("decoder_v_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38342208)))];
-            tensor<fp32, [256, 256]> decoder_v_proj_1_weight = const()[name = tensor<string, []>("decoder_v_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38343296)))];
-            tensor<fp32, [256]> decoder_g_proj_1_bias = const()[name = tensor<string, []>("decoder_g_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38605504)))];
-            tensor<fp32, [256, 256]> decoder_g_proj_1_weight = const()[name = tensor<string, []>("decoder_g_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38606592)))];
-            tensor<fp32, [256]> decoder_out_proj_1_bias = const()[name = tensor<string, []>("decoder_out_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38868800)))];
-            tensor<fp32, [256, 256]> decoder_out_proj_1_weight = const()[name = tensor<string, []>("decoder_out_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38869888)))];
-            tensor<fp32, [256]> decoder_norm11_1_bias = const()[name = tensor<string, []>("decoder_norm11_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39132096)))];
-            tensor<fp32, [256]> decoder_norm11_1_weight = const()[name = tensor<string, []>("decoder_norm11_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39133184)))];
-            tensor<fp32, [256]> decoder_self_attn2_1_out_proj_bias = const()[name = tensor<string, []>("decoder_self_attn2_1_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39134272)))];
-            tensor<fp32, [256, 256]> decoder_self_attn2_1_out_proj_weight = const()[name = tensor<string, []>("decoder_self_attn2_1_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39135360)))];
-            tensor<fp32, [768]> decoder_self_attn2_1_in_proj_bias = const()[name = tensor<string, []>("decoder_self_attn2_1_in_proj_bias"), val = tensor<fp32, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39397568)))];
-            tensor<fp32, [768, 256]> decoder_self_attn2_1_in_proj_weight = const()[name = tensor<string, []>("decoder_self_attn2_1_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39400704)))];
-            tensor<fp32, [256]> decoder_norm21_1_bias = const()[name = tensor<string, []>("decoder_norm21_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40187200)))];
-            tensor<fp32, [256]> decoder_norm21_1_weight = const()[name = tensor<string, []>("decoder_norm21_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40188288)))];
-            tensor<fp32, [2048]> decoder_linear1_1_bias = const()[name = tensor<string, []>("decoder_linear1_1_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40189376)))];
-            tensor<fp32, [2048, 256]> decoder_linear1_1_weight = const()[name = tensor<string, []>("decoder_linear1_1_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40197632)))];
-            tensor<fp32, [256]> decoder_linear2_1_bias = const()[name = tensor<string, []>("decoder_linear2_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42294848)))];
-            tensor<fp32, [256, 2048]> decoder_linear2_1_weight = const()[name = tensor<string, []>("decoder_linear2_1_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42295936)))];
-            tensor<fp32, [256]> decoder_norm22_1_bias = const()[name = tensor<string, []>("decoder_norm22_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44393152)))];
-            tensor<fp32, [256]> decoder_norm22_1_weight = const()[name = tensor<string, []>("decoder_norm22_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44394240)))];
-            tensor<fp32, []> var_11 = const()[name = tensor<string, []>("op_11"), val = tensor<fp32, []>(0x1.5798eep-27)];
-            tensor<fp32, []> var_17 = const()[name = tensor<string, []>("op_17"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
-            tensor<int32, []> var_20 = const()[name = tensor<string, []>("op_20"), val = tensor<int32, []>(2)];
-            tensor<int32, []> var_23 = const()[name = tensor<string, []>("op_23"), val = tensor<int32, []>(0)];
-            tensor<int32, []> var_26 = const()[name = tensor<string, []>("op_26"), val = tensor<int32, []>(1)];
-            tensor<fp32, []> var_28 = const()[name = tensor<string, []>("op_28"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
-            tensor<bool, []> var_29 = const()[name = tensor<string, []>("op_29"), val = tensor<bool, []>(true)];
-            tensor<fp32, [1, 2, 256]> input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor<string, []>("linear_0")];
-            tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor<string, []>("input_3")];
+    func main<ios17>(tensor<fp32, [1, 256, 18]> cnn_window, tensor<fp32, [2, 6, 4, 64, 64]> dec_kv, tensor<fp32, [2, 1]> dec_scale, tensor<fp32, [4, 1, 16, 256]> enc_conv_cache, tensor<fp32, [4, 1, 4, 64, 64]> enc_kv, tensor<fp32, [4, 1]> enc_scale, tensor<fp32, [1, 25, 23]> features, tensor<fp32, [2]> valid_mask) {
+            tensor<fp32, [256]> inner_encoder_cnn_bias = const()[name = tensor<string, []>("inner_encoder_cnn_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [256, 256, 19]> inner_encoder_cnn_weight = const()[name = tensor<string, []>("inner_encoder_cnn_weight"), val = tensor<fp32, [256, 256, 19]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1152)))];
+            tensor<fp32, [2, 2]> inner_encoder__causal_mask = const()[name = tensor<string, []>("inner_encoder__causal_mask"), val = tensor<fp32, [2, 2]>([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])];
+            tensor<fp32, [2]> inner_encoder__t_index = const()[name = tensor<string, []>("inner_encoder__t_index"), val = tensor<fp32, [2]>([0x1p+0, 0x1p+1])];
+            tensor<fp32, [256]> inner_encoder_input_projection_linear_bias = const()[name = tensor<string, []>("inner_encoder_input_projection_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4981952)))];
+            tensor<fp32, [256, 345]> inner_encoder_input_projection_linear_weight = const()[name = tensor<string, []>("inner_encoder_input_projection_linear_weight"), val = tensor<fp32, [256, 345]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4983040)))];
+            tensor<fp32, [256]> inner_encoder_pre_layer_norm_bias = const()[name = tensor<string, []>("inner_encoder_pre_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5336384)))];
+            tensor<fp32, [256]> inner_encoder_pre_layer_norm_weight = const()[name = tensor<string, []>("inner_encoder_pre_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5337472)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5338560)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5339648)))];
+            tensor<fp32, [1024]> inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5340736)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5344896)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6393536)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6394624)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_0_bias = const()[name = tensor<string, []>("inner_encoder_ret_lns_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7443264)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_0_weight = const()[name = tensor<string, []>("inner_encoder_ret_lns_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7444352)))];
+            tensor<fp32, [256]> inner_encoder_q_proj_0_bias = const()[name = tensor<string, []>("inner_encoder_q_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7445440)))];
+            tensor<fp32, [256, 256]> inner_encoder_q_proj_0_weight = const()[name = tensor<string, []>("inner_encoder_q_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7446528)))];
+            tensor<fp32, [256]> inner_encoder_k_proj_0_bias = const()[name = tensor<string, []>("inner_encoder_k_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7708736)))];
+            tensor<fp32, [256, 256]> inner_encoder_k_proj_0_weight = const()[name = tensor<string, []>("inner_encoder_k_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7709824)))];
+            tensor<fp32, [256]> inner_encoder_v_proj_0_bias = const()[name = tensor<string, []>("inner_encoder_v_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7972032)))];
+            tensor<fp32, [256, 256]> inner_encoder_v_proj_0_weight = const()[name = tensor<string, []>("inner_encoder_v_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7973120)))];
+            tensor<fp32, [256]> inner_encoder_g_proj_0_bias = const()[name = tensor<string, []>("inner_encoder_g_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8235328)))];
+            tensor<fp32, [256, 256]> inner_encoder_g_proj_0_weight = const()[name = tensor<string, []>("inner_encoder_g_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8236416)))];
+            tensor<fp32, [256]> inner_encoder_out_proj_0_bias = const()[name = tensor<string, []>("inner_encoder_out_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8498624)))];
+            tensor<fp32, [256, 256]> inner_encoder_out_proj_0_weight = const()[name = tensor<string, []>("inner_encoder_out_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8499712)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8761920)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8763008)))];
+            tensor<fp32, [512]> inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8764096)))];
+            tensor<fp32, [512, 256, 1]> inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8766208)))];
+            tensor<fp32, [256, 1, 16]> inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9290560)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9307008)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9308096)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9309184)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9310272)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9311360)))];
+            tensor<fp32, [256, 256, 1]> inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9312448)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9574656)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9575744)))];
+            tensor<fp32, [1024]> inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9576832)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9580992)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10629632)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10630720)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_0_bias = const()[name = tensor<string, []>("inner_encoder_layer_norm_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11679360)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_0_weight = const()[name = tensor<string, []>("inner_encoder_layer_norm_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11680448)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11681536)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11682624)))];
+            tensor<fp32, [1024]> inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11683712)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11687872)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12736512)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12737600)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_1_bias = const()[name = tensor<string, []>("inner_encoder_ret_lns_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13786240)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_1_weight = const()[name = tensor<string, []>("inner_encoder_ret_lns_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13787328)))];
+            tensor<fp32, [256]> inner_encoder_q_proj_1_bias = const()[name = tensor<string, []>("inner_encoder_q_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13788416)))];
+            tensor<fp32, [256, 256]> inner_encoder_q_proj_1_weight = const()[name = tensor<string, []>("inner_encoder_q_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13789504)))];
+            tensor<fp32, [256]> inner_encoder_k_proj_1_bias = const()[name = tensor<string, []>("inner_encoder_k_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14051712)))];
+            tensor<fp32, [256, 256]> inner_encoder_k_proj_1_weight = const()[name = tensor<string, []>("inner_encoder_k_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14052800)))];
+            tensor<fp32, [256]> inner_encoder_v_proj_1_bias = const()[name = tensor<string, []>("inner_encoder_v_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14315008)))];
+            tensor<fp32, [256, 256]> inner_encoder_v_proj_1_weight = const()[name = tensor<string, []>("inner_encoder_v_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14316096)))];
+            tensor<fp32, [256]> inner_encoder_g_proj_1_bias = const()[name = tensor<string, []>("inner_encoder_g_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14578304)))];
+            tensor<fp32, [256, 256]> inner_encoder_g_proj_1_weight = const()[name = tensor<string, []>("inner_encoder_g_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14579392)))];
+            tensor<fp32, [256]> inner_encoder_out_proj_1_bias = const()[name = tensor<string, []>("inner_encoder_out_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14841600)))];
+            tensor<fp32, [256, 256]> inner_encoder_out_proj_1_weight = const()[name = tensor<string, []>("inner_encoder_out_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14842688)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15104896)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15105984)))];
+            tensor<fp32, [512]> inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15107072)))];
+            tensor<fp32, [512, 256, 1]> inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15109184)))];
+            tensor<fp32, [256, 1, 16]> inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15633536)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15649984)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15651072)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15652160)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15653248)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15654336)))];
+            tensor<fp32, [256, 256, 1]> inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15655424)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15917632)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15918720)))];
+            tensor<fp32, [1024]> inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15919808)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15923968)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16972608)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16973696)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_1_bias = const()[name = tensor<string, []>("inner_encoder_layer_norm_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18022336)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_1_weight = const()[name = tensor<string, []>("inner_encoder_layer_norm_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18023424)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18024512)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18025600)))];
+            tensor<fp32, [1024]> inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18026688)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18030848)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19079488)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19080576)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_2_bias = const()[name = tensor<string, []>("inner_encoder_ret_lns_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20129216)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_2_weight = const()[name = tensor<string, []>("inner_encoder_ret_lns_2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20130304)))];
+            tensor<fp32, [256]> inner_encoder_q_proj_2_bias = const()[name = tensor<string, []>("inner_encoder_q_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20131392)))];
+            tensor<fp32, [256, 256]> inner_encoder_q_proj_2_weight = const()[name = tensor<string, []>("inner_encoder_q_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20132480)))];
+            tensor<fp32, [256]> inner_encoder_k_proj_2_bias = const()[name = tensor<string, []>("inner_encoder_k_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20394688)))];
+            tensor<fp32, [256, 256]> inner_encoder_k_proj_2_weight = const()[name = tensor<string, []>("inner_encoder_k_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20395776)))];
+            tensor<fp32, [256]> inner_encoder_v_proj_2_bias = const()[name = tensor<string, []>("inner_encoder_v_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20657984)))];
+            tensor<fp32, [256, 256]> inner_encoder_v_proj_2_weight = const()[name = tensor<string, []>("inner_encoder_v_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20659072)))];
+            tensor<fp32, [256]> inner_encoder_g_proj_2_bias = const()[name = tensor<string, []>("inner_encoder_g_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20921280)))];
+            tensor<fp32, [256, 256]> inner_encoder_g_proj_2_weight = const()[name = tensor<string, []>("inner_encoder_g_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20922368)))];
+            tensor<fp32, [256]> inner_encoder_out_proj_2_bias = const()[name = tensor<string, []>("inner_encoder_out_proj_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21184576)))];
+            tensor<fp32, [256, 256]> inner_encoder_out_proj_2_weight = const()[name = tensor<string, []>("inner_encoder_out_proj_2_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21185664)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21447872)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21448960)))];
+            tensor<fp32, [512]> inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21450048)))];
+            tensor<fp32, [512, 256, 1]> inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21452160)))];
+            tensor<fp32, [256, 1, 16]> inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21976512)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21992960)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21994048)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21995136)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21996224)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21997312)))];
+            tensor<fp32, [256, 256, 1]> inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21998400)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22260608)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22261696)))];
+            tensor<fp32, [1024]> inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22262784)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22266944)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23315584)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23316672)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_2_bias = const()[name = tensor<string, []>("inner_encoder_layer_norm_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24365312)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_2_weight = const()[name = tensor<string, []>("inner_encoder_layer_norm_2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24366400)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24367488)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24368576)))];
+            tensor<fp32, [1024]> inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24369664)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24373824)))];
+            tensor<fp32, [256]> inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25422464)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25423552)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_3_bias = const()[name = tensor<string, []>("inner_encoder_ret_lns_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26472192)))];
+            tensor<fp32, [256]> inner_encoder_ret_lns_3_weight = const()[name = tensor<string, []>("inner_encoder_ret_lns_3_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26473280)))];
+            tensor<fp32, [256]> inner_encoder_q_proj_3_bias = const()[name = tensor<string, []>("inner_encoder_q_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26474368)))];
+            tensor<fp32, [256, 256]> inner_encoder_q_proj_3_weight = const()[name = tensor<string, []>("inner_encoder_q_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26475456)))];
+            tensor<fp32, [256]> inner_encoder_k_proj_3_bias = const()[name = tensor<string, []>("inner_encoder_k_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26737664)))];
+            tensor<fp32, [256, 256]> inner_encoder_k_proj_3_weight = const()[name = tensor<string, []>("inner_encoder_k_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26738752)))];
+            tensor<fp32, [256]> inner_encoder_v_proj_3_bias = const()[name = tensor<string, []>("inner_encoder_v_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27000960)))];
+            tensor<fp32, [256, 256]> inner_encoder_v_proj_3_weight = const()[name = tensor<string, []>("inner_encoder_v_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27002048)))];
+            tensor<fp32, [256]> inner_encoder_g_proj_3_bias = const()[name = tensor<string, []>("inner_encoder_g_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27264256)))];
+            tensor<fp32, [256, 256]> inner_encoder_g_proj_3_weight = const()[name = tensor<string, []>("inner_encoder_g_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27265344)))];
+            tensor<fp32, [256]> inner_encoder_out_proj_3_bias = const()[name = tensor<string, []>("inner_encoder_out_proj_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27527552)))];
+            tensor<fp32, [256, 256]> inner_encoder_out_proj_3_weight = const()[name = tensor<string, []>("inner_encoder_out_proj_3_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27528640)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27790848)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27791936)))];
+            tensor<fp32, [512]> inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27793024)))];
+            tensor<fp32, [512, 256, 1]> inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27795136)))];
+            tensor<fp32, [256, 1, 16]> inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28319488)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28335936)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28337024)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28338112)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28339200)))];
+            tensor<fp32, [256]> inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28340288)))];
+            tensor<fp32, [256, 256, 1]> inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor<string, []>("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28341376)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28603584)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28604672)))];
+            tensor<fp32, [1024]> inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28605760)))];
+            tensor<fp32, [1024, 256]> inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28609920)))];
+            tensor<fp32, [256]> inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29658560)))];
+            tensor<fp32, [256, 1024]> inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor<string, []>("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29659648)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_3_bias = const()[name = tensor<string, []>("inner_encoder_layer_norm_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30708288)))];
+            tensor<fp32, [256]> inner_encoder_layer_norm_3_weight = const()[name = tensor<string, []>("inner_encoder_layer_norm_3_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30709376)))];
+            tensor<fp32, [2, 2]> inner_decoder__causal_mask = const()[name = tensor<string, []>("inner_decoder__causal_mask"), val = tensor<fp32, [2, 2]>([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])];
+            tensor<fp32, [256]> inner_decoder_convert_bias = const()[name = tensor<string, []>("inner_decoder_convert_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30710464)))];
+            tensor<fp32, [256, 512]> inner_decoder_convert_weight = const()[name = tensor<string, []>("inner_decoder_convert_weight"), val = tensor<fp32, [256, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30711552)))];
+            tensor<fp32, [256]> inner_decoder_q_proj_0_bias = const()[name = tensor<string, []>("inner_decoder_q_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31235904)))];
+            tensor<fp32, [256, 256]> inner_decoder_q_proj_0_weight = const()[name = tensor<string, []>("inner_decoder_q_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31236992)))];
+            tensor<fp32, [256]> inner_decoder_k_proj_0_bias = const()[name = tensor<string, []>("inner_decoder_k_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31499200)))];
+            tensor<fp32, [256, 256]> inner_decoder_k_proj_0_weight = const()[name = tensor<string, []>("inner_decoder_k_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31500288)))];
+            tensor<fp32, [256]> inner_decoder_v_proj_0_bias = const()[name = tensor<string, []>("inner_decoder_v_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31762496)))];
+            tensor<fp32, [256, 256]> inner_decoder_v_proj_0_weight = const()[name = tensor<string, []>("inner_decoder_v_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31763584)))];
+            tensor<fp32, [256]> inner_decoder_g_proj_0_bias = const()[name = tensor<string, []>("inner_decoder_g_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32025792)))];
+            tensor<fp32, [256, 256]> inner_decoder_g_proj_0_weight = const()[name = tensor<string, []>("inner_decoder_g_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32026880)))];
+            tensor<fp32, [256]> inner_decoder_out_proj_0_bias = const()[name = tensor<string, []>("inner_decoder_out_proj_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32289088)))];
+            tensor<fp32, [256, 256]> inner_decoder_out_proj_0_weight = const()[name = tensor<string, []>("inner_decoder_out_proj_0_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32290176)))];
+            tensor<fp32, [256]> inner_decoder_norm11_0_bias = const()[name = tensor<string, []>("inner_decoder_norm11_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32552384)))];
+            tensor<fp32, [256]> inner_decoder_norm11_0_weight = const()[name = tensor<string, []>("inner_decoder_norm11_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32553472)))];
+            tensor<fp32, [256]> inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor<string, []>("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32554560)))];
+            tensor<fp32, [256, 256]> inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor<string, []>("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32555648)))];
+            tensor<fp32, [768]> inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor<string, []>("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor<fp32, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32817856)))];
+            tensor<fp32, [768, 256]> inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor<string, []>("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32820992)))];
+            tensor<fp32, [256]> inner_decoder_norm21_0_bias = const()[name = tensor<string, []>("inner_decoder_norm21_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33607488)))];
+            tensor<fp32, [256]> inner_decoder_norm21_0_weight = const()[name = tensor<string, []>("inner_decoder_norm21_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33608576)))];
+            tensor<fp32, [2048]> inner_decoder_linear1_0_bias = const()[name = tensor<string, []>("inner_decoder_linear1_0_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33609664)))];
+            tensor<fp32, [2048, 256]> inner_decoder_linear1_0_weight = const()[name = tensor<string, []>("inner_decoder_linear1_0_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33617920)))];
+            tensor<fp32, [256]> inner_decoder_linear2_0_bias = const()[name = tensor<string, []>("inner_decoder_linear2_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35715136)))];
+            tensor<fp32, [256, 2048]> inner_decoder_linear2_0_weight = const()[name = tensor<string, []>("inner_decoder_linear2_0_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35716224)))];
+            tensor<fp32, [256]> inner_decoder_norm22_0_bias = const()[name = tensor<string, []>("inner_decoder_norm22_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37813440)))];
+            tensor<fp32, [256]> inner_decoder_norm22_0_weight = const()[name = tensor<string, []>("inner_decoder_norm22_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37814528)))];
+            tensor<fp32, [256]> inner_decoder_q_proj_1_bias = const()[name = tensor<string, []>("inner_decoder_q_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37815616)))];
+            tensor<fp32, [256, 256]> inner_decoder_q_proj_1_weight = const()[name = tensor<string, []>("inner_decoder_q_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37816704)))];
+            tensor<fp32, [256]> inner_decoder_k_proj_1_bias = const()[name = tensor<string, []>("inner_decoder_k_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38078912)))];
+            tensor<fp32, [256, 256]> inner_decoder_k_proj_1_weight = const()[name = tensor<string, []>("inner_decoder_k_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38080000)))];
+            tensor<fp32, [256]> inner_decoder_v_proj_1_bias = const()[name = tensor<string, []>("inner_decoder_v_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38342208)))];
+            tensor<fp32, [256, 256]> inner_decoder_v_proj_1_weight = const()[name = tensor<string, []>("inner_decoder_v_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38343296)))];
+            tensor<fp32, [256]> inner_decoder_g_proj_1_bias = const()[name = tensor<string, []>("inner_decoder_g_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38605504)))];
+            tensor<fp32, [256, 256]> inner_decoder_g_proj_1_weight = const()[name = tensor<string, []>("inner_decoder_g_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38606592)))];
+            tensor<fp32, [256]> inner_decoder_out_proj_1_bias = const()[name = tensor<string, []>("inner_decoder_out_proj_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38868800)))];
+            tensor<fp32, [256, 256]> inner_decoder_out_proj_1_weight = const()[name = tensor<string, []>("inner_decoder_out_proj_1_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38869888)))];
+            tensor<fp32, [256]> inner_decoder_norm11_1_bias = const()[name = tensor<string, []>("inner_decoder_norm11_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39132096)))];
+            tensor<fp32, [256]> inner_decoder_norm11_1_weight = const()[name = tensor<string, []>("inner_decoder_norm11_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39133184)))];
+            tensor<fp32, [256]> inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor<string, []>("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39134272)))];
+            tensor<fp32, [256, 256]> inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor<string, []>("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39135360)))];
+            tensor<fp32, [768]> inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor<string, []>("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor<fp32, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39397568)))];
+            tensor<fp32, [768, 256]> inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor<string, []>("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39400704)))];
+            tensor<fp32, [256]> inner_decoder_norm21_1_bias = const()[name = tensor<string, []>("inner_decoder_norm21_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40187200)))];
+            tensor<fp32, [256]> inner_decoder_norm21_1_weight = const()[name = tensor<string, []>("inner_decoder_norm21_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40188288)))];
+            tensor<fp32, [2048]> inner_decoder_linear1_1_bias = const()[name = tensor<string, []>("inner_decoder_linear1_1_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40189376)))];
+            tensor<fp32, [2048, 256]> inner_decoder_linear1_1_weight = const()[name = tensor<string, []>("inner_decoder_linear1_1_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40197632)))];
+            tensor<fp32, [256]> inner_decoder_linear2_1_bias = const()[name = tensor<string, []>("inner_decoder_linear2_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42294848)))];
+            tensor<fp32, [256, 2048]> inner_decoder_linear2_1_weight = const()[name = tensor<string, []>("inner_decoder_linear2_1_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42295936)))];
+            tensor<fp32, [256]> inner_decoder_norm22_1_bias = const()[name = tensor<string, []>("inner_decoder_norm22_1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44393152)))];
+            tensor<fp32, [256]> inner_decoder_norm22_1_weight = const()[name = tensor<string, []>("inner_decoder_norm22_1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44394240)))];
+            tensor<int32, [3]> var_19_begin_0 = const()[name = tensor<string, []>("op_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_19_end_0 = const()[name = tensor<string, []>("op_19_end_0"), val = tensor<int32, [3]>([1, 15, 23])];
+            tensor<bool, [3]> var_19_end_mask_0 = const()[name = tensor<string, []>("op_19_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 15, 23]> var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor<string, []>("op_19")];
+            tensor<int32, [3]> var_29_begin_0 = const()[name = tensor<string, []>("op_29_begin_0"), val = tensor<int32, [3]>([0, 10, 0])];
+            tensor<int32, [3]> var_29_end_0 = const()[name = tensor<string, []>("op_29_end_0"), val = tensor<int32, [3]>([1, 1, 23])];
+            tensor<bool, [3]> var_29_end_mask_0 = const()[name = tensor<string, []>("op_29_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 23]> var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor<string, []>("op_29")];
+            tensor<int32, []> stacked_axis_0 = const()[name = tensor<string, []>("stacked_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 2, 15, 23]> stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor<string, []>("stacked")];
+            tensor<int32, [3]> var_36 = const()[name = tensor<string, []>("op_36"), val = tensor<int32, [3]>([1, 2, 345])];
+            tensor<fp32, [1, 2, 345]> input_1 = reshape(shape = var_36, x = stacked)[name = tensor<string, []>("input_1")];
+            tensor<fp32, []> var_38 = const()[name = tensor<string, []>("op_38"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<bool, []> var_44 = const()[name = tensor<string, []>("op_44"), val = tensor<bool, []>(true)];
+            tensor<fp32, []> var_45 = const()[name = tensor<string, []>("op_45"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
+            tensor<int32, []> var_48 = const()[name = tensor<string, []>("op_48"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_50 = const()[name = tensor<string, []>("op_50"), val = tensor<int32, []>(2)];
+            tensor<int32, []> var_51 = const()[name = tensor<string, []>("op_51"), val = tensor<int32, []>(-1)];
+            tensor<fp32, []> var_53 = const()[name = tensor<string, []>("op_53"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, []> var_58 = const()[name = tensor<string, []>("op_58"), val = tensor<fp32, []>(0x1.5798eep-27)];
+            tensor<int32, []> var_62 = const()[name = tensor<string, []>("op_62"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 2, 256]> input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor<string, []>("linear_0")];
             tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor<string, []>("input_5")];
-            tensor<fp32, [1, 2, 1024]> inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor<string, []>("linear_1")];
-            tensor<fp32, [1, 2, 1024]> input_7 = silu(x = inputs_1)[name = tensor<string, []>("input_7")];
-            tensor<fp32, [1, 2, 256]> input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor<string, []>("linear_2")];
-            tensor<fp32, []> var_147 = const()[name = tensor<string, []>("op_147"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_148 = mul(x = input_11, y = var_147)[name = tensor<string, []>("op_148")];
-            tensor<fp32, [1, 2, 256]> input_13 = add(x = var_148, y = input_3)[name = tensor<string, []>("input_13")];
+            tensor<fp32, [1, 2, 256]> input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_45, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor<string, []>("input_5")];
+            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 2, 1024]> inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor<string, []>("linear_1")];
+            tensor<fp32, [1, 2, 1024]> input_9 = silu(x = inputs_1)[name = tensor<string, []>("input_9")];
+            tensor<fp32, [1, 2, 256]> input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor<string, []>("linear_2")];
+            tensor<fp32, []> var_183 = const()[name = tensor<string, []>("op_183"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_184 = mul(x = input_13, y = var_183)[name = tensor<string, []>("op_184")];
+            tensor<fp32, [1, 2, 256]> input_15 = add(x = var_184, y = input_5)[name = tensor<string, []>("input_15")];
             tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor<string, []>("x_1")];
+            tensor<fp32, [1, 2, 256]> x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor<string, []>("x_1")];
             tensor<int32, [5]> prev_kv_1_begin_0 = const()[name = tensor<string, []>("prev_kv_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_1_end_0 = const()[name = tensor<string, []>("prev_kv_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_1_end_mask_0 = const()[name = tensor<string, []>("prev_kv_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -239,153 +253,153 @@ program(1.0)
             tensor<bool, [2]> prev_scale_1_end_mask_0 = const()[name = tensor<string, []>("prev_scale_1_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_1_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_1_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor<string, []>("prev_scale_1")];
-            tensor<fp32, [1, 2, 256]> var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_3")];
-            tensor<int32, [4]> var_163 = const()[name = tensor<string, []>("op_163"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_164 = reshape(shape = var_163, x = var_162)[name = tensor<string, []>("op_164")];
+            tensor<fp32, [1, 2, 256]> var_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_3")];
+            tensor<int32, [4]> var_199 = const()[name = tensor<string, []>("op_199"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_200 = reshape(shape = var_199, x = var_198)[name = tensor<string, []>("op_200")];
             tensor<int32, [4]> q_1_perm_0 = const()[name = tensor<string, []>("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_4")];
-            tensor<fp32, []> var_169 = const()[name = tensor<string, []>("op_169"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 2, 256]> var_170 = mul(x = var_168, y = var_169)[name = tensor<string, []>("op_170")];
-            tensor<int32, [4]> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_172 = reshape(shape = var_171, x = var_170)[name = tensor<string, []>("op_172")];
+            tensor<fp32, [1, 2, 256]> var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_4")];
+            tensor<fp32, []> var_205 = const()[name = tensor<string, []>("op_205"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 2, 256]> var_206 = mul(x = var_204, y = var_205)[name = tensor<string, []>("op_206")];
+            tensor<int32, [4]> var_207 = const()[name = tensor<string, []>("op_207"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_208 = reshape(shape = var_207, x = var_206)[name = tensor<string, []>("op_208")];
             tensor<int32, [4]> k_1_perm_0 = const()[name = tensor<string, []>("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_5")];
-            tensor<int32, [4]> var_177 = const()[name = tensor<string, []>("op_177"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_178 = reshape(shape = var_177, x = var_176)[name = tensor<string, []>("op_178")];
+            tensor<fp32, [1, 2, 256]> var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_5")];
+            tensor<int32, [4]> var_213 = const()[name = tensor<string, []>("op_213"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_214 = reshape(shape = var_213, x = var_212)[name = tensor<string, []>("op_214")];
             tensor<int32, [4]> v_1_perm_0 = const()[name = tensor<string, []>("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_6")];
+            tensor<fp32, [1, 2, 256]> input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor<string, []>("linear_6")];
             tensor<fp32, [1]> sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor<string, []>("sqrt_s0_1")];
-            tensor<fp32, [2]> s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor<string, []>("s_t_1")];
+            tensor<fp32, [2]> s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_1")];
             tensor<fp32, [2]> sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor<string, []>("sqrt_s_t_1")];
             tensor<bool, []> qk_1_transpose_x_1 = const()[name = tensor<string, []>("qk_1_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_1_transpose_y_1 = const()[name = tensor<string, []>("qk_1_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [1, 4, 2, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor<string, []>("transpose_57")];
-            tensor<fp32, [1, 4, 2, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor<string, []>("transpose_58")];
+            tensor<fp32, [1, 4, 2, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_208)[name = tensor<string, []>("transpose_57")];
+            tensor<fp32, [1, 4, 2, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_200)[name = tensor<string, []>("transpose_58")];
             tensor<fp32, [1, 4, 2, 2]> qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor<string, []>("qk_1")];
-            tensor<int32, [2]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor<string, []>("op_189")];
-            tensor<fp32, [2, 2]> M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor<string, []>("M_1")];
-            tensor<fp32, [1, 4, 2, 2]> var_191 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_191")];
+            tensor<int32, [2]> var_224 = const()[name = tensor<string, []>("op_224"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor<string, []>("op_225")];
+            tensor<fp32, [2, 2]> M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor<string, []>("M_1")];
+            tensor<fp32, [1, 4, 2, 2]> var_227 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_227")];
             tensor<bool, []> inner_1_transpose_x_0 = const()[name = tensor<string, []>("inner_1_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_1_transpose_y_0 = const()[name = tensor<string, []>("inner_1_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor<string, []>("transpose_56")];
-            tensor<fp32, [1, 4, 2, 64]> inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor<string, []>("inner_1")];
-            tensor<bool, []> var_193_transpose_x_0 = const()[name = tensor<string, []>("op_193_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_193_transpose_y_0 = const()[name = tensor<string, []>("op_193_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor<string, []>("op_193")];
-            tensor<fp32, [2]> var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor<string, []>("op_194")];
-            tensor<int32, [4]> var_195 = const()[name = tensor<string, []>("op_195"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_196 = reshape(shape = var_195, x = var_194)[name = tensor<string, []>("op_196")];
-            tensor<fp32, [1, 4, 2, 64]> cross_1 = mul(x = var_193, y = var_196)[name = tensor<string, []>("cross_1")];
+            tensor<fp32, [1, 4, 2, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_214)[name = tensor<string, []>("transpose_56")];
+            tensor<fp32, [1, 4, 2, 64]> inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor<string, []>("inner_1")];
+            tensor<bool, []> var_229_transpose_x_0 = const()[name = tensor<string, []>("op_229_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_229_transpose_y_0 = const()[name = tensor<string, []>("op_229_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 2, 64]> var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor<string, []>("op_229")];
+            tensor<fp32, [2]> var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor<string, []>("op_230")];
+            tensor<int32, [4]> var_231 = const()[name = tensor<string, []>("op_231"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_232 = reshape(shape = var_231, x = var_230)[name = tensor<string, []>("op_232")];
+            tensor<fp32, [1, 4, 2, 64]> cross_1 = mul(x = var_229, y = var_232)[name = tensor<string, []>("cross_1")];
             tensor<fp32, [1, 4, 2, 64]> out_1 = add(x = inner_1, y = cross_1)[name = tensor<string, []>("out_1")];
-            tensor<fp32, [1, 4, 64, 64]> var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor<string, []>("op_199")];
-            tensor<bool, []> var_201_transpose_x_1 = const()[name = tensor<string, []>("op_201_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_201_transpose_y_1 = const()[name = tensor<string, []>("op_201_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor<string, []>("op_201")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor<string, []>("new_kv_unnorm_1")];
-            tensor<fp32, []> var_203 = const()[name = tensor<string, []>("op_203"), val = tensor<fp32, []>(0x1p+1)];
-            tensor<fp32, [1]> new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor<string, []>("new_scale_1")];
-            tensor<fp32, [1]> var_205 = sqrt(x = new_scale_1)[name = tensor<string, []>("op_205")];
-            tensor<fp32, [1, 4, 64, 64]> var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor<string, []>("op_206")];
-            tensor<int32, [4]> var_207_perm_0 = const()[name = tensor<string, []>("op_207_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor<string, []>("op_235")];
+            tensor<bool, []> var_237_transpose_x_1 = const()[name = tensor<string, []>("op_237_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_237_transpose_y_1 = const()[name = tensor<string, []>("op_237_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor<string, []>("op_237")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor<string, []>("new_kv_unnorm_1")];
+            tensor<fp32, []> var_239 = const()[name = tensor<string, []>("op_239"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor<string, []>("new_scale_1")];
+            tensor<fp32, [1]> var_241 = sqrt(x = new_scale_1)[name = tensor<string, []>("op_241")];
+            tensor<fp32, [1, 4, 64, 64]> var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor<string, []>("op_242")];
+            tensor<int32, [4]> var_243_perm_0 = const()[name = tensor<string, []>("op_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_3_axes_0 = const()[name = tensor<string, []>("out_3_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 4, 64]> var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
-            tensor<fp32, [1, 2, 4, 64]> out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor<string, []>("out_3")];
-            tensor<int32, [3]> var_211 = const()[name = tensor<string, []>("op_211"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<fp32, [1, 2, 256]> out_5 = reshape(shape = var_211, x = out_3)[name = tensor<string, []>("out_5")];
-            tensor<fp32, [1, 2, 256]> var_213 = silu(x = input_17)[name = tensor<string, []>("op_213")];
-            tensor<fp32, [1, 2, 256]> input_19 = mul(x = var_213, y = out_5)[name = tensor<string, []>("input_19")];
-            tensor<fp32, [1, 2, 256]> ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor<string, []>("linear_7")];
-            tensor<fp32, [1, 2, 256]> x_3 = add(x = input_13, y = ret_out_1)[name = tensor<string, []>("x_3")];
+            tensor<fp32, [1, 2, 4, 64]> var_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
+            tensor<fp32, [1, 2, 4, 64]> out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_53, x = var_243)[name = tensor<string, []>("out_3")];
+            tensor<int32, [3]> var_247 = const()[name = tensor<string, []>("op_247"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<fp32, [1, 2, 256]> out_5 = reshape(shape = var_247, x = out_3)[name = tensor<string, []>("out_5")];
+            tensor<fp32, [1, 2, 256]> var_249 = silu(x = input_19)[name = tensor<string, []>("op_249")];
+            tensor<fp32, [1, 2, 256]> input_21 = mul(x = var_249, y = out_5)[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 2, 256]> ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor<string, []>("linear_7")];
+            tensor<fp32, [1, 2, 256]> x_3 = add(x = input_15, y = ret_out_1)[name = tensor<string, []>("x_3")];
             tensor<int32, [4]> window_1_begin_0 = const()[name = tensor<string, []>("window_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [4]> window_1_end_0 = const()[name = tensor<string, []>("window_1_end_0"), val = tensor<int32, [4]>([1, 1, 16, 256])];
             tensor<bool, [4]> window_1_end_mask_0 = const()[name = tensor<string, []>("window_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_1_squeeze_mask_0 = const()[name = tensor<string, []>("window_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_1")];
-            tensor<int32, [3]> var_221_begin_0 = const()[name = tensor<string, []>("op_221_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_221_end_0 = const()[name = tensor<string, []>("op_221_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_221_end_mask_0 = const()[name = tensor<string, []>("op_221_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor<string, []>("op_221")];
-            tensor<int32, [3]> var_224_begin_0 = const()[name = tensor<string, []>("op_224_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_224_end_0 = const()[name = tensor<string, []>("op_224_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_224_end_mask_0 = const()[name = tensor<string, []>("op_224_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor<string, []>("op_224")];
+            tensor<int32, [3]> var_257_begin_0 = const()[name = tensor<string, []>("op_257_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_257_end_0 = const()[name = tensor<string, []>("op_257_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_257_end_mask_0 = const()[name = tensor<string, []>("op_257_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor<string, []>("op_257")];
+            tensor<int32, [3]> var_260_begin_0 = const()[name = tensor<string, []>("op_260_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_260_end_0 = const()[name = tensor<string, []>("op_260_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_260_end_mask_0 = const()[name = tensor<string, []>("op_260_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor<string, []>("op_260")];
             tensor<bool, []> window_3_interleave_0 = const()[name = tensor<string, []>("window_3_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor<string, []>("window_3")];
-            tensor<int32, [3]> var_229_begin_0 = const()[name = tensor<string, []>("op_229_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_229_end_0 = const()[name = tensor<string, []>("op_229_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_229_end_mask_0 = const()[name = tensor<string, []>("op_229_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor<string, []>("op_229")];
-            tensor<int32, [3]> var_232_begin_0 = const()[name = tensor<string, []>("op_232_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_232_end_0 = const()[name = tensor<string, []>("op_232_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_232_end_mask_0 = const()[name = tensor<string, []>("op_232_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor<string, []>("op_232")];
+            tensor<fp32, [1, 16, 256]> window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor<string, []>("window_3")];
+            tensor<int32, [3]> var_265_begin_0 = const()[name = tensor<string, []>("op_265_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_265_end_0 = const()[name = tensor<string, []>("op_265_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_265_end_mask_0 = const()[name = tensor<string, []>("op_265_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor<string, []>("op_265")];
+            tensor<int32, [3]> var_268_begin_0 = const()[name = tensor<string, []>("op_268_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_268_end_0 = const()[name = tensor<string, []>("op_268_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_268_end_mask_0 = const()[name = tensor<string, []>("op_268_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor<string, []>("op_268")];
             tensor<bool, []> window_5_interleave_0 = const()[name = tensor<string, []>("window_5_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor<string, []>("window_5")];
-            tensor<bool, []> input_21_interleave_0 = const()[name = tensor<string, []>("input_21_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [2, 16, 256]> input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 16, 256]> window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor<string, []>("window_5")];
+            tensor<bool, []> input_23_interleave_0 = const()[name = tensor<string, []>("input_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [2, 16, 256]> input_23 = concat(axis = var_48, interleave = input_23_interleave_0, values = (window_3, window_5))[name = tensor<string, []>("input_23")];
             tensor<int32, [1]> x_5_axes_0 = const()[name = tensor<string, []>("x_5_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 16, 256]> x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor<string, []>("x_5")];
-            tensor<int32, [3]> input_23_perm_0 = const()[name = tensor<string, []>("input_23_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [2, 16, 256]> x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor<string, []>("x_5")];
+            tensor<int32, [3]> input_25_perm_0 = const()[name = tensor<string, []>("input_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> inputs_3_pad_type_0 = const()[name = tensor<string, []>("inputs_3_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> inputs_3_strides_0 = const()[name = tensor<string, []>("inputs_3_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> inputs_3_pad_0 = const()[name = tensor<string, []>("inputs_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> inputs_3_dilations_0 = const()[name = tensor<string, []>("inputs_3_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> inputs_3_groups_0 = const()[name = tensor<string, []>("inputs_3_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
-            tensor<fp32, [2, 512, 16]> inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor<string, []>("inputs_3")];
-            tensor<int32, [2]> var_257_split_sizes_0 = const()[name = tensor<string, []>("op_257_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_257_axis_0 = const()[name = tensor<string, []>("op_257_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> var_257_0, tensor<fp32, [2, 256, 16]> var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_257")];
-            tensor<fp32, [2, 256, 16]> var_259 = sigmoid(x = var_257_1)[name = tensor<string, []>("op_259")];
-            tensor<fp32, [2, 256, 16]> inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor<string, []>("inputs_5")];
+            tensor<fp32, [2, 256, 16]> input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
+            tensor<fp32, [2, 512, 16]> inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor<string, []>("inputs_3")];
+            tensor<int32, [2]> var_293_split_sizes_0 = const()[name = tensor<string, []>("op_293_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_293_axis_0 = const()[name = tensor<string, []>("op_293_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> var_293_0, tensor<fp32, [2, 256, 16]> var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_293")];
+            tensor<fp32, [2, 256, 16]> var_295 = sigmoid(x = var_293_1)[name = tensor<string, []>("op_295")];
+            tensor<fp32, [2, 256, 16]> inputs_5 = mul(x = var_293_0, y = var_295)[name = tensor<string, []>("inputs_5")];
             tensor<string, []> outputs_aug_1_pad_type_0 = const()[name = tensor<string, []>("outputs_aug_1_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> outputs_aug_1_pad_0 = const()[name = tensor<string, []>("outputs_aug_1_pad_0"), val = tensor<int32, [2]>([15, 15])];
             tensor<int32, []> outputs_aug_1_groups_0 = const()[name = tensor<string, []>("outputs_aug_1_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [1]> outputs_aug_1_strides_0 = const()[name = tensor<string, []>("outputs_aug_1_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [1]> outputs_aug_1_dilations_0 = const()[name = tensor<string, []>("outputs_aug_1_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [2, 256, 31]> outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor<string, []>("outputs_aug_1")];
-            tensor<int32, [3]> input_25_begin_0 = const()[name = tensor<string, []>("input_25_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> input_25_end_0 = const()[name = tensor<string, []>("input_25_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
-            tensor<bool, [3]> input_25_end_mask_0 = const()[name = tensor<string, []>("input_25_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
-            tensor<fp32, [2, 256, 16]> input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor<string, []>("input_25")];
-            tensor<fp32, [2, 256, 16]> inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor<string, []>("inputs_7")];
-            tensor<fp32, [2, 256, 16]> input_27 = silu(x = inputs_7)[name = tensor<string, []>("input_27")];
-            tensor<string, []> input_29_pad_type_0 = const()[name = tensor<string, []>("input_29_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [1]> input_29_strides_0 = const()[name = tensor<string, []>("input_29_strides_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [2]> input_29_pad_0 = const()[name = tensor<string, []>("input_29_pad_0"), val = tensor<int32, [2]>([0, 0])];
-            tensor<int32, [1]> input_29_dilations_0 = const()[name = tensor<string, []>("input_29_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, []> input_29_groups_0 = const()[name = tensor<string, []>("input_29_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor<string, []>("input_29")];
+            tensor<fp32, [2, 256, 31]> outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor<string, []>("outputs_aug_1")];
+            tensor<int32, [3]> input_27_begin_0 = const()[name = tensor<string, []>("input_27_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_27_end_0 = const()[name = tensor<string, []>("input_27_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
+            tensor<bool, [3]> input_27_end_mask_0 = const()[name = tensor<string, []>("input_27_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp32, [2, 256, 16]> input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor<string, []>("input_27")];
+            tensor<fp32, [2, 256, 16]> inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor<string, []>("inputs_7")];
+            tensor<fp32, [2, 256, 16]> input_29 = silu(x = inputs_7)[name = tensor<string, []>("input_29")];
+            tensor<string, []> input_31_pad_type_0 = const()[name = tensor<string, []>("input_31_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_31_strides_0 = const()[name = tensor<string, []>("input_31_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_31_pad_0 = const()[name = tensor<string, []>("input_31_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_31_dilations_0 = const()[name = tensor<string, []>("input_31_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_31_groups_0 = const()[name = tensor<string, []>("input_31_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor<string, []>("input_31")];
             tensor<int32, [3]> conv_out_1_perm_0 = const()[name = tensor<string, []>("conv_out_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<int32, [3]> var_290_begin_0 = const()[name = tensor<string, []>("op_290_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_290_end_0 = const()[name = tensor<string, []>("op_290_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
-            tensor<bool, [3]> var_290_end_mask_0 = const()[name = tensor<string, []>("op_290_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [2, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor<string, []>("transpose_53")];
-            tensor<fp32, [2, 1, 256]> var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_290")];
-            tensor<int32, [3]> var_292_perm_0 = const()[name = tensor<string, []>("op_292_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 2, 256]> var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor<string, []>("transpose_52")];
-            tensor<fp32, [1, 2, 256]> input_31 = add(x = x_3, y = var_292)[name = tensor<string, []>("input_31")];
-            tensor<int32, [1]> input_33_axes_0 = const()[name = tensor<string, []>("input_33_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor<string, []>("input_33")];
-            tensor<fp32, [1, 2, 1024]> inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor<string, []>("linear_8")];
-            tensor<fp32, [1, 2, 1024]> input_35 = silu(x = inputs_9)[name = tensor<string, []>("input_35")];
-            tensor<fp32, [1, 2, 256]> input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor<string, []>("linear_9")];
-            tensor<fp32, []> var_315 = const()[name = tensor<string, []>("op_315"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_316 = mul(x = input_39, y = var_315)[name = tensor<string, []>("op_316")];
-            tensor<fp32, [1, 2, 256]> input_41 = add(x = var_316, y = input_31)[name = tensor<string, []>("input_41")];
-            tensor<int32, [1]> input_43_axes_0 = const()[name = tensor<string, []>("input_43_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor<string, []>("input_43")];
+            tensor<int32, [3]> var_326_begin_0 = const()[name = tensor<string, []>("op_326_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_326_end_0 = const()[name = tensor<string, []>("op_326_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
+            tensor<bool, [3]> var_326_end_mask_0 = const()[name = tensor<string, []>("op_326_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [2, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor<string, []>("transpose_53")];
+            tensor<fp32, [2, 1, 256]> var_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_326")];
+            tensor<int32, [3]> var_328_perm_0 = const()[name = tensor<string, []>("op_328_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 2, 256]> var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor<string, []>("transpose_52")];
+            tensor<fp32, [1, 2, 256]> input_33 = add(x = x_3, y = var_328)[name = tensor<string, []>("input_33")];
+            tensor<int32, [1]> input_35_axes_0 = const()[name = tensor<string, []>("input_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor<string, []>("input_35")];
+            tensor<fp32, [1, 2, 1024]> inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor<string, []>("linear_8")];
+            tensor<fp32, [1, 2, 1024]> input_37 = silu(x = inputs_9)[name = tensor<string, []>("input_37")];
+            tensor<fp32, [1, 2, 256]> input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor<string, []>("linear_9")];
+            tensor<fp32, []> var_351 = const()[name = tensor<string, []>("op_351"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_352 = mul(x = input_41, y = var_351)[name = tensor<string, []>("op_352")];
+            tensor<fp32, [1, 2, 256]> input_43 = add(x = var_352, y = input_33)[name = tensor<string, []>("input_43")];
             tensor<int32, [1]> input_45_axes_0 = const()[name = tensor<string, []>("input_45_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor<string, []>("input_45")];
-            tensor<fp32, [1, 2, 1024]> inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor<string, []>("linear_10")];
-            tensor<fp32, [1, 2, 1024]> input_47 = silu(x = inputs_11)[name = tensor<string, []>("input_47")];
-            tensor<fp32, [1, 2, 256]> input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor<string, []>("linear_11")];
-            tensor<fp32, []> var_345 = const()[name = tensor<string, []>("op_345"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_346 = mul(x = input_51, y = var_345)[name = tensor<string, []>("op_346")];
-            tensor<fp32, [1, 2, 256]> input_53 = add(x = var_346, y = input_43)[name = tensor<string, []>("input_53")];
+            tensor<fp32, [1, 2, 256]> input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor<string, []>("input_45")];
+            tensor<int32, [1]> input_47_axes_0 = const()[name = tensor<string, []>("input_47_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor<string, []>("input_47")];
+            tensor<fp32, [1, 2, 1024]> inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor<string, []>("linear_10")];
+            tensor<fp32, [1, 2, 1024]> input_49 = silu(x = inputs_11)[name = tensor<string, []>("input_49")];
+            tensor<fp32, [1, 2, 256]> input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor<string, []>("linear_11")];
+            tensor<fp32, []> var_381 = const()[name = tensor<string, []>("op_381"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_382 = mul(x = input_53, y = var_381)[name = tensor<string, []>("op_382")];
+            tensor<fp32, [1, 2, 256]> input_55 = add(x = var_382, y = input_45)[name = tensor<string, []>("input_55")];
             tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor<string, []>("x_7")];
+            tensor<fp32, [1, 2, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor<string, []>("x_7")];
             tensor<int32, [5]> prev_kv_3_begin_0 = const()[name = tensor<string, []>("prev_kv_3_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_3_end_0 = const()[name = tensor<string, []>("prev_kv_3_end_0"), val = tensor<int32, [5]>([2, 1, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_3_end_mask_0 = const()[name = tensor<string, []>("prev_kv_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -396,153 +410,153 @@ program(1.0)
             tensor<bool, [2]> prev_scale_3_end_mask_0 = const()[name = tensor<string, []>("prev_scale_3_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_3_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_3_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor<string, []>("prev_scale_3")];
-            tensor<fp32, [1, 2, 256]> var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_12")];
-            tensor<int32, [4]> var_361 = const()[name = tensor<string, []>("op_361"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_362 = reshape(shape = var_361, x = var_360)[name = tensor<string, []>("op_362")];
+            tensor<fp32, [1, 2, 256]> var_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_12")];
+            tensor<int32, [4]> var_397 = const()[name = tensor<string, []>("op_397"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_398 = reshape(shape = var_397, x = var_396)[name = tensor<string, []>("op_398")];
             tensor<int32, [4]> q_3_perm_0 = const()[name = tensor<string, []>("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_13")];
-            tensor<fp32, []> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 2, 256]> var_368 = mul(x = var_366, y = var_367)[name = tensor<string, []>("op_368")];
-            tensor<int32, [4]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_370 = reshape(shape = var_369, x = var_368)[name = tensor<string, []>("op_370")];
+            tensor<fp32, [1, 2, 256]> var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_13")];
+            tensor<fp32, []> var_403 = const()[name = tensor<string, []>("op_403"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 2, 256]> var_404 = mul(x = var_402, y = var_403)[name = tensor<string, []>("op_404")];
+            tensor<int32, [4]> var_405 = const()[name = tensor<string, []>("op_405"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_406 = reshape(shape = var_405, x = var_404)[name = tensor<string, []>("op_406")];
             tensor<int32, [4]> k_3_perm_0 = const()[name = tensor<string, []>("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_14")];
-            tensor<int32, [4]> var_375 = const()[name = tensor<string, []>("op_375"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_376 = reshape(shape = var_375, x = var_374)[name = tensor<string, []>("op_376")];
+            tensor<fp32, [1, 2, 256]> var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_14")];
+            tensor<int32, [4]> var_411 = const()[name = tensor<string, []>("op_411"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_412 = reshape(shape = var_411, x = var_410)[name = tensor<string, []>("op_412")];
             tensor<int32, [4]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_15")];
+            tensor<fp32, [1, 2, 256]> input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_15")];
             tensor<fp32, [1]> sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor<string, []>("sqrt_s0_3")];
-            tensor<fp32, [2]> s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor<string, []>("s_t_3")];
+            tensor<fp32, [2]> s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_3")];
             tensor<fp32, [2]> sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor<string, []>("sqrt_s_t_3")];
             tensor<bool, []> qk_3_transpose_x_1 = const()[name = tensor<string, []>("qk_3_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_3_transpose_y_1 = const()[name = tensor<string, []>("qk_3_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [1, 4, 2, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor<string, []>("transpose_50")];
-            tensor<fp32, [1, 4, 2, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor<string, []>("transpose_51")];
+            tensor<fp32, [1, 4, 2, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_406)[name = tensor<string, []>("transpose_50")];
+            tensor<fp32, [1, 4, 2, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_398)[name = tensor<string, []>("transpose_51")];
             tensor<fp32, [1, 4, 2, 2]> qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor<string, []>("qk_3")];
-            tensor<int32, [2]> var_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor<string, []>("op_387")];
-            tensor<fp32, [2, 2]> M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor<string, []>("M_3")];
-            tensor<fp32, [1, 4, 2, 2]> var_389 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_389")];
+            tensor<int32, [2]> var_422 = const()[name = tensor<string, []>("op_422"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor<string, []>("op_423")];
+            tensor<fp32, [2, 2]> M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor<string, []>("M_3")];
+            tensor<fp32, [1, 4, 2, 2]> var_425 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_425")];
             tensor<bool, []> inner_3_transpose_x_0 = const()[name = tensor<string, []>("inner_3_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_3_transpose_y_0 = const()[name = tensor<string, []>("inner_3_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor<string, []>("transpose_49")];
-            tensor<fp32, [1, 4, 2, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor<string, []>("inner_3")];
-            tensor<bool, []> var_391_transpose_x_0 = const()[name = tensor<string, []>("op_391_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_391_transpose_y_0 = const()[name = tensor<string, []>("op_391_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_391")];
-            tensor<fp32, [2]> var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_392")];
-            tensor<int32, [4]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_394 = reshape(shape = var_393, x = var_392)[name = tensor<string, []>("op_394")];
-            tensor<fp32, [1, 4, 2, 64]> cross_3 = mul(x = var_391, y = var_394)[name = tensor<string, []>("cross_3")];
+            tensor<fp32, [1, 4, 2, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_412)[name = tensor<string, []>("transpose_49")];
+            tensor<fp32, [1, 4, 2, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor<string, []>("inner_3")];
+            tensor<bool, []> var_427_transpose_x_0 = const()[name = tensor<string, []>("op_427_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_427_transpose_y_0 = const()[name = tensor<string, []>("op_427_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 2, 64]> var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_427")];
+            tensor<fp32, [2]> var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_428")];
+            tensor<int32, [4]> var_429 = const()[name = tensor<string, []>("op_429"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_430 = reshape(shape = var_429, x = var_428)[name = tensor<string, []>("op_430")];
+            tensor<fp32, [1, 4, 2, 64]> cross_3 = mul(x = var_427, y = var_430)[name = tensor<string, []>("cross_3")];
             tensor<fp32, [1, 4, 2, 64]> out_7 = add(x = inner_3, y = cross_3)[name = tensor<string, []>("out_7")];
-            tensor<fp32, [1, 4, 64, 64]> var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_397")];
-            tensor<bool, []> var_399_transpose_x_1 = const()[name = tensor<string, []>("op_399_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_399_transpose_y_1 = const()[name = tensor<string, []>("op_399_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_399")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor<string, []>("new_kv_unnorm_3")];
-            tensor<fp32, []> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<fp32, []>(0x1p+1)];
-            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor<string, []>("new_scale_3")];
-            tensor<fp32, [1]> var_403 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_403")];
-            tensor<fp32, [1, 4, 64, 64]> var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor<string, []>("op_404")];
-            tensor<int32, [4]> var_405_perm_0 = const()[name = tensor<string, []>("op_405_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_433")];
+            tensor<bool, []> var_435_transpose_x_1 = const()[name = tensor<string, []>("op_435_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_435_transpose_y_1 = const()[name = tensor<string, []>("op_435_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_435")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor<string, []>("new_kv_unnorm_3")];
+            tensor<fp32, []> var_437 = const()[name = tensor<string, []>("op_437"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor<string, []>("new_scale_3")];
+            tensor<fp32, [1]> var_439 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_439")];
+            tensor<fp32, [1, 4, 64, 64]> var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor<string, []>("op_440")];
+            tensor<int32, [4]> var_441_perm_0 = const()[name = tensor<string, []>("op_441_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_9_axes_0 = const()[name = tensor<string, []>("out_9_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 4, 64]> var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
-            tensor<fp32, [1, 2, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor<string, []>("out_9")];
-            tensor<int32, [3]> var_409 = const()[name = tensor<string, []>("op_409"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<fp32, [1, 2, 256]> out_11 = reshape(shape = var_409, x = out_9)[name = tensor<string, []>("out_11")];
-            tensor<fp32, [1, 2, 256]> var_411 = silu(x = input_57)[name = tensor<string, []>("op_411")];
-            tensor<fp32, [1, 2, 256]> input_59 = mul(x = var_411, y = out_11)[name = tensor<string, []>("input_59")];
-            tensor<fp32, [1, 2, 256]> ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor<string, []>("linear_16")];
-            tensor<fp32, [1, 2, 256]> x_9 = add(x = input_53, y = ret_out_3)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 2, 4, 64]> var_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
+            tensor<fp32, [1, 2, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_53, x = var_441)[name = tensor<string, []>("out_9")];
+            tensor<int32, [3]> var_445 = const()[name = tensor<string, []>("op_445"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<fp32, [1, 2, 256]> out_11 = reshape(shape = var_445, x = out_9)[name = tensor<string, []>("out_11")];
+            tensor<fp32, [1, 2, 256]> var_447 = silu(x = input_59)[name = tensor<string, []>("op_447")];
+            tensor<fp32, [1, 2, 256]> input_61 = mul(x = var_447, y = out_11)[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 2, 256]> ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor<string, []>("linear_16")];
+            tensor<fp32, [1, 2, 256]> x_9 = add(x = input_55, y = ret_out_3)[name = tensor<string, []>("x_9")];
             tensor<int32, [4]> window_7_begin_0 = const()[name = tensor<string, []>("window_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> window_7_end_0 = const()[name = tensor<string, []>("window_7_end_0"), val = tensor<int32, [4]>([2, 1, 16, 256])];
             tensor<bool, [4]> window_7_end_mask_0 = const()[name = tensor<string, []>("window_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_7_squeeze_mask_0 = const()[name = tensor<string, []>("window_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_7")];
-            tensor<int32, [3]> var_419_begin_0 = const()[name = tensor<string, []>("op_419_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_419_end_0 = const()[name = tensor<string, []>("op_419_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_419_end_mask_0 = const()[name = tensor<string, []>("op_419_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor<string, []>("op_419")];
-            tensor<int32, [3]> var_422_begin_0 = const()[name = tensor<string, []>("op_422_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_422_end_0 = const()[name = tensor<string, []>("op_422_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_422_end_mask_0 = const()[name = tensor<string, []>("op_422_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor<string, []>("op_422")];
+            tensor<int32, [3]> var_455_begin_0 = const()[name = tensor<string, []>("op_455_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_455_end_0 = const()[name = tensor<string, []>("op_455_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_455_end_mask_0 = const()[name = tensor<string, []>("op_455_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor<string, []>("op_455")];
+            tensor<int32, [3]> var_458_begin_0 = const()[name = tensor<string, []>("op_458_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_458_end_0 = const()[name = tensor<string, []>("op_458_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_458_end_mask_0 = const()[name = tensor<string, []>("op_458_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor<string, []>("op_458")];
             tensor<bool, []> window_9_interleave_0 = const()[name = tensor<string, []>("window_9_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor<string, []>("window_9")];
-            tensor<int32, [3]> var_427_begin_0 = const()[name = tensor<string, []>("op_427_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_427_end_0 = const()[name = tensor<string, []>("op_427_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_427_end_mask_0 = const()[name = tensor<string, []>("op_427_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor<string, []>("op_427")];
-            tensor<int32, [3]> var_430_begin_0 = const()[name = tensor<string, []>("op_430_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_430_end_0 = const()[name = tensor<string, []>("op_430_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_430_end_mask_0 = const()[name = tensor<string, []>("op_430_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor<string, []>("op_430")];
+            tensor<fp32, [1, 16, 256]> window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor<string, []>("window_9")];
+            tensor<int32, [3]> var_463_begin_0 = const()[name = tensor<string, []>("op_463_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_463_end_0 = const()[name = tensor<string, []>("op_463_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_463_end_mask_0 = const()[name = tensor<string, []>("op_463_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor<string, []>("op_463")];
+            tensor<int32, [3]> var_466_begin_0 = const()[name = tensor<string, []>("op_466_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_466_end_0 = const()[name = tensor<string, []>("op_466_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_466_end_mask_0 = const()[name = tensor<string, []>("op_466_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor<string, []>("op_466")];
             tensor<bool, []> window_11_interleave_0 = const()[name = tensor<string, []>("window_11_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor<string, []>("window_11")];
-            tensor<bool, []> input_61_interleave_0 = const()[name = tensor<string, []>("input_61_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [2, 16, 256]> input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 16, 256]> window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor<string, []>("window_11")];
+            tensor<bool, []> input_63_interleave_0 = const()[name = tensor<string, []>("input_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [2, 16, 256]> input_63 = concat(axis = var_48, interleave = input_63_interleave_0, values = (window_9, window_11))[name = tensor<string, []>("input_63")];
             tensor<int32, [1]> x_11_axes_0 = const()[name = tensor<string, []>("x_11_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 16, 256]> x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor<string, []>("x_11")];
-            tensor<int32, [3]> input_63_perm_0 = const()[name = tensor<string, []>("input_63_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [2, 16, 256]> x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor<string, []>("x_11")];
+            tensor<int32, [3]> input_65_perm_0 = const()[name = tensor<string, []>("input_65_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> inputs_13_pad_type_0 = const()[name = tensor<string, []>("inputs_13_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> inputs_13_strides_0 = const()[name = tensor<string, []>("inputs_13_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> inputs_13_pad_0 = const()[name = tensor<string, []>("inputs_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> inputs_13_dilations_0 = const()[name = tensor<string, []>("inputs_13_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> inputs_13_groups_0 = const()[name = tensor<string, []>("inputs_13_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
-            tensor<fp32, [2, 512, 16]> inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor<string, []>("inputs_13")];
-            tensor<int32, [2]> var_455_split_sizes_0 = const()[name = tensor<string, []>("op_455_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_455_axis_0 = const()[name = tensor<string, []>("op_455_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> var_455_0, tensor<fp32, [2, 256, 16]> var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_455")];
-            tensor<fp32, [2, 256, 16]> var_457 = sigmoid(x = var_455_1)[name = tensor<string, []>("op_457")];
-            tensor<fp32, [2, 256, 16]> inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor<string, []>("inputs_15")];
+            tensor<fp32, [2, 256, 16]> input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
+            tensor<fp32, [2, 512, 16]> inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor<string, []>("inputs_13")];
+            tensor<int32, [2]> var_491_split_sizes_0 = const()[name = tensor<string, []>("op_491_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_491_axis_0 = const()[name = tensor<string, []>("op_491_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> var_491_0, tensor<fp32, [2, 256, 16]> var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_491")];
+            tensor<fp32, [2, 256, 16]> var_493 = sigmoid(x = var_491_1)[name = tensor<string, []>("op_493")];
+            tensor<fp32, [2, 256, 16]> inputs_15 = mul(x = var_491_0, y = var_493)[name = tensor<string, []>("inputs_15")];
             tensor<string, []> outputs_aug_3_pad_type_0 = const()[name = tensor<string, []>("outputs_aug_3_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> outputs_aug_3_pad_0 = const()[name = tensor<string, []>("outputs_aug_3_pad_0"), val = tensor<int32, [2]>([15, 15])];
             tensor<int32, []> outputs_aug_3_groups_0 = const()[name = tensor<string, []>("outputs_aug_3_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [1]> outputs_aug_3_strides_0 = const()[name = tensor<string, []>("outputs_aug_3_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [1]> outputs_aug_3_dilations_0 = const()[name = tensor<string, []>("outputs_aug_3_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [2, 256, 31]> outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor<string, []>("outputs_aug_3")];
-            tensor<int32, [3]> input_65_begin_0 = const()[name = tensor<string, []>("input_65_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> input_65_end_0 = const()[name = tensor<string, []>("input_65_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
-            tensor<bool, [3]> input_65_end_mask_0 = const()[name = tensor<string, []>("input_65_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
-            tensor<fp32, [2, 256, 16]> input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor<string, []>("input_65")];
-            tensor<fp32, [2, 256, 16]> inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor<string, []>("inputs_17")];
-            tensor<fp32, [2, 256, 16]> input_67 = silu(x = inputs_17)[name = tensor<string, []>("input_67")];
-            tensor<string, []> input_69_pad_type_0 = const()[name = tensor<string, []>("input_69_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [1]> input_69_strides_0 = const()[name = tensor<string, []>("input_69_strides_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [2]> input_69_pad_0 = const()[name = tensor<string, []>("input_69_pad_0"), val = tensor<int32, [2]>([0, 0])];
-            tensor<int32, [1]> input_69_dilations_0 = const()[name = tensor<string, []>("input_69_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, []> input_69_groups_0 = const()[name = tensor<string, []>("input_69_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor<string, []>("input_69")];
+            tensor<fp32, [2, 256, 31]> outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor<string, []>("outputs_aug_3")];
+            tensor<int32, [3]> input_67_begin_0 = const()[name = tensor<string, []>("input_67_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_67_end_0 = const()[name = tensor<string, []>("input_67_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
+            tensor<bool, [3]> input_67_end_mask_0 = const()[name = tensor<string, []>("input_67_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp32, [2, 256, 16]> input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor<string, []>("input_67")];
+            tensor<fp32, [2, 256, 16]> inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor<string, []>("inputs_17")];
+            tensor<fp32, [2, 256, 16]> input_69 = silu(x = inputs_17)[name = tensor<string, []>("input_69")];
+            tensor<string, []> input_71_pad_type_0 = const()[name = tensor<string, []>("input_71_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_71_strides_0 = const()[name = tensor<string, []>("input_71_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_71_pad_0 = const()[name = tensor<string, []>("input_71_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_71_dilations_0 = const()[name = tensor<string, []>("input_71_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_71_groups_0 = const()[name = tensor<string, []>("input_71_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor<string, []>("input_71")];
             tensor<int32, [3]> conv_out_3_perm_0 = const()[name = tensor<string, []>("conv_out_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<int32, [3]> var_488_begin_0 = const()[name = tensor<string, []>("op_488_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_488_end_0 = const()[name = tensor<string, []>("op_488_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
-            tensor<bool, [3]> var_488_end_mask_0 = const()[name = tensor<string, []>("op_488_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [2, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor<string, []>("transpose_46")];
-            tensor<fp32, [2, 1, 256]> var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_488")];
-            tensor<int32, [3]> var_490_perm_0 = const()[name = tensor<string, []>("op_490_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 2, 256]> var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor<string, []>("transpose_45")];
-            tensor<fp32, [1, 2, 256]> input_71 = add(x = x_9, y = var_490)[name = tensor<string, []>("input_71")];
-            tensor<int32, [1]> input_73_axes_0 = const()[name = tensor<string, []>("input_73_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor<string, []>("input_73")];
-            tensor<fp32, [1, 2, 1024]> inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor<string, []>("linear_17")];
-            tensor<fp32, [1, 2, 1024]> input_75 = silu(x = inputs_19)[name = tensor<string, []>("input_75")];
-            tensor<fp32, [1, 2, 256]> input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor<string, []>("linear_18")];
-            tensor<fp32, []> var_513 = const()[name = tensor<string, []>("op_513"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_514 = mul(x = input_79, y = var_513)[name = tensor<string, []>("op_514")];
-            tensor<fp32, [1, 2, 256]> input_81 = add(x = var_514, y = input_71)[name = tensor<string, []>("input_81")];
-            tensor<int32, [1]> input_83_axes_0 = const()[name = tensor<string, []>("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor<string, []>("input_83")];
+            tensor<int32, [3]> var_524_begin_0 = const()[name = tensor<string, []>("op_524_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_524_end_0 = const()[name = tensor<string, []>("op_524_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
+            tensor<bool, [3]> var_524_end_mask_0 = const()[name = tensor<string, []>("op_524_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [2, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor<string, []>("transpose_46")];
+            tensor<fp32, [2, 1, 256]> var_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_524")];
+            tensor<int32, [3]> var_526_perm_0 = const()[name = tensor<string, []>("op_526_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 2, 256]> var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor<string, []>("transpose_45")];
+            tensor<fp32, [1, 2, 256]> input_73 = add(x = x_9, y = var_526)[name = tensor<string, []>("input_73")];
+            tensor<int32, [1]> input_75_axes_0 = const()[name = tensor<string, []>("input_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor<string, []>("input_75")];
+            tensor<fp32, [1, 2, 1024]> inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor<string, []>("linear_17")];
+            tensor<fp32, [1, 2, 1024]> input_77 = silu(x = inputs_19)[name = tensor<string, []>("input_77")];
+            tensor<fp32, [1, 2, 256]> input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor<string, []>("linear_18")];
+            tensor<fp32, []> var_549 = const()[name = tensor<string, []>("op_549"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_550 = mul(x = input_81, y = var_549)[name = tensor<string, []>("op_550")];
+            tensor<fp32, [1, 2, 256]> input_83 = add(x = var_550, y = input_73)[name = tensor<string, []>("input_83")];
             tensor<int32, [1]> input_85_axes_0 = const()[name = tensor<string, []>("input_85_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor<string, []>("input_85")];
-            tensor<fp32, [1, 2, 1024]> inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor<string, []>("linear_19")];
-            tensor<fp32, [1, 2, 1024]> input_87 = silu(x = inputs_21)[name = tensor<string, []>("input_87")];
-            tensor<fp32, [1, 2, 256]> input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor<string, []>("linear_20")];
-            tensor<fp32, []> var_543 = const()[name = tensor<string, []>("op_543"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_544 = mul(x = input_91, y = var_543)[name = tensor<string, []>("op_544")];
-            tensor<fp32, [1, 2, 256]> input_93 = add(x = var_544, y = input_83)[name = tensor<string, []>("input_93")];
+            tensor<fp32, [1, 2, 256]> input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor<string, []>("input_85")];
+            tensor<int32, [1]> input_87_axes_0 = const()[name = tensor<string, []>("input_87_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor<string, []>("input_87")];
+            tensor<fp32, [1, 2, 1024]> inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor<string, []>("linear_19")];
+            tensor<fp32, [1, 2, 1024]> input_89 = silu(x = inputs_21)[name = tensor<string, []>("input_89")];
+            tensor<fp32, [1, 2, 256]> input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor<string, []>("linear_20")];
+            tensor<fp32, []> var_579 = const()[name = tensor<string, []>("op_579"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_580 = mul(x = input_93, y = var_579)[name = tensor<string, []>("op_580")];
+            tensor<fp32, [1, 2, 256]> input_95 = add(x = var_580, y = input_85)[name = tensor<string, []>("input_95")];
             tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor<string, []>("x_13")];
+            tensor<fp32, [1, 2, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor<string, []>("x_13")];
             tensor<int32, [5]> prev_kv_5_begin_0 = const()[name = tensor<string, []>("prev_kv_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_5_end_0 = const()[name = tensor<string, []>("prev_kv_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_5_end_mask_0 = const()[name = tensor<string, []>("prev_kv_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -553,153 +567,153 @@ program(1.0)
             tensor<bool, [2]> prev_scale_5_end_mask_0 = const()[name = tensor<string, []>("prev_scale_5_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_5_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_5_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor<string, []>("prev_scale_5")];
-            tensor<fp32, [1, 2, 256]> var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_21")];
-            tensor<int32, [4]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_560 = reshape(shape = var_559, x = var_558)[name = tensor<string, []>("op_560")];
+            tensor<fp32, [1, 2, 256]> var_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_21")];
+            tensor<int32, [4]> var_595 = const()[name = tensor<string, []>("op_595"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_596 = reshape(shape = var_595, x = var_594)[name = tensor<string, []>("op_596")];
             tensor<int32, [4]> q_5_perm_0 = const()[name = tensor<string, []>("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_22")];
-            tensor<fp32, []> var_565 = const()[name = tensor<string, []>("op_565"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 2, 256]> var_566 = mul(x = var_564, y = var_565)[name = tensor<string, []>("op_566")];
-            tensor<int32, [4]> var_567 = const()[name = tensor<string, []>("op_567"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_568 = reshape(shape = var_567, x = var_566)[name = tensor<string, []>("op_568")];
+            tensor<fp32, [1, 2, 256]> var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_22")];
+            tensor<fp32, []> var_601 = const()[name = tensor<string, []>("op_601"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 2, 256]> var_602 = mul(x = var_600, y = var_601)[name = tensor<string, []>("op_602")];
+            tensor<int32, [4]> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_604 = reshape(shape = var_603, x = var_602)[name = tensor<string, []>("op_604")];
             tensor<int32, [4]> k_5_perm_0 = const()[name = tensor<string, []>("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_23")];
-            tensor<int32, [4]> var_573 = const()[name = tensor<string, []>("op_573"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_574 = reshape(shape = var_573, x = var_572)[name = tensor<string, []>("op_574")];
+            tensor<fp32, [1, 2, 256]> var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_23")];
+            tensor<int32, [4]> var_609 = const()[name = tensor<string, []>("op_609"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_610 = reshape(shape = var_609, x = var_608)[name = tensor<string, []>("op_610")];
             tensor<int32, [4]> v_5_perm_0 = const()[name = tensor<string, []>("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_24")];
+            tensor<fp32, [1, 2, 256]> input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_24")];
             tensor<fp32, [1]> sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor<string, []>("sqrt_s0_5")];
-            tensor<fp32, [2]> s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor<string, []>("s_t_5")];
+            tensor<fp32, [2]> s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_5")];
             tensor<fp32, [2]> sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor<string, []>("sqrt_s_t_5")];
             tensor<bool, []> qk_5_transpose_x_1 = const()[name = tensor<string, []>("qk_5_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_5_transpose_y_1 = const()[name = tensor<string, []>("qk_5_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [1, 4, 2, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor<string, []>("transpose_43")];
-            tensor<fp32, [1, 4, 2, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor<string, []>("transpose_44")];
+            tensor<fp32, [1, 4, 2, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_604)[name = tensor<string, []>("transpose_43")];
+            tensor<fp32, [1, 4, 2, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_596)[name = tensor<string, []>("transpose_44")];
             tensor<fp32, [1, 4, 2, 2]> qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor<string, []>("qk_5")];
-            tensor<int32, [2]> var_584 = const()[name = tensor<string, []>("op_584"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor<string, []>("op_585")];
-            tensor<fp32, [2, 2]> M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor<string, []>("M_5")];
-            tensor<fp32, [1, 4, 2, 2]> var_587 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_587")];
+            tensor<int32, [2]> var_620 = const()[name = tensor<string, []>("op_620"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor<string, []>("op_621")];
+            tensor<fp32, [2, 2]> M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor<string, []>("M_5")];
+            tensor<fp32, [1, 4, 2, 2]> var_623 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_623")];
             tensor<bool, []> inner_5_transpose_x_0 = const()[name = tensor<string, []>("inner_5_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_5_transpose_y_0 = const()[name = tensor<string, []>("inner_5_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor<string, []>("transpose_42")];
-            tensor<fp32, [1, 4, 2, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor<string, []>("inner_5")];
-            tensor<bool, []> var_589_transpose_x_0 = const()[name = tensor<string, []>("op_589_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_589_transpose_y_0 = const()[name = tensor<string, []>("op_589_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_589")];
-            tensor<fp32, [2]> var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_590")];
-            tensor<int32, [4]> var_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_592 = reshape(shape = var_591, x = var_590)[name = tensor<string, []>("op_592")];
-            tensor<fp32, [1, 4, 2, 64]> cross_5 = mul(x = var_589, y = var_592)[name = tensor<string, []>("cross_5")];
+            tensor<fp32, [1, 4, 2, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_610)[name = tensor<string, []>("transpose_42")];
+            tensor<fp32, [1, 4, 2, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor<string, []>("inner_5")];
+            tensor<bool, []> var_625_transpose_x_0 = const()[name = tensor<string, []>("op_625_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_625_transpose_y_0 = const()[name = tensor<string, []>("op_625_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 2, 64]> var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_625")];
+            tensor<fp32, [2]> var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_626")];
+            tensor<int32, [4]> var_627 = const()[name = tensor<string, []>("op_627"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_628 = reshape(shape = var_627, x = var_626)[name = tensor<string, []>("op_628")];
+            tensor<fp32, [1, 4, 2, 64]> cross_5 = mul(x = var_625, y = var_628)[name = tensor<string, []>("cross_5")];
             tensor<fp32, [1, 4, 2, 64]> out_13 = add(x = inner_5, y = cross_5)[name = tensor<string, []>("out_13")];
-            tensor<fp32, [1, 4, 64, 64]> var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_595")];
-            tensor<bool, []> var_597_transpose_x_1 = const()[name = tensor<string, []>("op_597_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_597_transpose_y_1 = const()[name = tensor<string, []>("op_597_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_597")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor<string, []>("new_kv_unnorm_5")];
-            tensor<fp32, []> var_599 = const()[name = tensor<string, []>("op_599"), val = tensor<fp32, []>(0x1p+1)];
-            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor<string, []>("new_scale_5")];
-            tensor<fp32, [1]> var_601 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_601")];
-            tensor<fp32, [1, 4, 64, 64]> var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor<string, []>("op_602")];
-            tensor<int32, [4]> var_603_perm_0 = const()[name = tensor<string, []>("op_603_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_631")];
+            tensor<bool, []> var_633_transpose_x_1 = const()[name = tensor<string, []>("op_633_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_633_transpose_y_1 = const()[name = tensor<string, []>("op_633_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_633")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor<string, []>("new_kv_unnorm_5")];
+            tensor<fp32, []> var_635 = const()[name = tensor<string, []>("op_635"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor<string, []>("new_scale_5")];
+            tensor<fp32, [1]> var_637 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_637")];
+            tensor<fp32, [1, 4, 64, 64]> var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor<string, []>("op_638")];
+            tensor<int32, [4]> var_639_perm_0 = const()[name = tensor<string, []>("op_639_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_15_axes_0 = const()[name = tensor<string, []>("out_15_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 4, 64]> var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
-            tensor<fp32, [1, 2, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor<string, []>("out_15")];
-            tensor<int32, [3]> var_607 = const()[name = tensor<string, []>("op_607"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<fp32, [1, 2, 256]> out_17 = reshape(shape = var_607, x = out_15)[name = tensor<string, []>("out_17")];
-            tensor<fp32, [1, 2, 256]> var_609 = silu(x = input_97)[name = tensor<string, []>("op_609")];
-            tensor<fp32, [1, 2, 256]> input_99 = mul(x = var_609, y = out_17)[name = tensor<string, []>("input_99")];
-            tensor<fp32, [1, 2, 256]> ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor<string, []>("linear_25")];
-            tensor<fp32, [1, 2, 256]> x_15 = add(x = input_93, y = ret_out_5)[name = tensor<string, []>("x_15")];
+            tensor<fp32, [1, 2, 4, 64]> var_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
+            tensor<fp32, [1, 2, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_53, x = var_639)[name = tensor<string, []>("out_15")];
+            tensor<int32, [3]> var_643 = const()[name = tensor<string, []>("op_643"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<fp32, [1, 2, 256]> out_17 = reshape(shape = var_643, x = out_15)[name = tensor<string, []>("out_17")];
+            tensor<fp32, [1, 2, 256]> var_645 = silu(x = input_99)[name = tensor<string, []>("op_645")];
+            tensor<fp32, [1, 2, 256]> input_101 = mul(x = var_645, y = out_17)[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 2, 256]> ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor<string, []>("linear_25")];
+            tensor<fp32, [1, 2, 256]> x_15 = add(x = input_95, y = ret_out_5)[name = tensor<string, []>("x_15")];
             tensor<int32, [4]> window_13_begin_0 = const()[name = tensor<string, []>("window_13_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> window_13_end_0 = const()[name = tensor<string, []>("window_13_end_0"), val = tensor<int32, [4]>([3, 1, 16, 256])];
             tensor<bool, [4]> window_13_end_mask_0 = const()[name = tensor<string, []>("window_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_13_squeeze_mask_0 = const()[name = tensor<string, []>("window_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_13")];
-            tensor<int32, [3]> var_617_begin_0 = const()[name = tensor<string, []>("op_617_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_617_end_0 = const()[name = tensor<string, []>("op_617_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_617_end_mask_0 = const()[name = tensor<string, []>("op_617_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor<string, []>("op_617")];
-            tensor<int32, [3]> var_620_begin_0 = const()[name = tensor<string, []>("op_620_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_620_end_0 = const()[name = tensor<string, []>("op_620_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_620_end_mask_0 = const()[name = tensor<string, []>("op_620_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor<string, []>("op_620")];
+            tensor<int32, [3]> var_653_begin_0 = const()[name = tensor<string, []>("op_653_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_653_end_0 = const()[name = tensor<string, []>("op_653_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_653_end_mask_0 = const()[name = tensor<string, []>("op_653_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor<string, []>("op_653")];
+            tensor<int32, [3]> var_656_begin_0 = const()[name = tensor<string, []>("op_656_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_656_end_0 = const()[name = tensor<string, []>("op_656_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_656_end_mask_0 = const()[name = tensor<string, []>("op_656_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor<string, []>("op_656")];
             tensor<bool, []> window_15_interleave_0 = const()[name = tensor<string, []>("window_15_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor<string, []>("window_15")];
-            tensor<int32, [3]> var_625_begin_0 = const()[name = tensor<string, []>("op_625_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_625_end_0 = const()[name = tensor<string, []>("op_625_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_625_end_mask_0 = const()[name = tensor<string, []>("op_625_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor<string, []>("op_625")];
-            tensor<int32, [3]> var_628_begin_0 = const()[name = tensor<string, []>("op_628_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_628_end_0 = const()[name = tensor<string, []>("op_628_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_628_end_mask_0 = const()[name = tensor<string, []>("op_628_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor<string, []>("op_628")];
+            tensor<fp32, [1, 16, 256]> window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor<string, []>("window_15")];
+            tensor<int32, [3]> var_661_begin_0 = const()[name = tensor<string, []>("op_661_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_661_end_0 = const()[name = tensor<string, []>("op_661_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_661_end_mask_0 = const()[name = tensor<string, []>("op_661_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor<string, []>("op_661")];
+            tensor<int32, [3]> var_664_begin_0 = const()[name = tensor<string, []>("op_664_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_664_end_0 = const()[name = tensor<string, []>("op_664_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_664_end_mask_0 = const()[name = tensor<string, []>("op_664_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor<string, []>("op_664")];
             tensor<bool, []> window_17_interleave_0 = const()[name = tensor<string, []>("window_17_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor<string, []>("window_17")];
-            tensor<bool, []> input_101_interleave_0 = const()[name = tensor<string, []>("input_101_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [2, 16, 256]> input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 16, 256]> window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor<string, []>("window_17")];
+            tensor<bool, []> input_103_interleave_0 = const()[name = tensor<string, []>("input_103_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [2, 16, 256]> input_103 = concat(axis = var_48, interleave = input_103_interleave_0, values = (window_15, window_17))[name = tensor<string, []>("input_103")];
             tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 16, 256]> x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor<string, []>("x_17")];
-            tensor<int32, [3]> input_103_perm_0 = const()[name = tensor<string, []>("input_103_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [2, 16, 256]> x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor<string, []>("x_17")];
+            tensor<int32, [3]> input_105_perm_0 = const()[name = tensor<string, []>("input_105_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> inputs_23_pad_type_0 = const()[name = tensor<string, []>("inputs_23_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> inputs_23_strides_0 = const()[name = tensor<string, []>("inputs_23_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> inputs_23_pad_0 = const()[name = tensor<string, []>("inputs_23_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> inputs_23_dilations_0 = const()[name = tensor<string, []>("inputs_23_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> inputs_23_groups_0 = const()[name = tensor<string, []>("inputs_23_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
-            tensor<fp32, [2, 512, 16]> inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor<string, []>("inputs_23")];
-            tensor<int32, [2]> var_653_split_sizes_0 = const()[name = tensor<string, []>("op_653_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_653_axis_0 = const()[name = tensor<string, []>("op_653_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> var_653_0, tensor<fp32, [2, 256, 16]> var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_653")];
-            tensor<fp32, [2, 256, 16]> var_655 = sigmoid(x = var_653_1)[name = tensor<string, []>("op_655")];
-            tensor<fp32, [2, 256, 16]> inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor<string, []>("inputs_25")];
+            tensor<fp32, [2, 256, 16]> input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
+            tensor<fp32, [2, 512, 16]> inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor<string, []>("inputs_23")];
+            tensor<int32, [2]> var_689_split_sizes_0 = const()[name = tensor<string, []>("op_689_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_689_axis_0 = const()[name = tensor<string, []>("op_689_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> var_689_0, tensor<fp32, [2, 256, 16]> var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_689")];
+            tensor<fp32, [2, 256, 16]> var_691 = sigmoid(x = var_689_1)[name = tensor<string, []>("op_691")];
+            tensor<fp32, [2, 256, 16]> inputs_25 = mul(x = var_689_0, y = var_691)[name = tensor<string, []>("inputs_25")];
             tensor<string, []> outputs_aug_5_pad_type_0 = const()[name = tensor<string, []>("outputs_aug_5_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> outputs_aug_5_pad_0 = const()[name = tensor<string, []>("outputs_aug_5_pad_0"), val = tensor<int32, [2]>([15, 15])];
             tensor<int32, []> outputs_aug_5_groups_0 = const()[name = tensor<string, []>("outputs_aug_5_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [1]> outputs_aug_5_strides_0 = const()[name = tensor<string, []>("outputs_aug_5_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [1]> outputs_aug_5_dilations_0 = const()[name = tensor<string, []>("outputs_aug_5_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [2, 256, 31]> outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor<string, []>("outputs_aug_5")];
-            tensor<int32, [3]> input_105_begin_0 = const()[name = tensor<string, []>("input_105_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> input_105_end_0 = const()[name = tensor<string, []>("input_105_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
-            tensor<bool, [3]> input_105_end_mask_0 = const()[name = tensor<string, []>("input_105_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
-            tensor<fp32, [2, 256, 16]> input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor<string, []>("input_105")];
-            tensor<fp32, [2, 256, 16]> inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor<string, []>("inputs_27")];
-            tensor<fp32, [2, 256, 16]> input_107 = silu(x = inputs_27)[name = tensor<string, []>("input_107")];
-            tensor<string, []> input_109_pad_type_0 = const()[name = tensor<string, []>("input_109_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [1]> input_109_strides_0 = const()[name = tensor<string, []>("input_109_strides_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [2]> input_109_pad_0 = const()[name = tensor<string, []>("input_109_pad_0"), val = tensor<int32, [2]>([0, 0])];
-            tensor<int32, [1]> input_109_dilations_0 = const()[name = tensor<string, []>("input_109_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, []> input_109_groups_0 = const()[name = tensor<string, []>("input_109_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor<string, []>("input_109")];
+            tensor<fp32, [2, 256, 31]> outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor<string, []>("outputs_aug_5")];
+            tensor<int32, [3]> input_107_begin_0 = const()[name = tensor<string, []>("input_107_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_107_end_0 = const()[name = tensor<string, []>("input_107_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
+            tensor<bool, [3]> input_107_end_mask_0 = const()[name = tensor<string, []>("input_107_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp32, [2, 256, 16]> input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor<string, []>("input_107")];
+            tensor<fp32, [2, 256, 16]> inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor<string, []>("inputs_27")];
+            tensor<fp32, [2, 256, 16]> input_109 = silu(x = inputs_27)[name = tensor<string, []>("input_109")];
+            tensor<string, []> input_111_pad_type_0 = const()[name = tensor<string, []>("input_111_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_111_strides_0 = const()[name = tensor<string, []>("input_111_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_111_pad_0 = const()[name = tensor<string, []>("input_111_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_111_dilations_0 = const()[name = tensor<string, []>("input_111_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_111_groups_0 = const()[name = tensor<string, []>("input_111_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor<string, []>("input_111")];
             tensor<int32, [3]> conv_out_5_perm_0 = const()[name = tensor<string, []>("conv_out_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<int32, [3]> var_686_begin_0 = const()[name = tensor<string, []>("op_686_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_686_end_0 = const()[name = tensor<string, []>("op_686_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
-            tensor<bool, [3]> var_686_end_mask_0 = const()[name = tensor<string, []>("op_686_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [2, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor<string, []>("transpose_39")];
-            tensor<fp32, [2, 1, 256]> var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_686")];
-            tensor<int32, [3]> var_688_perm_0 = const()[name = tensor<string, []>("op_688_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 2, 256]> var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor<string, []>("transpose_38")];
-            tensor<fp32, [1, 2, 256]> input_111 = add(x = x_15, y = var_688)[name = tensor<string, []>("input_111")];
-            tensor<int32, [1]> input_113_axes_0 = const()[name = tensor<string, []>("input_113_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor<string, []>("input_113")];
-            tensor<fp32, [1, 2, 1024]> inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor<string, []>("linear_26")];
-            tensor<fp32, [1, 2, 1024]> input_115 = silu(x = inputs_29)[name = tensor<string, []>("input_115")];
-            tensor<fp32, [1, 2, 256]> input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor<string, []>("linear_27")];
-            tensor<fp32, []> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_712 = mul(x = input_119, y = var_711)[name = tensor<string, []>("op_712")];
-            tensor<fp32, [1, 2, 256]> input_121 = add(x = var_712, y = input_111)[name = tensor<string, []>("input_121")];
-            tensor<int32, [1]> input_123_axes_0 = const()[name = tensor<string, []>("input_123_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor<string, []>("input_123")];
+            tensor<int32, [3]> var_722_begin_0 = const()[name = tensor<string, []>("op_722_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_722_end_0 = const()[name = tensor<string, []>("op_722_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
+            tensor<bool, [3]> var_722_end_mask_0 = const()[name = tensor<string, []>("op_722_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [2, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor<string, []>("transpose_39")];
+            tensor<fp32, [2, 1, 256]> var_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_722")];
+            tensor<int32, [3]> var_724_perm_0 = const()[name = tensor<string, []>("op_724_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 2, 256]> var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor<string, []>("transpose_38")];
+            tensor<fp32, [1, 2, 256]> input_113 = add(x = x_15, y = var_724)[name = tensor<string, []>("input_113")];
+            tensor<int32, [1]> input_115_axes_0 = const()[name = tensor<string, []>("input_115_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor<string, []>("input_115")];
+            tensor<fp32, [1, 2, 1024]> inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor<string, []>("linear_26")];
+            tensor<fp32, [1, 2, 1024]> input_117 = silu(x = inputs_29)[name = tensor<string, []>("input_117")];
+            tensor<fp32, [1, 2, 256]> input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor<string, []>("linear_27")];
+            tensor<fp32, []> var_747 = const()[name = tensor<string, []>("op_747"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_748 = mul(x = input_121, y = var_747)[name = tensor<string, []>("op_748")];
+            tensor<fp32, [1, 2, 256]> input_123 = add(x = var_748, y = input_113)[name = tensor<string, []>("input_123")];
             tensor<int32, [1]> input_125_axes_0 = const()[name = tensor<string, []>("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor<string, []>("input_125")];
-            tensor<fp32, [1, 2, 1024]> inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor<string, []>("linear_28")];
-            tensor<fp32, [1, 2, 1024]> input_127 = silu(x = inputs_31)[name = tensor<string, []>("input_127")];
-            tensor<fp32, [1, 2, 256]> input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor<string, []>("linear_29")];
-            tensor<fp32, []> var_741 = const()[name = tensor<string, []>("op_741"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_742 = mul(x = input_131, y = var_741)[name = tensor<string, []>("op_742")];
-            tensor<fp32, [1, 2, 256]> input_133 = add(x = var_742, y = input_123)[name = tensor<string, []>("input_133")];
+            tensor<fp32, [1, 2, 256]> input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor<string, []>("input_125")];
+            tensor<int32, [1]> input_127_axes_0 = const()[name = tensor<string, []>("input_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor<string, []>("input_127")];
+            tensor<fp32, [1, 2, 1024]> inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor<string, []>("linear_28")];
+            tensor<fp32, [1, 2, 1024]> input_129 = silu(x = inputs_31)[name = tensor<string, []>("input_129")];
+            tensor<fp32, [1, 2, 256]> input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor<string, []>("linear_29")];
+            tensor<fp32, []> var_777 = const()[name = tensor<string, []>("op_777"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_778 = mul(x = input_133, y = var_777)[name = tensor<string, []>("op_778")];
+            tensor<fp32, [1, 2, 256]> input_135 = add(x = var_778, y = input_125)[name = tensor<string, []>("input_135")];
             tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor<string, []>("x_19")];
+            tensor<fp32, [1, 2, 256]> x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor<string, []>("x_19")];
             tensor<int32, [5]> prev_kv_7_begin_0 = const()[name = tensor<string, []>("prev_kv_7_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_7_end_0 = const()[name = tensor<string, []>("prev_kv_7_end_0"), val = tensor<int32, [5]>([4, 1, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_7_end_mask_0 = const()[name = tensor<string, []>("prev_kv_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -710,189 +724,182 @@ program(1.0)
             tensor<bool, [2]> prev_scale_7_end_mask_0 = const()[name = tensor<string, []>("prev_scale_7_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_7_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_7_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor<string, []>("prev_scale_7")];
-            tensor<fp32, [1, 2, 256]> var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_30")];
-            tensor<int32, [4]> var_757 = const()[name = tensor<string, []>("op_757"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_758 = reshape(shape = var_757, x = var_756)[name = tensor<string, []>("op_758")];
+            tensor<fp32, [1, 2, 256]> var_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_30")];
+            tensor<int32, [4]> var_793 = const()[name = tensor<string, []>("op_793"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_794 = reshape(shape = var_793, x = var_792)[name = tensor<string, []>("op_794")];
             tensor<int32, [4]> q_7_perm_0 = const()[name = tensor<string, []>("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_31")];
-            tensor<fp32, []> var_763 = const()[name = tensor<string, []>("op_763"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 2, 256]> var_764 = mul(x = var_762, y = var_763)[name = tensor<string, []>("op_764")];
-            tensor<int32, [4]> var_765 = const()[name = tensor<string, []>("op_765"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_766 = reshape(shape = var_765, x = var_764)[name = tensor<string, []>("op_766")];
+            tensor<fp32, [1, 2, 256]> var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_31")];
+            tensor<fp32, []> var_799 = const()[name = tensor<string, []>("op_799"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 2, 256]> var_800 = mul(x = var_798, y = var_799)[name = tensor<string, []>("op_800")];
+            tensor<int32, [4]> var_801 = const()[name = tensor<string, []>("op_801"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_802 = reshape(shape = var_801, x = var_800)[name = tensor<string, []>("op_802")];
             tensor<int32, [4]> k_7_perm_0 = const()[name = tensor<string, []>("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_32")];
-            tensor<int32, [4]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [4]>([1, 2, 4, 64])];
-            tensor<fp32, [1, 2, 4, 64]> var_772 = reshape(shape = var_771, x = var_770)[name = tensor<string, []>("op_772")];
+            tensor<fp32, [1, 2, 256]> var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_32")];
+            tensor<int32, [4]> var_807 = const()[name = tensor<string, []>("op_807"), val = tensor<int32, [4]>([1, 2, 4, 64])];
+            tensor<fp32, [1, 2, 4, 64]> var_808 = reshape(shape = var_807, x = var_806)[name = tensor<string, []>("op_808")];
             tensor<int32, [4]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [1, 2, 256]> input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_33")];
+            tensor<fp32, [1, 2, 256]> input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_33")];
             tensor<fp32, [1]> sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor<string, []>("sqrt_s0_7")];
-            tensor<fp32, [2]> s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor<string, []>("s_t_7")];
+            tensor<fp32, [2]> s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_7")];
             tensor<fp32, [2]> sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor<string, []>("sqrt_s_t_7")];
             tensor<bool, []> qk_7_transpose_x_1 = const()[name = tensor<string, []>("qk_7_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_7_transpose_y_1 = const()[name = tensor<string, []>("qk_7_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [1, 4, 2, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor<string, []>("transpose_36")];
-            tensor<fp32, [1, 4, 2, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor<string, []>("transpose_37")];
+            tensor<fp32, [1, 4, 2, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_802)[name = tensor<string, []>("transpose_36")];
+            tensor<fp32, [1, 4, 2, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_794)[name = tensor<string, []>("transpose_37")];
             tensor<fp32, [1, 4, 2, 2]> qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor<string, []>("qk_7")];
-            tensor<int32, [2]> var_782 = const()[name = tensor<string, []>("op_782"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor<string, []>("op_783")];
-            tensor<fp32, [2, 2]> M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor<string, []>("M_7")];
-            tensor<fp32, [1, 4, 2, 2]> var_785 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_785")];
+            tensor<int32, [2]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor<string, []>("op_819")];
+            tensor<fp32, [2, 2]> M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor<string, []>("M_7")];
+            tensor<fp32, [1, 4, 2, 2]> var_821 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_821")];
             tensor<bool, []> inner_7_transpose_x_0 = const()[name = tensor<string, []>("inner_7_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_7_transpose_y_0 = const()[name = tensor<string, []>("inner_7_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor<string, []>("transpose_35")];
-            tensor<fp32, [1, 4, 2, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor<string, []>("inner_7")];
-            tensor<bool, []> var_787_transpose_x_0 = const()[name = tensor<string, []>("op_787_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_787_transpose_y_0 = const()[name = tensor<string, []>("op_787_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 2, 64]> var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_787")];
-            tensor<fp32, [2]> var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_788")];
-            tensor<int32, [4]> var_789 = const()[name = tensor<string, []>("op_789"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_790 = reshape(shape = var_789, x = var_788)[name = tensor<string, []>("op_790")];
-            tensor<fp32, [1, 4, 2, 64]> cross_7 = mul(x = var_787, y = var_790)[name = tensor<string, []>("cross_7")];
+            tensor<fp32, [1, 4, 2, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_808)[name = tensor<string, []>("transpose_35")];
+            tensor<fp32, [1, 4, 2, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor<string, []>("inner_7")];
+            tensor<bool, []> var_823_transpose_x_0 = const()[name = tensor<string, []>("op_823_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_823_transpose_y_0 = const()[name = tensor<string, []>("op_823_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 2, 64]> var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_823")];
+            tensor<fp32, [2]> var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_824")];
+            tensor<int32, [4]> var_825 = const()[name = tensor<string, []>("op_825"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_826 = reshape(shape = var_825, x = var_824)[name = tensor<string, []>("op_826")];
+            tensor<fp32, [1, 4, 2, 64]> cross_7 = mul(x = var_823, y = var_826)[name = tensor<string, []>("cross_7")];
             tensor<fp32, [1, 4, 2, 64]> out_19 = add(x = inner_7, y = cross_7)[name = tensor<string, []>("out_19")];
-            tensor<fp32, [1, 4, 64, 64]> var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_793")];
-            tensor<bool, []> var_795_transpose_x_1 = const()[name = tensor<string, []>("op_795_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_795_transpose_y_1 = const()[name = tensor<string, []>("op_795_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_795")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor<string, []>("new_kv_unnorm_7")];
-            tensor<fp32, []> var_797 = const()[name = tensor<string, []>("op_797"), val = tensor<fp32, []>(0x1p+1)];
-            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor<string, []>("new_scale_7")];
-            tensor<fp32, [1]> var_799 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_799")];
-            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor<string, []>("nkv_1")];
-            tensor<int32, [4]> var_801_perm_0 = const()[name = tensor<string, []>("op_801_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_829")];
+            tensor<bool, []> var_831_transpose_x_1 = const()[name = tensor<string, []>("op_831_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_831_transpose_y_1 = const()[name = tensor<string, []>("op_831_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_831")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor<string, []>("new_kv_unnorm_7")];
+            tensor<fp32, []> var_833 = const()[name = tensor<string, []>("op_833"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor<string, []>("new_scale_7")];
+            tensor<fp32, [1]> var_835 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_835")];
+            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor<string, []>("nkv_1")];
+            tensor<int32, [4]> var_837_perm_0 = const()[name = tensor<string, []>("op_837_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_21_axes_0 = const()[name = tensor<string, []>("out_21_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 4, 64]> var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
-            tensor<fp32, [1, 2, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor<string, []>("out_21")];
-            tensor<int32, [3]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<fp32, [1, 2, 256]> out_23 = reshape(shape = var_805, x = out_21)[name = tensor<string, []>("out_23")];
-            tensor<fp32, [1, 2, 256]> var_807 = silu(x = input_137)[name = tensor<string, []>("op_807")];
-            tensor<fp32, [1, 2, 256]> input_139 = mul(x = var_807, y = out_23)[name = tensor<string, []>("input_139")];
-            tensor<fp32, [1, 2, 256]> ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor<string, []>("linear_34")];
-            tensor<fp32, [1, 2, 256]> x_21 = add(x = input_133, y = ret_out_7)[name = tensor<string, []>("x_21")];
+            tensor<fp32, [1, 2, 4, 64]> var_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
+            tensor<fp32, [1, 2, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_53, x = var_837)[name = tensor<string, []>("out_21")];
+            tensor<int32, [3]> var_841 = const()[name = tensor<string, []>("op_841"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<fp32, [1, 2, 256]> out_23 = reshape(shape = var_841, x = out_21)[name = tensor<string, []>("out_23")];
+            tensor<fp32, [1, 2, 256]> var_843 = silu(x = input_139)[name = tensor<string, []>("op_843")];
+            tensor<fp32, [1, 2, 256]> input_141 = mul(x = var_843, y = out_23)[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 2, 256]> ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor<string, []>("linear_34")];
+            tensor<fp32, [1, 2, 256]> x_21 = add(x = input_135, y = ret_out_7)[name = tensor<string, []>("x_21")];
             tensor<int32, [4]> window_19_begin_0 = const()[name = tensor<string, []>("window_19_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
             tensor<int32, [4]> window_19_end_0 = const()[name = tensor<string, []>("window_19_end_0"), val = tensor<int32, [4]>([4, 1, 16, 256])];
             tensor<bool, [4]> window_19_end_mask_0 = const()[name = tensor<string, []>("window_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_19_squeeze_mask_0 = const()[name = tensor<string, []>("window_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_19")];
-            tensor<int32, [3]> var_815_begin_0 = const()[name = tensor<string, []>("op_815_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_815_end_0 = const()[name = tensor<string, []>("op_815_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_815_end_mask_0 = const()[name = tensor<string, []>("op_815_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor<string, []>("op_815")];
-            tensor<int32, [3]> var_818_begin_0 = const()[name = tensor<string, []>("op_818_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_818_end_0 = const()[name = tensor<string, []>("op_818_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_818_end_mask_0 = const()[name = tensor<string, []>("op_818_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor<string, []>("op_818")];
+            tensor<int32, [3]> var_851_begin_0 = const()[name = tensor<string, []>("op_851_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_851_end_0 = const()[name = tensor<string, []>("op_851_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_851_end_mask_0 = const()[name = tensor<string, []>("op_851_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor<string, []>("op_851")];
+            tensor<int32, [3]> var_854_begin_0 = const()[name = tensor<string, []>("op_854_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_854_end_0 = const()[name = tensor<string, []>("op_854_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_854_end_mask_0 = const()[name = tensor<string, []>("op_854_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor<string, []>("op_854")];
             tensor<bool, []> window_21_interleave_0 = const()[name = tensor<string, []>("window_21_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor<string, []>("window_21")];
-            tensor<int32, [3]> var_823_begin_0 = const()[name = tensor<string, []>("op_823_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_823_end_0 = const()[name = tensor<string, []>("op_823_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_823_end_mask_0 = const()[name = tensor<string, []>("op_823_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor<string, []>("op_823")];
-            tensor<int32, [3]> var_826_begin_0 = const()[name = tensor<string, []>("op_826_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_826_end_0 = const()[name = tensor<string, []>("op_826_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_826_end_mask_0 = const()[name = tensor<string, []>("op_826_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor<string, []>("op_826")];
+            tensor<fp32, [1, 16, 256]> window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor<string, []>("window_21")];
+            tensor<int32, [3]> var_859_begin_0 = const()[name = tensor<string, []>("op_859_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_859_end_0 = const()[name = tensor<string, []>("op_859_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_859_end_mask_0 = const()[name = tensor<string, []>("op_859_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor<string, []>("op_859")];
+            tensor<int32, [3]> var_862_begin_0 = const()[name = tensor<string, []>("op_862_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_862_end_0 = const()[name = tensor<string, []>("op_862_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_862_end_mask_0 = const()[name = tensor<string, []>("op_862_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor<string, []>("op_862")];
             tensor<bool, []> window_interleave_0 = const()[name = tensor<string, []>("window_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor<string, []>("window")];
-            tensor<bool, []> input_141_interleave_0 = const()[name = tensor<string, []>("input_141_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [2, 16, 256]> input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 16, 256]> window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor<string, []>("window")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [2, 16, 256]> input_143 = concat(axis = var_48, interleave = input_143_interleave_0, values = (window_21, window))[name = tensor<string, []>("input_143")];
             tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 16, 256]> x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor<string, []>("x_23")];
-            tensor<int32, [3]> input_143_perm_0 = const()[name = tensor<string, []>("input_143_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [2, 16, 256]> x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor<string, []>("x_23")];
+            tensor<int32, [3]> input_145_perm_0 = const()[name = tensor<string, []>("input_145_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> inputs_33_pad_type_0 = const()[name = tensor<string, []>("inputs_33_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> inputs_33_strides_0 = const()[name = tensor<string, []>("inputs_33_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> inputs_33_pad_0 = const()[name = tensor<string, []>("inputs_33_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> inputs_33_dilations_0 = const()[name = tensor<string, []>("inputs_33_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> inputs_33_groups_0 = const()[name = tensor<string, []>("inputs_33_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
-            tensor<fp32, [2, 512, 16]> inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor<string, []>("inputs_33")];
-            tensor<int32, [2]> var_851_split_sizes_0 = const()[name = tensor<string, []>("op_851_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_851_axis_0 = const()[name = tensor<string, []>("op_851_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> var_851_0, tensor<fp32, [2, 256, 16]> var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_851")];
-            tensor<fp32, [2, 256, 16]> var_853 = sigmoid(x = var_851_1)[name = tensor<string, []>("op_853")];
-            tensor<fp32, [2, 256, 16]> inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor<string, []>("inputs_35")];
+            tensor<fp32, [2, 256, 16]> input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
+            tensor<fp32, [2, 512, 16]> inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor<string, []>("inputs_33")];
+            tensor<int32, [2]> var_887_split_sizes_0 = const()[name = tensor<string, []>("op_887_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_887_axis_0 = const()[name = tensor<string, []>("op_887_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> var_887_0, tensor<fp32, [2, 256, 16]> var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_887")];
+            tensor<fp32, [2, 256, 16]> var_889 = sigmoid(x = var_887_1)[name = tensor<string, []>("op_889")];
+            tensor<fp32, [2, 256, 16]> inputs_35 = mul(x = var_887_0, y = var_889)[name = tensor<string, []>("inputs_35")];
             tensor<string, []> outputs_aug_pad_type_0 = const()[name = tensor<string, []>("outputs_aug_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> outputs_aug_pad_0 = const()[name = tensor<string, []>("outputs_aug_pad_0"), val = tensor<int32, [2]>([15, 15])];
             tensor<int32, []> outputs_aug_groups_0 = const()[name = tensor<string, []>("outputs_aug_groups_0"), val = tensor<int32, []>(256)];
             tensor<int32, [1]> outputs_aug_strides_0 = const()[name = tensor<string, []>("outputs_aug_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [1]> outputs_aug_dilations_0 = const()[name = tensor<string, []>("outputs_aug_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [2, 256, 31]> outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor<string, []>("outputs_aug")];
-            tensor<int32, [3]> input_145_begin_0 = const()[name = tensor<string, []>("input_145_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> input_145_end_0 = const()[name = tensor<string, []>("input_145_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
-            tensor<bool, [3]> input_145_end_mask_0 = const()[name = tensor<string, []>("input_145_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
-            tensor<fp32, [2, 256, 16]> input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor<string, []>("input_145")];
-            tensor<fp32, [2, 256, 16]> inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor<string, []>("inputs_37")];
-            tensor<fp32, [2, 256, 16]> input_147 = silu(x = inputs_37)[name = tensor<string, []>("input_147")];
-            tensor<string, []> input_149_pad_type_0 = const()[name = tensor<string, []>("input_149_pad_type_0"), val = tensor<string, []>("valid")];
-            tensor<int32, [1]> input_149_strides_0 = const()[name = tensor<string, []>("input_149_strides_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [2]> input_149_pad_0 = const()[name = tensor<string, []>("input_149_pad_0"), val = tensor<int32, [2]>([0, 0])];
-            tensor<int32, [1]> input_149_dilations_0 = const()[name = tensor<string, []>("input_149_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, []> input_149_groups_0 = const()[name = tensor<string, []>("input_149_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [2, 256, 16]> input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor<string, []>("input_149")];
+            tensor<fp32, [2, 256, 31]> outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor<string, []>("outputs_aug")];
+            tensor<int32, [3]> input_147_begin_0 = const()[name = tensor<string, []>("input_147_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_147_end_0 = const()[name = tensor<string, []>("input_147_end_0"), val = tensor<int32, [3]>([2, 256, 16])];
+            tensor<bool, [3]> input_147_end_mask_0 = const()[name = tensor<string, []>("input_147_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp32, [2, 256, 16]> input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor<string, []>("input_147")];
+            tensor<fp32, [2, 256, 16]> inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor<string, []>("inputs_37")];
+            tensor<fp32, [2, 256, 16]> input_149 = silu(x = inputs_37)[name = tensor<string, []>("input_149")];
+            tensor<string, []> input_151_pad_type_0 = const()[name = tensor<string, []>("input_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_151_strides_0 = const()[name = tensor<string, []>("input_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_151_pad_0 = const()[name = tensor<string, []>("input_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_151_dilations_0 = const()[name = tensor<string, []>("input_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_151_groups_0 = const()[name = tensor<string, []>("input_151_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [2, 256, 16]> input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor<string, []>("input_151")];
             tensor<int32, [3]> conv_out_7_perm_0 = const()[name = tensor<string, []>("conv_out_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<int32, [3]> var_884_begin_0 = const()[name = tensor<string, []>("op_884_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_884_end_0 = const()[name = tensor<string, []>("op_884_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
-            tensor<bool, [3]> var_884_end_mask_0 = const()[name = tensor<string, []>("op_884_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [2, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor<string, []>("transpose_32")];
-            tensor<fp32, [2, 1, 256]> var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_884")];
-            tensor<int32, [3]> var_886_perm_0 = const()[name = tensor<string, []>("op_886_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 2, 256]> var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor<string, []>("transpose_31")];
-            tensor<fp32, [1, 2, 256]> input_151 = add(x = x_21, y = var_886)[name = tensor<string, []>("input_151")];
-            tensor<int32, [1]> input_153_axes_0 = const()[name = tensor<string, []>("input_153_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor<string, []>("input_153")];
-            tensor<fp32, [1, 2, 1024]> inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor<string, []>("linear_35")];
-            tensor<fp32, [1, 2, 1024]> input_155 = silu(x = inputs)[name = tensor<string, []>("input_155")];
-            tensor<fp32, [1, 2, 256]> input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor<string, []>("linear_36")];
-            tensor<fp32, []> var_909 = const()[name = tensor<string, []>("op_909"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 2, 256]> var_910 = mul(x = input_159, y = var_909)[name = tensor<string, []>("op_910")];
-            tensor<fp32, [1, 2, 256]> input_161 = add(x = var_910, y = input_151)[name = tensor<string, []>("input_161")];
+            tensor<int32, [3]> var_920_begin_0 = const()[name = tensor<string, []>("op_920_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_920_end_0 = const()[name = tensor<string, []>("op_920_end_0"), val = tensor<int32, [3]>([2, 16, 256])];
+            tensor<bool, [3]> var_920_end_mask_0 = const()[name = tensor<string, []>("op_920_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [2, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor<string, []>("transpose_32")];
+            tensor<fp32, [2, 1, 256]> var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_920")];
+            tensor<int32, [3]> var_922_perm_0 = const()[name = tensor<string, []>("op_922_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 2, 256]> var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor<string, []>("transpose_31")];
+            tensor<fp32, [1, 2, 256]> input_153 = add(x = x_21, y = var_922)[name = tensor<string, []>("input_153")];
+            tensor<int32, [1]> input_155_axes_0 = const()[name = tensor<string, []>("input_155_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor<string, []>("input_155")];
+            tensor<fp32, [1, 2, 1024]> inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor<string, []>("linear_35")];
+            tensor<fp32, [1, 2, 1024]> input_157 = silu(x = inputs)[name = tensor<string, []>("input_157")];
+            tensor<fp32, [1, 2, 256]> input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor<string, []>("linear_36")];
+            tensor<fp32, []> var_945 = const()[name = tensor<string, []>("op_945"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 256]> var_946 = mul(x = input_161, y = var_945)[name = tensor<string, []>("op_946")];
+            tensor<fp32, [1, 2, 256]> input_163 = add(x = var_946, y = input_153)[name = tensor<string, []>("input_163")];
             tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor<string, []>("x_25")];
+            tensor<fp32, [1, 2, 256]> x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor<string, []>("x_25")];
             tensor<int32, [3]> x_ct_perm_0 = const()[name = tensor<string, []>("x_ct_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<bool, []> cat_interleave_0 = const()[name = tensor<string, []>("cat_interleave_0"), val = tensor<bool, []>(false)];
             tensor<fp32, [1, 256, 2]> x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor<string, []>("transpose_30")];
-            tensor<fp32, [1, 256, 20]> cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor<string, []>("cat")];
+            tensor<fp32, [1, 256, 20]> cat = concat(axis = var_50, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor<string, []>("cat")];
             tensor<string, []> conv_out_pad_type_0 = const()[name = tensor<string, []>("conv_out_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> conv_out_strides_0 = const()[name = tensor<string, []>("conv_out_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> conv_out_pad_0 = const()[name = tensor<string, []>("conv_out_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> conv_out_dilations_0 = const()[name = tensor<string, []>("conv_out_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> conv_out_groups_0 = const()[name = tensor<string, []>("conv_out_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [1, 256, 2]> conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor<string, []>("conv_out")];
-            tensor<int32, [3]> var_928_begin_0 = const()[name = tensor<string, []>("op_928_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
-            tensor<int32, [3]> var_928_end_0 = const()[name = tensor<string, []>("op_928_end_0"), val = tensor<int32, [3]>([1, 256, 20])];
-            tensor<bool, [3]> var_928_end_mask_0 = const()[name = tensor<string, []>("op_928_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor<string, []>("op_928")];
-            tensor<int32, [3]> input_163_perm_0 = const()[name = tensor<string, []>("input_163_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<int32, [1]> var_930 = const()[name = tensor<string, []>("op_930"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 256]> input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
-            tensor<fp32, [1, 2, 1]> var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor<string, []>("op_931")];
+            tensor<fp32, [1, 256, 2]> conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor<string, []>("conv_out")];
+            tensor<int32, [3]> var_964_begin_0 = const()[name = tensor<string, []>("op_964_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_964_end_0 = const()[name = tensor<string, []>("op_964_end_0"), val = tensor<int32, [3]>([1, 256, 20])];
+            tensor<bool, [3]> var_964_end_mask_0 = const()[name = tensor<string, []>("op_964_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor<string, []>("op_964")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_966 = const()[name = tensor<string, []>("op_966"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 256]> input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
+            tensor<fp32, [1, 2, 1]> var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_44, x = input_165)[name = tensor<string, []>("op_967")];
             tensor<fp32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 2, 1]> clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor<string, []>("clip_0")];
-            tensor<fp32, [1, 2, 256]> emb = real_div(x = input_163, y = clip_0)[name = tensor<string, []>("emb")];
-            tensor<int32, []> var_935_axis_0 = const()[name = tensor<string, []>("op_935_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor<string, []>("op_935")];
-            tensor<int32, []> var_937_axis_0 = const()[name = tensor<string, []>("op_937_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_937")];
-            tensor<int32, []> var_939_axis_0 = const()[name = tensor<string, []>("op_939_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor<string, []>("op_939")];
-            tensor<fp32, []> var_948 = const()[name = tensor<string, []>("op_948"), val = tensor<fp32, []>(0x1.5798eep-27)];
-            tensor<fp32, []> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
-            tensor<fp32, []> var_955 = const()[name = tensor<string, []>("op_955"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
-            tensor<bool, []> var_956 = const()[name = tensor<string, []>("op_956"), val = tensor<bool, []>(true)];
-            tensor<fp32, []> var_958 = const()[name = tensor<string, []>("op_958"), val = tensor<fp32, []>(0x1p+0)];
-            tensor<int32, []> var_962 = const()[name = tensor<string, []>("op_962"), val = tensor<int32, []>(-1)];
-            tensor<int32, []> var_968 = const()[name = tensor<string, []>("op_968"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 2, 1]> clip_0 = clip(alpha = var_58, beta = const_12, x = var_967)[name = tensor<string, []>("clip_0")];
+            tensor<fp32, [1, 2, 256]> emb = real_div(x = input_165, y = clip_0)[name = tensor<string, []>("emb")];
+            tensor<int32, []> var_971_axis_0 = const()[name = tensor<string, []>("op_971_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor<string, []>("op_971")];
+            tensor<int32, []> var_973_axis_0 = const()[name = tensor<string, []>("op_973_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_973")];
+            tensor<int32, []> var_975_axis_0 = const()[name = tensor<string, []>("op_975_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor<string, []>("op_975")];
             tensor<fp32, [1, 2, 6, 256]> pos = const()[name = tensor<string, []>("pos"), val = tensor<fp32, [1, 2, 6, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44395328)))];
-            tensor<int32, [1]> var_1030_axes_0 = const()[name = tensor<string, []>("op_1030_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<fp32, [1, 2, 1, 256]> var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor<string, []>("op_1030")];
+            tensor<int32, [1]> var_1043_axes_0 = const()[name = tensor<string, []>("op_1043_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 2, 1, 256]> var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor<string, []>("op_1043")];
             tensor<int32, [4]> emb_exp_reps_0 = const()[name = tensor<string, []>("emb_exp_reps_0"), val = tensor<int32, [4]>([1, 1, 6, 1])];
-            tensor<fp32, [1, 2, 6, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor<string, []>("emb_exp")];
-            tensor<bool, []> input_165_interleave_0 = const()[name = tensor<string, []>("input_165_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 2, 6, 512]> input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_165")];
-            tensor<fp32, [1, 2, 6, 256]> x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor<string, []>("linear_37")];
-            tensor<int32, [4]> var_1038_perm_0 = const()[name = tensor<string, []>("op_1038_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1042 = const()[name = tensor<string, []>("op_1042"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [1, 6, 2, 256]> var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
-            tensor<fp32, [6, 2, 256]> x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor<string, []>("x_29")];
+            tensor<fp32, [1, 2, 6, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1043)[name = tensor<string, []>("emb_exp")];
+            tensor<bool, []> input_167_interleave_0 = const()[name = tensor<string, []>("input_167_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 2, 6, 512]> input_167 = concat(axis = var_51, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_167")];
+            tensor<fp32, [1, 2, 6, 256]> x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor<string, []>("linear_37")];
+            tensor<int32, [4]> var_1051_perm_0 = const()[name = tensor<string, []>("op_1051_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1055 = const()[name = tensor<string, []>("op_1055"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [1, 6, 2, 256]> var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
+            tensor<fp32, [6, 2, 256]> x_29 = reshape(shape = var_1055, x = var_1051)[name = tensor<string, []>("x_29")];
             tensor<int32, [5]> prev_kv_9_begin_0 = const()[name = tensor<string, []>("prev_kv_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_9_end_0 = const()[name = tensor<string, []>("prev_kv_9_end_0"), val = tensor<int32, [5]>([1, 6, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_9_end_mask_0 = const()[name = tensor<string, []>("prev_kv_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -903,132 +910,132 @@ program(1.0)
             tensor<bool, [2]> prev_scale_9_end_mask_0 = const()[name = tensor<string, []>("prev_scale_9_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_9_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_9_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor<string, []>("prev_scale_9")];
-            tensor<fp32, [6, 2, 256]> var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_38")];
-            tensor<int32, [4]> var_1051 = const()[name = tensor<string, []>("op_1051"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor<string, []>("op_1052")];
+            tensor<fp32, [6, 2, 256]> var_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_38")];
+            tensor<int32, [4]> var_1064 = const()[name = tensor<string, []>("op_1064"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor<string, []>("op_1065")];
             tensor<int32, [4]> q_9_perm_0 = const()[name = tensor<string, []>("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_39")];
-            tensor<fp32, []> var_1057 = const()[name = tensor<string, []>("op_1057"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [6, 2, 256]> var_1058 = mul(x = var_1056, y = var_1057)[name = tensor<string, []>("op_1058")];
-            tensor<int32, [4]> var_1059 = const()[name = tensor<string, []>("op_1059"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor<string, []>("op_1060")];
+            tensor<fp32, [6, 2, 256]> var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_39")];
+            tensor<fp32, []> var_1070 = const()[name = tensor<string, []>("op_1070"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [6, 2, 256]> var_1071 = mul(x = var_1069, y = var_1070)[name = tensor<string, []>("op_1071")];
+            tensor<int32, [4]> var_1072 = const()[name = tensor<string, []>("op_1072"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor<string, []>("op_1073")];
             tensor<int32, [4]> k_9_perm_0 = const()[name = tensor<string, []>("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_40")];
-            tensor<int32, [4]> var_1065 = const()[name = tensor<string, []>("op_1065"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor<string, []>("op_1066")];
+            tensor<fp32, [6, 2, 256]> var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_40")];
+            tensor<int32, [4]> var_1078 = const()[name = tensor<string, []>("op_1078"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor<string, []>("op_1079")];
             tensor<int32, [4]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_41")];
+            tensor<fp32, [6, 2, 256]> input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_41")];
             tensor<bool, []> cumsum_mask_1_exclusive_0 = const()[name = tensor<string, []>("cumsum_mask_1_exclusive_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> cumsum_mask_1_reverse_0 = const()[name = tensor<string, []>("cumsum_mask_1_reverse_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [2]> cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor<string, []>("cumsum_mask_1")];
+            tensor<fp32, [2]> cumsum_mask_1 = cumsum(axis = var_48, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor<string, []>("cumsum_mask_1")];
             tensor<fp32, [1]> sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor<string, []>("sqrt_s0_9")];
             tensor<fp32, [2]> s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor<string, []>("s_t_9")];
             tensor<fp32, []> const_20 = const()[name = tensor<string, []>("const_20"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [2]> clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
+            tensor<fp32, [2]> clip_1 = clip(alpha = var_38, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
             tensor<fp32, [2]> sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor<string, []>("sqrt_s_t_9")];
             tensor<bool, []> qk_9_transpose_x_1 = const()[name = tensor<string, []>("qk_9_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_9_transpose_y_1 = const()[name = tensor<string, []>("qk_9_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [6, 4, 2, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor<string, []>("transpose_26")];
-            tensor<fp32, [6, 4, 2, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor<string, []>("transpose_27")];
+            tensor<fp32, [6, 4, 2, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1073)[name = tensor<string, []>("transpose_26")];
+            tensor<fp32, [6, 4, 2, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1065)[name = tensor<string, []>("transpose_27")];
             tensor<fp32, [6, 4, 2, 2]> qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor<string, []>("qk_9")];
-            tensor<int32, [2]> var_1078 = const()[name = tensor<string, []>("op_1078"), val = tensor<int32, [2]>([1, 2])];
-            tensor<fp32, [1, 2]> var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor<string, []>("op_1079")];
-            tensor<fp32, [2, 2]> causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor<string, []>("causal_with_valid_1")];
-            tensor<int32, [2]> var_1081 = const()[name = tensor<string, []>("op_1081"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor<string, []>("op_1082")];
-            tensor<fp32, [2, 2]> M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor<string, []>("M_9")];
-            tensor<fp32, [6, 4, 2, 2]> var_1084 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1084")];
+            tensor<int32, [2]> var_1091 = const()[name = tensor<string, []>("op_1091"), val = tensor<int32, [2]>([1, 2])];
+            tensor<fp32, [1, 2]> var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor<string, []>("op_1092")];
+            tensor<fp32, [2, 2]> causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor<string, []>("causal_with_valid_1")];
+            tensor<int32, [2]> var_1094 = const()[name = tensor<string, []>("op_1094"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor<string, []>("op_1095")];
+            tensor<fp32, [2, 2]> M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor<string, []>("M_9")];
+            tensor<fp32, [6, 4, 2, 2]> var_1097 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1097")];
             tensor<bool, []> inner_9_transpose_x_0 = const()[name = tensor<string, []>("inner_9_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_9_transpose_y_0 = const()[name = tensor<string, []>("inner_9_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 2, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor<string, []>("transpose_25")];
-            tensor<fp32, [6, 4, 2, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor<string, []>("inner_9")];
-            tensor<bool, []> var_1086_transpose_x_0 = const()[name = tensor<string, []>("op_1086_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1086_transpose_y_0 = const()[name = tensor<string, []>("op_1086_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 2, 64]> var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1086")];
-            tensor<fp32, [2]> var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1087")];
-            tensor<int32, [4]> var_1088 = const()[name = tensor<string, []>("op_1088"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor<string, []>("op_1089")];
-            tensor<fp32, [6, 4, 2, 64]> cross_9 = mul(x = var_1086, y = var_1089)[name = tensor<string, []>("cross_9")];
+            tensor<fp32, [6, 4, 2, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1079)[name = tensor<string, []>("transpose_25")];
+            tensor<fp32, [6, 4, 2, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor<string, []>("inner_9")];
+            tensor<bool, []> var_1099_transpose_x_0 = const()[name = tensor<string, []>("op_1099_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1099_transpose_y_0 = const()[name = tensor<string, []>("op_1099_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 2, 64]> var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1099")];
+            tensor<fp32, [2]> var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1100")];
+            tensor<int32, [4]> var_1101 = const()[name = tensor<string, []>("op_1101"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor<string, []>("op_1102")];
+            tensor<fp32, [6, 4, 2, 64]> cross_9 = mul(x = var_1099, y = var_1102)[name = tensor<string, []>("cross_9")];
             tensor<fp32, [6, 4, 2, 64]> out_25 = add(x = inner_9, y = cross_9)[name = tensor<string, []>("out_25")];
-            tensor<int32, [4]> var_1092 = const()[name = tensor<string, []>("op_1092"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor<string, []>("op_1093")];
-            tensor<fp32, [6, 4, 2, 64]> v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor<string, []>("v_masked_1")];
-            tensor<fp32, [6, 4, 64, 64]> var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1095")];
-            tensor<bool, []> var_1097_transpose_x_1 = const()[name = tensor<string, []>("op_1097_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1097_transpose_y_1 = const()[name = tensor<string, []>("op_1097_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 64, 64]> var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1097")];
-            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor<string, []>("new_kv_unnorm_9")];
-            tensor<bool, []> var_1099_keep_dims_0 = const()[name = tensor<string, []>("op_1099_keep_dims_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, []> var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1099")];
-            tensor<int32, [1]> var_1100 = const()[name = tensor<string, []>("op_1100"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [1]> var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor<string, []>("op_1101")];
-            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor<string, []>("new_scale_9")];
+            tensor<int32, [4]> var_1105 = const()[name = tensor<string, []>("op_1105"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor<string, []>("op_1106")];
+            tensor<fp32, [6, 4, 2, 64]> v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor<string, []>("v_masked_1")];
+            tensor<fp32, [6, 4, 64, 64]> var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1108")];
+            tensor<bool, []> var_1110_transpose_x_1 = const()[name = tensor<string, []>("op_1110_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1110_transpose_y_1 = const()[name = tensor<string, []>("op_1110_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 64, 64]> var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1110")];
+            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor<string, []>("new_kv_unnorm_9")];
+            tensor<bool, []> var_1112_keep_dims_0 = const()[name = tensor<string, []>("op_1112_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, []> var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1112")];
+            tensor<int32, [1]> var_1113 = const()[name = tensor<string, []>("op_1113"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1]> var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor<string, []>("op_1114")];
+            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor<string, []>("new_scale_9")];
             tensor<fp32, []> const_21 = const()[name = tensor<string, []>("const_21"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1]> clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor<string, []>("clip_2")];
+            tensor<fp32, [1]> clip_2 = clip(alpha = var_38, beta = const_21, x = new_scale_9)[name = tensor<string, []>("clip_2")];
             tensor<fp32, [1]> sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor<string, []>("sqrt_new_scale_1")];
-            tensor<fp32, [6, 4, 64, 64]> var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1105")];
-            tensor<int32, [4]> var_1106_perm_0 = const()[name = tensor<string, []>("op_1106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [6, 4, 64, 64]> var_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1118")];
+            tensor<int32, [4]> var_1119_perm_0 = const()[name = tensor<string, []>("op_1119_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 2, 4, 64]> var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
-            tensor<fp32, [6, 2, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor<string, []>("out_27")];
-            tensor<int32, [3]> var_1110 = const()[name = tensor<string, []>("op_1110"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [6, 2, 256]> out_29 = reshape(shape = var_1110, x = out_27)[name = tensor<string, []>("out_29")];
-            tensor<fp32, [6, 2, 256]> var_1112 = silu(x = input_169)[name = tensor<string, []>("op_1112")];
-            tensor<fp32, [6, 2, 256]> input_171 = mul(x = var_1112, y = out_29)[name = tensor<string, []>("input_171")];
-            tensor<fp32, [6, 2, 256]> ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor<string, []>("linear_42")];
-            tensor<fp32, [6, 2, 256]> input_173 = add(x = x_29, y = ret_out_9)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [6, 2, 4, 64]> var_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
+            tensor<fp32, [6, 2, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_53, x = var_1119)[name = tensor<string, []>("out_27")];
+            tensor<int32, [3]> var_1123 = const()[name = tensor<string, []>("op_1123"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [6, 2, 256]> out_29 = reshape(shape = var_1123, x = out_27)[name = tensor<string, []>("out_29")];
+            tensor<fp32, [6, 2, 256]> var_1125 = silu(x = input_171)[name = tensor<string, []>("op_1125")];
+            tensor<fp32, [6, 2, 256]> input_173 = mul(x = var_1125, y = out_29)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [6, 2, 256]> ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor<string, []>("linear_42")];
+            tensor<fp32, [6, 2, 256]> input_175 = add(x = x_29, y = ret_out_9)[name = tensor<string, []>("input_175")];
             tensor<int32, [1]> xt_1_axes_0 = const()[name = tensor<string, []>("xt_1_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 2, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor<string, []>("xt_1")];
-            tensor<int32, [4]> var_1122 = const()[name = tensor<string, []>("op_1122"), val = tensor<int32, [4]>([1, 6, 2, 256])];
-            tensor<fp32, [1, 6, 2, 256]> var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor<string, []>("op_1123")];
-            tensor<int32, [4]> var_1124_perm_0 = const()[name = tensor<string, []>("op_1124_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [3]>([2, 6, 256])];
-            tensor<fp32, [1, 2, 6, 256]> var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor<string, []>("transpose_23")];
-            tensor<fp32, [2, 6, 256]> query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor<string, []>("query_1")];
+            tensor<fp32, [6, 2, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_45, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor<string, []>("xt_1")];
+            tensor<int32, [4]> var_1135 = const()[name = tensor<string, []>("op_1135"), val = tensor<int32, [4]>([1, 6, 2, 256])];
+            tensor<fp32, [1, 6, 2, 256]> var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor<string, []>("op_1136")];
+            tensor<int32, [4]> var_1137_perm_0 = const()[name = tensor<string, []>("op_1137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1140 = const()[name = tensor<string, []>("op_1140"), val = tensor<int32, [3]>([2, 6, 256])];
+            tensor<fp32, [1, 2, 6, 256]> var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor<string, []>("transpose_23")];
+            tensor<fp32, [2, 6, 256]> query_1 = reshape(shape = var_1140, x = var_1137)[name = tensor<string, []>("query_1")];
             tensor<int32, [3]> query_3_perm_0 = const()[name = tensor<string, []>("query_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [6, 2, 256]> query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor<string, []>("transpose_22")];
-            tensor<fp32, [6, 2, 768]> var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor<string, []>("linear_43")];
+            tensor<fp32, [6, 2, 768]> var_1163 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor<string, []>("linear_43")];
             tensor<int32, [4]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<int32, [4]>([6, 2, 3, 256])];
-            tensor<fp32, [6, 2, 3, 256]> var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor<string, []>("op_1152")];
-            tensor<int32, [1]> var_1153_axes_0 = const()[name = tensor<string, []>("op_1153_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 6, 2, 3, 256]> var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor<string, []>("op_1153")];
-            tensor<int32, [5]> var_1154_perm_0 = const()[name = tensor<string, []>("op_1154_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1155_axes_0 = const()[name = tensor<string, []>("op_1155_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 6, 2, 1, 256]> var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor<string, []>("transpose_21")];
-            tensor<fp32, [3, 6, 2, 256]> var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor<string, []>("op_1155")];
+            tensor<fp32, [6, 2, 3, 256]> var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor<string, []>("op_1165")];
+            tensor<int32, [1]> var_1166_axes_0 = const()[name = tensor<string, []>("op_1166_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 6, 2, 3, 256]> var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor<string, []>("op_1166")];
+            tensor<int32, [5]> var_1167_perm_0 = const()[name = tensor<string, []>("op_1167_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1168_axes_0 = const()[name = tensor<string, []>("op_1168_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 6, 2, 1, 256]> var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor<string, []>("transpose_21")];
+            tensor<fp32, [3, 6, 2, 256]> var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor<string, []>("op_1168")];
             tensor<int32, [4]> q_11_begin_0 = const()[name = tensor<string, []>("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [4]> q_11_end_0 = const()[name = tensor<string, []>("q_11_end_0"), val = tensor<int32, [4]>([1, 6, 2, 256])];
             tensor<bool, [4]> q_11_end_mask_0 = const()[name = tensor<string, []>("q_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> q_11_squeeze_mask_0 = const()[name = tensor<string, []>("q_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor<string, []>("q_11")];
+            tensor<fp32, [6, 2, 256]> q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1168)[name = tensor<string, []>("q_11")];
             tensor<int32, [4]> k_11_begin_0 = const()[name = tensor<string, []>("k_11_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> k_11_end_0 = const()[name = tensor<string, []>("k_11_end_0"), val = tensor<int32, [4]>([2, 6, 2, 256])];
             tensor<bool, [4]> k_11_end_mask_0 = const()[name = tensor<string, []>("k_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> k_11_squeeze_mask_0 = const()[name = tensor<string, []>("k_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor<string, []>("k_11")];
+            tensor<fp32, [6, 2, 256]> k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1168)[name = tensor<string, []>("k_11")];
             tensor<int32, [4]> v_11_begin_0 = const()[name = tensor<string, []>("v_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> v_11_end_0 = const()[name = tensor<string, []>("v_11_end_0"), val = tensor<int32, [4]>([3, 6, 2, 256])];
             tensor<bool, [4]> v_11_end_mask_0 = const()[name = tensor<string, []>("v_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> v_11_squeeze_mask_0 = const()[name = tensor<string, []>("v_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor<string, []>("v_11")];
-            tensor<int32, [3]> var_1163 = const()[name = tensor<string, []>("op_1163"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor<string, []>("op_1164")];
+            tensor<fp32, [6, 2, 256]> v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1168)[name = tensor<string, []>("v_11")];
+            tensor<int32, [3]> var_1176 = const()[name = tensor<string, []>("op_1176"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor<string, []>("op_1177")];
             tensor<int32, [3]> q_13_perm_0 = const()[name = tensor<string, []>("q_13_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [3]> var_1170 = const()[name = tensor<string, []>("op_1170"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor<string, []>("op_1171")];
+            tensor<int32, [3]> var_1183 = const()[name = tensor<string, []>("op_1183"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor<string, []>("op_1184")];
             tensor<int32, [3]> k_13_perm_0 = const()[name = tensor<string, []>("k_13_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [3]> var_1177 = const()[name = tensor<string, []>("op_1177"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor<string, []>("op_1178")];
+            tensor<int32, [3]> var_1190 = const()[name = tensor<string, []>("op_1190"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor<string, []>("op_1191")];
             tensor<int32, [3]> v_13_perm_0 = const()[name = tensor<string, []>("v_13_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [4]> var_1181 = const()[name = tensor<string, []>("op_1181"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor<string, []>("transpose_20")];
-            tensor<fp32, [2, 4, 6, 64]> q_15 = reshape(shape = var_1181, x = q_13)[name = tensor<string, []>("q_15")];
-            tensor<int32, [4]> var_1183 = const()[name = tensor<string, []>("op_1183"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor<string, []>("transpose_19")];
-            tensor<fp32, [2, 4, 6, 64]> k_15 = reshape(shape = var_1183, x = k_13)[name = tensor<string, []>("k_15")];
-            tensor<int32, [4]> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor<string, []>("transpose_18")];
-            tensor<fp32, [2, 4, 6, 64]> v_15 = reshape(shape = var_1185, x = v_13)[name = tensor<string, []>("v_15")];
+            tensor<int32, [4]> var_1194 = const()[name = tensor<string, []>("op_1194"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor<string, []>("transpose_20")];
+            tensor<fp32, [2, 4, 6, 64]> q_15 = reshape(shape = var_1194, x = q_13)[name = tensor<string, []>("q_15")];
+            tensor<int32, [4]> var_1196 = const()[name = tensor<string, []>("op_1196"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor<string, []>("transpose_19")];
+            tensor<fp32, [2, 4, 6, 64]> k_15 = reshape(shape = var_1196, x = k_13)[name = tensor<string, []>("k_15")];
+            tensor<int32, [4]> var_1198 = const()[name = tensor<string, []>("op_1198"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor<string, []>("transpose_18")];
+            tensor<fp32, [2, 4, 6, 64]> v_15 = reshape(shape = var_1198, x = v_13)[name = tensor<string, []>("v_15")];
             tensor<fp32, []> mul_1_y_0 = const()[name = tensor<string, []>("mul_1_y_0"), val = tensor<fp32, []>(0x1p-3)];
             tensor<fp32, [2, 4, 6, 64]> mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor<string, []>("mul_1")];
             tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
@@ -1039,30 +1046,30 @@ program(1.0)
             tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
             tensor<fp32, [2, 4, 6, 64]> attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor<string, []>("attn_output_1")];
-            tensor<int32, [4]> var_1188 = const()[name = tensor<string, []>("op_1188"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1193 = const()[name = tensor<string, []>("op_1193"), val = tensor<int32, [2]>([12, 256])];
-            tensor<fp32, [6, 2, 4, 64]> var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
-            tensor<fp32, [12, 256]> attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor<string, []>("attn_output_3")];
-            tensor<fp32, [12, 256]> attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor<string, []>("linear_44")];
-            tensor<int32, [3]> var_1197 = const()[name = tensor<string, []>("op_1197"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [6, 2, 256]> attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor<string, []>("attn_output_7")];
+            tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1206 = const()[name = tensor<string, []>("op_1206"), val = tensor<int32, [2]>([12, 256])];
+            tensor<fp32, [6, 2, 4, 64]> var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
+            tensor<fp32, [12, 256]> attn_output_3 = reshape(shape = var_1206, x = var_1202)[name = tensor<string, []>("attn_output_3")];
+            tensor<fp32, [12, 256]> attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor<string, []>("linear_44")];
+            tensor<int32, [3]> var_1210 = const()[name = tensor<string, []>("op_1210"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [6, 2, 256]> attn_output_7 = reshape(shape = var_1210, x = attn_output_5)[name = tensor<string, []>("attn_output_7")];
             tensor<int32, [3]> sa2_out_1_perm_0 = const()[name = tensor<string, []>("sa2_out_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [2, 6, 256]> sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor<string, []>("transpose_16")];
-            tensor<fp32, [2, 6, 256]> input_175 = add(x = query_1, y = sa2_out_1)[name = tensor<string, []>("input_175")];
-            tensor<int32, [1]> input_177_axes_0 = const()[name = tensor<string, []>("input_177_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 6, 256]> input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor<string, []>("input_177")];
-            tensor<fp32, [2, 6, 2048]> input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor<string, []>("linear_45")];
-            tensor<fp32, [2, 6, 2048]> input_181 = relu(x = input_179)[name = tensor<string, []>("input_181")];
-            tensor<fp32, [2, 6, 256]> ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor<string, []>("linear_46")];
-            tensor<fp32, [2, 6, 256]> input_183 = add(x = input_177, y = ffn_out_1)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [2, 6, 256]> input_177 = add(x = query_1, y = sa2_out_1)[name = tensor<string, []>("input_177")];
+            tensor<int32, [1]> input_179_axes_0 = const()[name = tensor<string, []>("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [2, 6, 256]> input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_45, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor<string, []>("input_179")];
+            tensor<fp32, [2, 6, 2048]> input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor<string, []>("linear_45")];
+            tensor<fp32, [2, 6, 2048]> input_183 = relu(x = input_181)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [2, 6, 256]> ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor<string, []>("linear_46")];
+            tensor<fp32, [2, 6, 256]> input_185 = add(x = input_179, y = ffn_out_1)[name = tensor<string, []>("input_185")];
             tensor<int32, [1]> xt_3_axes_0 = const()[name = tensor<string, []>("xt_3_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 6, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor<string, []>("xt_3")];
-            tensor<int32, [4]> var_1217 = const()[name = tensor<string, []>("op_1217"), val = tensor<int32, [4]>([1, 2, 6, 256])];
-            tensor<fp32, [1, 2, 6, 256]> x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor<string, []>("x_31")];
-            tensor<int32, [4]> var_1219_perm_0 = const()[name = tensor<string, []>("op_1219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1223 = const()[name = tensor<string, []>("op_1223"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [1, 6, 2, 256]> var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
-            tensor<fp32, [6, 2, 256]> x = reshape(shape = var_1223, x = var_1219)[name = tensor<string, []>("x")];
+            tensor<fp32, [2, 6, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_45, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor<string, []>("xt_3")];
+            tensor<int32, [4]> var_1230 = const()[name = tensor<string, []>("op_1230"), val = tensor<int32, [4]>([1, 2, 6, 256])];
+            tensor<fp32, [1, 2, 6, 256]> x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor<string, []>("x_31")];
+            tensor<int32, [4]> var_1232_perm_0 = const()[name = tensor<string, []>("op_1232_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1236 = const()[name = tensor<string, []>("op_1236"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [1, 6, 2, 256]> var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
+            tensor<fp32, [6, 2, 256]> x = reshape(shape = var_1236, x = var_1232)[name = tensor<string, []>("x")];
             tensor<int32, [5]> prev_kv_begin_0 = const()[name = tensor<string, []>("prev_kv_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_end_0 = const()[name = tensor<string, []>("prev_kv_end_0"), val = tensor<int32, [5]>([2, 6, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_end_mask_0 = const()[name = tensor<string, []>("prev_kv_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -1073,120 +1080,120 @@ program(1.0)
             tensor<bool, [2]> prev_scale_end_mask_0 = const()[name = tensor<string, []>("prev_scale_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor<string, []>("prev_scale")];
-            tensor<fp32, [6, 2, 256]> var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor<string, []>("linear_47")];
-            tensor<int32, [4]> var_1232 = const()[name = tensor<string, []>("op_1232"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor<string, []>("op_1233")];
+            tensor<fp32, [6, 2, 256]> var_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor<string, []>("linear_47")];
+            tensor<int32, [4]> var_1245 = const()[name = tensor<string, []>("op_1245"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor<string, []>("op_1246")];
             tensor<int32, [4]> q_17_perm_0 = const()[name = tensor<string, []>("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor<string, []>("linear_48")];
-            tensor<fp32, []> var_1238 = const()[name = tensor<string, []>("op_1238"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [6, 2, 256]> var_1239 = mul(x = var_1237, y = var_1238)[name = tensor<string, []>("op_1239")];
-            tensor<int32, [4]> var_1240 = const()[name = tensor<string, []>("op_1240"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor<string, []>("op_1241")];
+            tensor<fp32, [6, 2, 256]> var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor<string, []>("linear_48")];
+            tensor<fp32, []> var_1251 = const()[name = tensor<string, []>("op_1251"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [6, 2, 256]> var_1252 = mul(x = var_1250, y = var_1251)[name = tensor<string, []>("op_1252")];
+            tensor<int32, [4]> var_1253 = const()[name = tensor<string, []>("op_1253"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor<string, []>("op_1254")];
             tensor<int32, [4]> k_17_perm_0 = const()[name = tensor<string, []>("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor<string, []>("linear_49")];
-            tensor<int32, [4]> var_1246 = const()[name = tensor<string, []>("op_1246"), val = tensor<int32, [4]>([6, 2, 4, 64])];
-            tensor<fp32, [6, 2, 4, 64]> var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor<string, []>("op_1247")];
+            tensor<fp32, [6, 2, 256]> var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor<string, []>("linear_49")];
+            tensor<int32, [4]> var_1259 = const()[name = tensor<string, []>("op_1259"), val = tensor<int32, [4]>([6, 2, 4, 64])];
+            tensor<fp32, [6, 2, 4, 64]> var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor<string, []>("op_1260")];
             tensor<int32, [4]> v_17_perm_0 = const()[name = tensor<string, []>("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 2, 256]> input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor<string, []>("linear_50")];
+            tensor<fp32, [6, 2, 256]> input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor<string, []>("linear_50")];
             tensor<fp32, [1]> sqrt_s0 = sqrt(x = prev_scale)[name = tensor<string, []>("sqrt_s0")];
             tensor<fp32, [2]> s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor<string, []>("s_t")];
             tensor<fp32, []> const_32 = const()[name = tensor<string, []>("const_32"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [2]> clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
+            tensor<fp32, [2]> clip_3 = clip(alpha = var_38, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
             tensor<fp32, [2]> sqrt_s_t = sqrt(x = clip_3)[name = tensor<string, []>("sqrt_s_t")];
             tensor<bool, []> qk_transpose_x_1 = const()[name = tensor<string, []>("qk_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_transpose_y_1 = const()[name = tensor<string, []>("qk_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [6, 4, 2, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor<string, []>("transpose_13")];
-            tensor<fp32, [6, 4, 2, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor<string, []>("transpose_14")];
+            tensor<fp32, [6, 4, 2, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1254)[name = tensor<string, []>("transpose_13")];
+            tensor<fp32, [6, 4, 2, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1246)[name = tensor<string, []>("transpose_14")];
             tensor<fp32, [6, 4, 2, 2]> qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor<string, []>("qk")];
-            tensor<int32, [2]> var_1262 = const()[name = tensor<string, []>("op_1262"), val = tensor<int32, [2]>([2, 1])];
-            tensor<fp32, [2, 1]> var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor<string, []>("op_1263")];
-            tensor<fp32, [2, 2]> M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor<string, []>("M")];
-            tensor<fp32, [6, 4, 2, 2]> var_1265 = mul(x = qk, y = M)[name = tensor<string, []>("op_1265")];
-            tensor<bool, []> inner_transpose_x_0 = const()[name = tensor<string, []>("inner_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> inner_transpose_y_0 = const()[name = tensor<string, []>("inner_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 2, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor<string, []>("transpose_12")];
-            tensor<fp32, [6, 4, 2, 64]> inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor<string, []>("inner")];
-            tensor<bool, []> var_1267_transpose_x_0 = const()[name = tensor<string, []>("op_1267_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1267_transpose_y_0 = const()[name = tensor<string, []>("op_1267_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 2, 64]> var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1267")];
-            tensor<fp32, [2]> var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1268")];
-            tensor<int32, [4]> var_1269 = const()[name = tensor<string, []>("op_1269"), val = tensor<int32, [4]>([1, 1, 2, 1])];
-            tensor<fp32, [1, 1, 2, 1]> var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor<string, []>("op_1270")];
-            tensor<fp32, [6, 4, 2, 64]> cross = mul(x = var_1267, y = var_1270)[name = tensor<string, []>("cross")];
-            tensor<fp32, [6, 4, 2, 64]> out_31 = add(x = inner, y = cross)[name = tensor<string, []>("out_31")];
-            tensor<fp32, [6, 4, 2, 64]> v_masked = mul(x = v_17, y = var_1093)[name = tensor<string, []>("v_masked")];
-            tensor<fp32, [6, 4, 64, 64]> var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1276")];
-            tensor<bool, []> var_1278_transpose_x_1 = const()[name = tensor<string, []>("op_1278_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1278_transpose_y_1 = const()[name = tensor<string, []>("op_1278_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 64, 64]> var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1278")];
-            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor<string, []>("new_kv_unnorm")];
-            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1101)[name = tensor<string, []>("new_scale")];
+            tensor<int32, [2]> var_1275 = const()[name = tensor<string, []>("op_1275"), val = tensor<int32, [2]>([2, 1])];
+            tensor<fp32, [2, 1]> var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor<string, []>("op_1276")];
+            tensor<fp32, [2, 2]> M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor<string, []>("M")];
+            tensor<fp32, [6, 4, 2, 2]> var_1278 = mul(x = qk, y = M)[name = tensor<string, []>("op_1278")];
+            tensor<bool, []> inner_11_transpose_x_0 = const()[name = tensor<string, []>("inner_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> inner_11_transpose_y_0 = const()[name = tensor<string, []>("inner_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 2, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1260)[name = tensor<string, []>("transpose_12")];
+            tensor<fp32, [6, 4, 2, 64]> inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor<string, []>("inner_11")];
+            tensor<bool, []> var_1280_transpose_x_0 = const()[name = tensor<string, []>("op_1280_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1280_transpose_y_0 = const()[name = tensor<string, []>("op_1280_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 2, 64]> var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1280")];
+            tensor<fp32, [2]> var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1281")];
+            tensor<int32, [4]> var_1282 = const()[name = tensor<string, []>("op_1282"), val = tensor<int32, [4]>([1, 1, 2, 1])];
+            tensor<fp32, [1, 1, 2, 1]> var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor<string, []>("op_1283")];
+            tensor<fp32, [6, 4, 2, 64]> cross = mul(x = var_1280, y = var_1283)[name = tensor<string, []>("cross")];
+            tensor<fp32, [6, 4, 2, 64]> out_31 = add(x = inner_11, y = cross)[name = tensor<string, []>("out_31")];
+            tensor<fp32, [6, 4, 2, 64]> v_masked = mul(x = v_17, y = var_1106)[name = tensor<string, []>("v_masked")];
+            tensor<fp32, [6, 4, 64, 64]> var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1289")];
+            tensor<bool, []> var_1291_transpose_x_1 = const()[name = tensor<string, []>("op_1291_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1291_transpose_y_1 = const()[name = tensor<string, []>("op_1291_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 64, 64]> var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1291")];
+            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor<string, []>("new_kv_unnorm")];
+            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1114)[name = tensor<string, []>("new_scale")];
             tensor<fp32, []> const_33 = const()[name = tensor<string, []>("const_33"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1]> clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor<string, []>("clip_4")];
+            tensor<fp32, [1]> clip_4 = clip(alpha = var_38, beta = const_33, x = new_scale)[name = tensor<string, []>("clip_4")];
             tensor<fp32, [1]> sqrt_new_scale = sqrt(x = clip_4)[name = tensor<string, []>("sqrt_new_scale")];
             tensor<fp32, [6, 4, 64, 64]> nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor<string, []>("nkv")];
-            tensor<int32, [4]> var_1287_perm_0 = const()[name = tensor<string, []>("op_1287_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_1300_perm_0 = const()[name = tensor<string, []>("op_1300_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 2, 4, 64]> var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
-            tensor<fp32, [6, 2, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor<string, []>("out_33")];
-            tensor<int32, [3]> var_1291 = const()[name = tensor<string, []>("op_1291"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [6, 2, 256]> out = reshape(shape = var_1291, x = out_33)[name = tensor<string, []>("out")];
-            tensor<fp32, [6, 2, 256]> var_1293 = silu(x = input_187)[name = tensor<string, []>("op_1293")];
-            tensor<fp32, [6, 2, 256]> input_189 = mul(x = var_1293, y = out)[name = tensor<string, []>("input_189")];
-            tensor<fp32, [6, 2, 256]> ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor<string, []>("linear_51")];
-            tensor<fp32, [6, 2, 256]> input_191 = add(x = x, y = ret_out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [6, 2, 4, 64]> var_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
+            tensor<fp32, [6, 2, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_53, x = var_1300)[name = tensor<string, []>("out_33")];
+            tensor<int32, [3]> var_1304 = const()[name = tensor<string, []>("op_1304"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [6, 2, 256]> out = reshape(shape = var_1304, x = out_33)[name = tensor<string, []>("out")];
+            tensor<fp32, [6, 2, 256]> var_1306 = silu(x = input_189)[name = tensor<string, []>("op_1306")];
+            tensor<fp32, [6, 2, 256]> input_191 = mul(x = var_1306, y = out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [6, 2, 256]> ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor<string, []>("linear_51")];
+            tensor<fp32, [6, 2, 256]> input_193 = add(x = x, y = ret_out)[name = tensor<string, []>("input_193")];
             tensor<int32, [1]> xt_5_axes_0 = const()[name = tensor<string, []>("xt_5_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 2, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor<string, []>("xt_5")];
-            tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, 6, 2, 256])];
-            tensor<fp32, [1, 6, 2, 256]> var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor<string, []>("op_1304")];
-            tensor<int32, [4]> var_1305_perm_0 = const()[name = tensor<string, []>("op_1305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1308 = const()[name = tensor<string, []>("op_1308"), val = tensor<int32, [3]>([2, 6, 256])];
-            tensor<fp32, [1, 2, 6, 256]> var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor<string, []>("transpose_10")];
-            tensor<fp32, [2, 6, 256]> query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor<string, []>("query_5")];
+            tensor<fp32, [6, 2, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_45, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor<string, []>("xt_5")];
+            tensor<int32, [4]> var_1316 = const()[name = tensor<string, []>("op_1316"), val = tensor<int32, [4]>([1, 6, 2, 256])];
+            tensor<fp32, [1, 6, 2, 256]> var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor<string, []>("op_1317")];
+            tensor<int32, [4]> var_1318_perm_0 = const()[name = tensor<string, []>("op_1318_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1321 = const()[name = tensor<string, []>("op_1321"), val = tensor<int32, [3]>([2, 6, 256])];
+            tensor<fp32, [1, 2, 6, 256]> var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor<string, []>("transpose_10")];
+            tensor<fp32, [2, 6, 256]> query_5 = reshape(shape = var_1321, x = var_1318)[name = tensor<string, []>("query_5")];
             tensor<int32, [3]> query_perm_0 = const()[name = tensor<string, []>("query_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [6, 2, 256]> query = transpose(perm = query_perm_0, x = query_5)[name = tensor<string, []>("transpose_9")];
-            tensor<fp32, [6, 2, 768]> var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor<string, []>("linear_52")];
+            tensor<fp32, [6, 2, 768]> var_1344 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor<string, []>("linear_52")];
             tensor<int32, [4]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [4]>([6, 2, 3, 256])];
-            tensor<fp32, [6, 2, 3, 256]> var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor<string, []>("op_1333")];
-            tensor<int32, [1]> var_1334_axes_0 = const()[name = tensor<string, []>("op_1334_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 6, 2, 3, 256]> var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor<string, []>("op_1334")];
-            tensor<int32, [5]> var_1335_perm_0 = const()[name = tensor<string, []>("op_1335_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1336_axes_0 = const()[name = tensor<string, []>("op_1336_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 6, 2, 1, 256]> var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor<string, []>("transpose_8")];
-            tensor<fp32, [3, 6, 2, 256]> var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor<string, []>("op_1336")];
+            tensor<fp32, [6, 2, 3, 256]> var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor<string, []>("op_1346")];
+            tensor<int32, [1]> var_1347_axes_0 = const()[name = tensor<string, []>("op_1347_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 6, 2, 3, 256]> var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor<string, []>("op_1347")];
+            tensor<int32, [5]> var_1348_perm_0 = const()[name = tensor<string, []>("op_1348_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1349_axes_0 = const()[name = tensor<string, []>("op_1349_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 6, 2, 1, 256]> var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor<string, []>("transpose_8")];
+            tensor<fp32, [3, 6, 2, 256]> var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor<string, []>("op_1349")];
             tensor<int32, [4]> q_19_begin_0 = const()[name = tensor<string, []>("q_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [4]> q_19_end_0 = const()[name = tensor<string, []>("q_19_end_0"), val = tensor<int32, [4]>([1, 6, 2, 256])];
             tensor<bool, [4]> q_19_end_mask_0 = const()[name = tensor<string, []>("q_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> q_19_squeeze_mask_0 = const()[name = tensor<string, []>("q_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor<string, []>("q_19")];
+            tensor<fp32, [6, 2, 256]> q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1349)[name = tensor<string, []>("q_19")];
             tensor<int32, [4]> k_19_begin_0 = const()[name = tensor<string, []>("k_19_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> k_19_end_0 = const()[name = tensor<string, []>("k_19_end_0"), val = tensor<int32, [4]>([2, 6, 2, 256])];
             tensor<bool, [4]> k_19_end_mask_0 = const()[name = tensor<string, []>("k_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> k_19_squeeze_mask_0 = const()[name = tensor<string, []>("k_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor<string, []>("k_19")];
+            tensor<fp32, [6, 2, 256]> k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1349)[name = tensor<string, []>("k_19")];
             tensor<int32, [4]> v_19_begin_0 = const()[name = tensor<string, []>("v_19_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> v_19_end_0 = const()[name = tensor<string, []>("v_19_end_0"), val = tensor<int32, [4]>([3, 6, 2, 256])];
             tensor<bool, [4]> v_19_end_mask_0 = const()[name = tensor<string, []>("v_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> v_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 2, 256]> v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor<string, []>("v_19")];
-            tensor<int32, [3]> var_1344 = const()[name = tensor<string, []>("op_1344"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor<string, []>("op_1345")];
+            tensor<fp32, [6, 2, 256]> v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1349)[name = tensor<string, []>("v_19")];
+            tensor<int32, [3]> var_1357 = const()[name = tensor<string, []>("op_1357"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor<string, []>("op_1358")];
             tensor<int32, [3]> q_21_perm_0 = const()[name = tensor<string, []>("q_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [3]> var_1351 = const()[name = tensor<string, []>("op_1351"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor<string, []>("op_1352")];
+            tensor<int32, [3]> var_1364 = const()[name = tensor<string, []>("op_1364"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor<string, []>("op_1365")];
             tensor<int32, [3]> k_21_perm_0 = const()[name = tensor<string, []>("k_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [3]> var_1358 = const()[name = tensor<string, []>("op_1358"), val = tensor<int32, [3]>([6, 8, 64])];
-            tensor<fp32, [6, 8, 64]> var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor<string, []>("op_1359")];
+            tensor<int32, [3]> var_1371 = const()[name = tensor<string, []>("op_1371"), val = tensor<int32, [3]>([6, 8, 64])];
+            tensor<fp32, [6, 8, 64]> var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor<string, []>("op_1372")];
             tensor<int32, [3]> v_21_perm_0 = const()[name = tensor<string, []>("v_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [4]> var_1362 = const()[name = tensor<string, []>("op_1362"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor<string, []>("transpose_7")];
-            tensor<fp32, [2, 4, 6, 64]> q = reshape(shape = var_1362, x = q_21)[name = tensor<string, []>("q")];
-            tensor<int32, [4]> var_1364 = const()[name = tensor<string, []>("op_1364"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor<string, []>("transpose_6")];
-            tensor<fp32, [2, 4, 6, 64]> k = reshape(shape = var_1364, x = k_21)[name = tensor<string, []>("k")];
-            tensor<int32, [4]> var_1366 = const()[name = tensor<string, []>("op_1366"), val = tensor<int32, [4]>([2, 4, 6, 64])];
-            tensor<fp32, [8, 6, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor<string, []>("transpose_5")];
-            tensor<fp32, [2, 4, 6, 64]> v = reshape(shape = var_1366, x = v_21)[name = tensor<string, []>("v")];
+            tensor<int32, [4]> var_1375 = const()[name = tensor<string, []>("op_1375"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor<string, []>("transpose_7")];
+            tensor<fp32, [2, 4, 6, 64]> q = reshape(shape = var_1375, x = q_21)[name = tensor<string, []>("q")];
+            tensor<int32, [4]> var_1377 = const()[name = tensor<string, []>("op_1377"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor<string, []>("transpose_6")];
+            tensor<fp32, [2, 4, 6, 64]> k = reshape(shape = var_1377, x = k_21)[name = tensor<string, []>("k")];
+            tensor<int32, [4]> var_1379 = const()[name = tensor<string, []>("op_1379"), val = tensor<int32, [4]>([2, 4, 6, 64])];
+            tensor<fp32, [8, 6, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor<string, []>("transpose_5")];
+            tensor<fp32, [2, 4, 6, 64]> v = reshape(shape = var_1379, x = v_21)[name = tensor<string, []>("v")];
             tensor<fp32, []> mul_3_y_0 = const()[name = tensor<string, []>("mul_3_y_0"), val = tensor<fp32, []>(0x1p-3)];
             tensor<fp32, [2, 4, 6, 64]> mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor<string, []>("mul_3")];
             tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
@@ -1197,36 +1204,36 @@ program(1.0)
             tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)];
             tensor<fp32, [2, 4, 6, 64]> attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor<string, []>("attn_output_9")];
-            tensor<int32, [4]> var_1369 = const()[name = tensor<string, []>("op_1369"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1374 = const()[name = tensor<string, []>("op_1374"), val = tensor<int32, [2]>([12, 256])];
-            tensor<fp32, [6, 2, 4, 64]> var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
-            tensor<fp32, [12, 256]> attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor<string, []>("attn_output_11")];
-            tensor<fp32, [12, 256]> attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor<string, []>("linear_53")];
-            tensor<int32, [3]> var_1378 = const()[name = tensor<string, []>("op_1378"), val = tensor<int32, [3]>([6, 2, 256])];
-            tensor<fp32, [6, 2, 256]> attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor<string, []>("attn_output")];
+            tensor<int32, [4]> var_1382 = const()[name = tensor<string, []>("op_1382"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1387 = const()[name = tensor<string, []>("op_1387"), val = tensor<int32, [2]>([12, 256])];
+            tensor<fp32, [6, 2, 4, 64]> var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
+            tensor<fp32, [12, 256]> attn_output_11 = reshape(shape = var_1387, x = var_1383)[name = tensor<string, []>("attn_output_11")];
+            tensor<fp32, [12, 256]> attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor<string, []>("linear_53")];
+            tensor<int32, [3]> var_1391 = const()[name = tensor<string, []>("op_1391"), val = tensor<int32, [3]>([6, 2, 256])];
+            tensor<fp32, [6, 2, 256]> attn_output = reshape(shape = var_1391, x = attn_output_13)[name = tensor<string, []>("attn_output")];
             tensor<int32, [3]> sa2_out_perm_0 = const()[name = tensor<string, []>("sa2_out_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [2, 6, 256]> sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor<string, []>("transpose_3")];
-            tensor<fp32, [2, 6, 256]> input_193 = add(x = query_5, y = sa2_out)[name = tensor<string, []>("input_193")];
-            tensor<int32, [1]> input_195_axes_0 = const()[name = tensor<string, []>("input_195_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 6, 256]> input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor<string, []>("input_195")];
-            tensor<fp32, [2, 6, 2048]> input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor<string, []>("linear_54")];
-            tensor<fp32, [2, 6, 2048]> input_199 = relu(x = input_197)[name = tensor<string, []>("input_199")];
-            tensor<fp32, [2, 6, 256]> ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor<string, []>("linear_55")];
-            tensor<fp32, [2, 6, 256]> input_201 = add(x = input_195, y = ffn_out)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [2, 6, 256]> input_195 = add(x = query_5, y = sa2_out)[name = tensor<string, []>("input_195")];
+            tensor<int32, [1]> input_197_axes_0 = const()[name = tensor<string, []>("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [2, 6, 256]> input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_45, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor<string, []>("input_197")];
+            tensor<fp32, [2, 6, 2048]> input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor<string, []>("linear_54")];
+            tensor<fp32, [2, 6, 2048]> input_201 = relu(x = input_199)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [2, 6, 256]> ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor<string, []>("linear_55")];
+            tensor<fp32, [2, 6, 256]> input_203 = add(x = input_197, y = ffn_out)[name = tensor<string, []>("input_203")];
             tensor<int32, [1]> xt_axes_0 = const()[name = tensor<string, []>("xt_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [2, 6, 256]> xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor<string, []>("xt")];
-            tensor<int32, [4]> var_1398 = const()[name = tensor<string, []>("op_1398"), val = tensor<int32, [4]>([1, 2, 6, 256])];
-            tensor<fp32, [1, 2, 6, 256]> input = reshape(shape = var_1398, x = xt)[name = tensor<string, []>("input")];
-            tensor<int32, [1]> var_1400 = const()[name = tensor<string, []>("op_1400"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 2, 6, 1]> var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor<string, []>("op_1401")];
+            tensor<fp32, [2, 6, 256]> xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_45, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor<string, []>("xt")];
+            tensor<int32, [4]> var_1411 = const()[name = tensor<string, []>("op_1411"), val = tensor<int32, [4]>([1, 2, 6, 256])];
+            tensor<fp32, [1, 2, 6, 256]> input = reshape(shape = var_1411, x = xt)[name = tensor<string, []>("input")];
+            tensor<int32, [1]> var_1413 = const()[name = tensor<string, []>("op_1413"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 6, 1]> var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_44, x = input)[name = tensor<string, []>("op_1414")];
             tensor<fp32, []> const_42 = const()[name = tensor<string, []>("const_42"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 2, 6, 1]> clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor<string, []>("clip_5")];
-            tensor<fp32, [1, 2, 6, 256]> var_1403 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1403")];
+            tensor<fp32, [1, 2, 6, 1]> clip_5 = clip(alpha = var_58, beta = const_42, x = var_1414)[name = tensor<string, []>("clip_5")];
+            tensor<fp32, [1, 2, 6, 256]> var_1416 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1416")];
             tensor<int32, [3]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<int32, [3]>([2, 1, 256])];
             tensor<fp32, [2, 1, 256]> reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor<string, []>("reshape_0")];
             tensor<int32, [4]> transpose_1_perm_0 = const()[name = tensor<string, []>("transpose_1_perm_0"), val = tensor<int32, [4]>([0, 1, 3, 2])];
             tensor<int32, [3]> concat_7 = const()[name = tensor<string, []>("concat_7"), val = tensor<int32, [3]>([2, 256, 6])];
-            tensor<fp32, [1, 2, 256, 6]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 2, 256, 6]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[name = tensor<string, []>("transpose_2")];
             tensor<fp32, [2, 256, 6]> reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor<string, []>("reshape_1")];
             tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(false)];
@@ -1237,10 +1244,10 @@ program(1.0)
             tensor<int32, [3]> output_end_0 = const()[name = tensor<string, []>("output_end_0"), val = tensor<int32, [3]>([1, 2, 5])];
             tensor<bool, [3]> output_end_mask_0 = const()[name = tensor<string, []>("output_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
             tensor<fp32, [1, 2, 4]> output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor<string, []>("output")];
-            tensor<fp32, [1, 2, 4]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1407")];
-            tensor<int32, []> var_1409_axis_0 = const()[name = tensor<string, []>("op_1409_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 6, 4, 64, 64]> dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor<string, []>("op_1409")];
-            tensor<int32, []> var_1411_axis_0 = const()[name = tensor<string, []>("op_1411_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1411")];
+            tensor<fp32, [1, 2, 4]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1420")];
+            tensor<int32, []> var_1422_axis_0 = const()[name = tensor<string, []>("op_1422_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 6, 4, 64, 64]> dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor<string, []>("op_1422")];
+            tensor<int32, []> var_1424_axis_0 = const()[name = tensor<string, []>("op_1424_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1424")];
         } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new);
 }
\ No newline at end of file