diff --git "a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil" "b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil"
--- "a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil"
+++ "b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil"
@@ -1,234 +1,252 @@
 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, 9, 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, 3, 345]> features, tensor<fp32, [3]> 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, [3, 3]> encoder__causal_mask = const()[name = tensor<string, []>("encoder__causal_mask"), val = tensor<fp32, [3, 3]>([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])];
-            tensor<fp32, [3]> encoder__t_index = const()[name = tensor<string, []>("encoder__t_index"), val = tensor<fp32, [3]>([0x1p+0, 0x1p+1, 0x1.8p+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, [3, 3]> decoder__causal_mask = const()[name = tensor<string, []>("decoder__causal_mask"), val = tensor<fp32, [3, 3]>([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+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, 3, 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, 3, 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, 9, 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, 35, 23]> features, tensor<fp32, [3]> 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, [3, 3]> inner_encoder__causal_mask = const()[name = tensor<string, []>("inner_encoder__causal_mask"), val = tensor<fp32, [3, 3]>([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])];
+            tensor<fp32, [3]> inner_encoder__t_index = const()[name = tensor<string, []>("inner_encoder__t_index"), val = tensor<fp32, [3]>([0x1p+0, 0x1p+1, 0x1.8p+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, [3, 3]> inner_decoder__causal_mask = const()[name = tensor<string, []>("inner_decoder__causal_mask"), val = tensor<fp32, [3, 3]>([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+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, 25, 23])];
+            tensor<bool, [3]> var_29_end_mask_0 = const()[name = tensor<string, []>("op_29_end_mask_0"), val = tensor<bool, [3]>([true, false, 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, [3]> var_39_begin_0 = const()[name = tensor<string, []>("op_39_begin_0"), val = tensor<int32, [3]>([0, 20, 0])];
+            tensor<int32, [3]> var_39_end_0 = const()[name = tensor<string, []>("op_39_end_0"), val = tensor<int32, [3]>([1, 1, 23])];
+            tensor<bool, [3]> var_39_end_mask_0 = const()[name = tensor<string, []>("op_39_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 23]> var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor<string, []>("op_39")];
+            tensor<int32, []> stacked_axis_0 = const()[name = tensor<string, []>("stacked_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 3, 15, 23]> stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor<string, []>("stacked")];
+            tensor<int32, [3]> var_46 = const()[name = tensor<string, []>("op_46"), val = tensor<int32, [3]>([1, 3, 345])];
+            tensor<fp32, [1, 3, 345]> input_1 = reshape(shape = var_46, x = stacked)[name = tensor<string, []>("input_1")];
+            tensor<fp32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<bool, []> var_55 = const()[name = tensor<string, []>("op_55"), val = tensor<bool, []>(true)];
+            tensor<fp32, []> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_61 = const()[name = tensor<string, []>("op_61"), val = tensor<int32, []>(2)];
+            tensor<int32, []> var_62 = const()[name = tensor<string, []>("op_62"), val = tensor<int32, []>(-1)];
+            tensor<fp32, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, []> var_69 = const()[name = tensor<string, []>("op_69"), val = tensor<fp32, []>(0x1.5798eep-27)];
+            tensor<int32, []> var_72 = const()[name = tensor<string, []>("op_72"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 3, 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, 3, 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, 3, 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, 3, 1024]> input_7 = silu(x = inputs_1)[name = tensor<string, []>("input_7")];
-            tensor<fp32, [1, 3, 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, 3, 256]> var_148 = mul(x = input_11, y = var_147)[name = tensor<string, []>("op_148")];
-            tensor<fp32, [1, 3, 256]> input_13 = add(x = var_148, y = input_3)[name = tensor<string, []>("input_13")];
+            tensor<fp32, [1, 3, 256]> input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_56, 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, 3, 256]> input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_9 = silu(x = inputs_1)[name = tensor<string, []>("input_9")];
+            tensor<fp32, [1, 3, 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_193 = const()[name = tensor<string, []>("op_193"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_194 = mul(x = input_13, y = var_193)[name = tensor<string, []>("op_194")];
+            tensor<fp32, [1, 3, 256]> input_15 = add(x = var_194, 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, 3, 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, 3, 256]> x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_56, 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])];
@@ -239,163 +257,163 @@ 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, 3, 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, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_164 = reshape(shape = var_163, x = var_162)[name = tensor<string, []>("op_164")];
+            tensor<fp32, [1, 3, 256]> var_208 = 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_209 = const()[name = tensor<string, []>("op_209"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_210 = reshape(shape = var_209, x = var_208)[name = tensor<string, []>("op_210")];
             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, 3, 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, 3, 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, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_172 = reshape(shape = var_171, x = var_170)[name = tensor<string, []>("op_172")];
+            tensor<fp32, [1, 3, 256]> var_214 = 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_215 = const()[name = tensor<string, []>("op_215"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 3, 256]> var_216 = mul(x = var_214, y = var_215)[name = tensor<string, []>("op_216")];
+            tensor<int32, [4]> var_217 = const()[name = tensor<string, []>("op_217"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_218 = reshape(shape = var_217, x = var_216)[name = tensor<string, []>("op_218")];
             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, 3, 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, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_178 = reshape(shape = var_177, x = var_176)[name = tensor<string, []>("op_178")];
+            tensor<fp32, [1, 3, 256]> var_222 = 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_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_224 = reshape(shape = var_223, x = var_222)[name = tensor<string, []>("op_224")];
             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, 3, 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, 3, 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, [3]> s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor<string, []>("s_t_1")];
+            tensor<fp32, [3]> s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_1")];
             tensor<fp32, [3]> 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, 3, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor<string, []>("transpose_57")];
-            tensor<fp32, [1, 4, 3, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor<string, []>("transpose_58")];
+            tensor<fp32, [1, 4, 3, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor<string, []>("transpose_57")];
+            tensor<fp32, [1, 4, 3, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor<string, []>("transpose_58")];
             tensor<fp32, [1, 4, 3, 3]> 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]>([3, 1])];
-            tensor<fp32, [3, 1]> var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor<string, []>("op_189")];
-            tensor<fp32, [3, 3]> M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor<string, []>("M_1")];
-            tensor<fp32, [1, 4, 3, 3]> var_191 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_191")];
+            tensor<int32, [2]> var_234 = const()[name = tensor<string, []>("op_234"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor<string, []>("op_235")];
+            tensor<fp32, [3, 3]> M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor<string, []>("M_1")];
+            tensor<fp32, [1, 4, 3, 3]> var_237 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_237")];
             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, 3, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor<string, []>("transpose_56")];
-            tensor<fp32, [1, 4, 3, 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, 3, 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, [3]> 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, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_196 = reshape(shape = var_195, x = var_194)[name = tensor<string, []>("op_196")];
-            tensor<fp32, [1, 4, 3, 64]> cross_1 = mul(x = var_193, y = var_196)[name = tensor<string, []>("cross_1")];
+            tensor<fp32, [1, 4, 3, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor<string, []>("transpose_56")];
+            tensor<fp32, [1, 4, 3, 64]> inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor<string, []>("inner_1")];
+            tensor<bool, []> var_239_transpose_x_0 = const()[name = tensor<string, []>("op_239_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_239_transpose_y_0 = const()[name = tensor<string, []>("op_239_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 3, 64]> var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor<string, []>("op_239")];
+            tensor<fp32, [3]> var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor<string, []>("op_240")];
+            tensor<int32, [4]> var_241 = const()[name = tensor<string, []>("op_241"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_242 = reshape(shape = var_241, x = var_240)[name = tensor<string, []>("op_242")];
+            tensor<fp32, [1, 4, 3, 64]> cross_1 = mul(x = var_239, y = var_242)[name = tensor<string, []>("cross_1")];
             tensor<fp32, [1, 4, 3, 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, []>(0x1.8p+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_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor<string, []>("op_245")];
+            tensor<bool, []> var_247_transpose_x_1 = const()[name = tensor<string, []>("op_247_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_247_transpose_y_1 = const()[name = tensor<string, []>("op_247_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor<string, []>("op_247")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor<string, []>("new_kv_unnorm_1")];
+            tensor<fp32, []> var_249 = const()[name = tensor<string, []>("op_249"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor<string, []>("new_scale_1")];
+            tensor<fp32, [1]> var_251 = sqrt(x = new_scale_1)[name = tensor<string, []>("op_251")];
+            tensor<fp32, [1, 4, 64, 64]> var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor<string, []>("op_252")];
+            tensor<int32, [4]> var_253_perm_0 = const()[name = tensor<string, []>("op_253_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, 3, 4, 64]> var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
-            tensor<fp32, [1, 3, 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, 3, 256])];
-            tensor<fp32, [1, 3, 256]> out_5 = reshape(shape = var_211, x = out_3)[name = tensor<string, []>("out_5")];
-            tensor<fp32, [1, 3, 256]> var_213 = silu(x = input_17)[name = tensor<string, []>("op_213")];
-            tensor<fp32, [1, 3, 256]> input_19 = mul(x = var_213, y = out_5)[name = tensor<string, []>("input_19")];
-            tensor<fp32, [1, 3, 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, 3, 256]> x_3 = add(x = input_13, y = ret_out_1)[name = tensor<string, []>("x_3")];
+            tensor<fp32, [1, 3, 4, 64]> var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
+            tensor<fp32, [1, 3, 4, 64]> out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_64, x = var_253)[name = tensor<string, []>("out_3")];
+            tensor<int32, [3]> var_257 = const()[name = tensor<string, []>("op_257"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<fp32, [1, 3, 256]> out_5 = reshape(shape = var_257, x = out_3)[name = tensor<string, []>("out_5")];
+            tensor<fp32, [1, 3, 256]> var_259 = silu(x = input_19)[name = tensor<string, []>("op_259")];
+            tensor<fp32, [1, 3, 256]> input_21 = mul(x = var_259, y = out_5)[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 3, 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, 3, 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_267_begin_0 = const()[name = tensor<string, []>("op_267_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_267_end_0 = const()[name = tensor<string, []>("op_267_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_267_end_mask_0 = const()[name = tensor<string, []>("op_267_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor<string, []>("op_267")];
+            tensor<int32, [3]> var_270_begin_0 = const()[name = tensor<string, []>("op_270_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_270_end_0 = const()[name = tensor<string, []>("op_270_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_270_end_mask_0 = const()[name = tensor<string, []>("op_270_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor<string, []>("op_270")];
             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, 2, 256])];
-            tensor<bool, [3]> var_229_end_mask_0 = const()[name = tensor<string, []>("op_229_end_mask_0"), val = tensor<bool, [3]>([true, false, 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_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor<string, []>("window_3")];
+            tensor<int32, [3]> var_275_begin_0 = const()[name = tensor<string, []>("op_275_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_275_end_0 = const()[name = tensor<string, []>("op_275_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_275_end_mask_0 = const()[name = tensor<string, []>("op_275_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor<string, []>("op_275")];
+            tensor<int32, [3]> var_278_begin_0 = const()[name = tensor<string, []>("op_278_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_278_end_0 = const()[name = tensor<string, []>("op_278_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_278_end_mask_0 = const()[name = tensor<string, []>("op_278_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor<string, []>("op_278")];
             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<int32, [3]> var_237_begin_0 = const()[name = tensor<string, []>("op_237_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_237_end_0 = const()[name = tensor<string, []>("op_237_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_237_end_mask_0 = const()[name = tensor<string, []>("op_237_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor<string, []>("op_237")];
-            tensor<int32, [3]> var_240_begin_0 = const()[name = tensor<string, []>("op_240_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_240_end_0 = const()[name = tensor<string, []>("op_240_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_240_end_mask_0 = const()[name = tensor<string, []>("op_240_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor<string, []>("op_240")];
+            tensor<fp32, [1, 16, 256]> window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor<string, []>("window_5")];
+            tensor<int32, [3]> var_283_begin_0 = const()[name = tensor<string, []>("op_283_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_283_end_0 = const()[name = tensor<string, []>("op_283_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_283_end_mask_0 = const()[name = tensor<string, []>("op_283_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor<string, []>("op_283")];
+            tensor<int32, [3]> var_286_begin_0 = const()[name = tensor<string, []>("op_286_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_286_end_0 = const()[name = tensor<string, []>("op_286_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_286_end_mask_0 = const()[name = tensor<string, []>("op_286_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor<string, []>("op_286")];
             tensor<bool, []> window_7_interleave_0 = const()[name = tensor<string, []>("window_7_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor<string, []>("window_7")];
-            tensor<bool, []> input_21_interleave_0 = const()[name = tensor<string, []>("input_21_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [3, 16, 256]> input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 16, 256]> window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor<string, []>("window_7")];
+            tensor<bool, []> input_23_interleave_0 = const()[name = tensor<string, []>("input_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [3, 16, 256]> input_23 = concat(axis = var_59, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[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, [3, 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, [3, 16, 256]> x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_56, 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, [3, 256, 16]> input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
-            tensor<fp32, [3, 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_265_split_sizes_0 = const()[name = tensor<string, []>("op_265_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_265_axis_0 = const()[name = tensor<string, []>("op_265_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [3, 256, 16]> var_265_0, tensor<fp32, [3, 256, 16]> var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_265")];
-            tensor<fp32, [3, 256, 16]> var_267 = sigmoid(x = var_265_1)[name = tensor<string, []>("op_267")];
-            tensor<fp32, [3, 256, 16]> inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor<string, []>("inputs_5")];
+            tensor<fp32, [3, 256, 16]> input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
+            tensor<fp32, [3, 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_311_split_sizes_0 = const()[name = tensor<string, []>("op_311_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_311_axis_0 = const()[name = tensor<string, []>("op_311_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [3, 256, 16]> var_311_0, tensor<fp32, [3, 256, 16]> var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_311")];
+            tensor<fp32, [3, 256, 16]> var_313 = sigmoid(x = var_311_1)[name = tensor<string, []>("op_313")];
+            tensor<fp32, [3, 256, 16]> inputs_5 = mul(x = var_311_0, y = var_313)[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, [3, 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]>([3, 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, [3, 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, [3, 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, [3, 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, [3, 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, [3, 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]>([3, 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, [3, 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, [3, 256, 16]> inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_56, 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, [3, 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, [3, 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_298_begin_0 = const()[name = tensor<string, []>("op_298_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_298_end_0 = const()[name = tensor<string, []>("op_298_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
-            tensor<bool, [3]> var_298_end_mask_0 = const()[name = tensor<string, []>("op_298_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [3, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor<string, []>("transpose_53")];
-            tensor<fp32, [3, 1, 256]> var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_298")];
-            tensor<int32, [3]> var_300_perm_0 = const()[name = tensor<string, []>("op_300_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 3, 256]> var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor<string, []>("transpose_52")];
-            tensor<fp32, [1, 3, 256]> input_31 = add(x = x_3, y = var_300)[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, 3, 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, 3, 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, 3, 1024]> input_35 = silu(x = inputs_9)[name = tensor<string, []>("input_35")];
-            tensor<fp32, [1, 3, 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_323 = const()[name = tensor<string, []>("op_323"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_324 = mul(x = input_39, y = var_323)[name = tensor<string, []>("op_324")];
-            tensor<fp32, [1, 3, 256]> input_41 = add(x = var_324, 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, 3, 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_344_begin_0 = const()[name = tensor<string, []>("op_344_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_344_end_0 = const()[name = tensor<string, []>("op_344_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
+            tensor<bool, [3]> var_344_end_mask_0 = const()[name = tensor<string, []>("op_344_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [3, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor<string, []>("transpose_53")];
+            tensor<fp32, [3, 1, 256]> var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_344")];
+            tensor<int32, [3]> var_346_perm_0 = const()[name = tensor<string, []>("op_346_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 3, 256]> var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor<string, []>("transpose_52")];
+            tensor<fp32, [1, 3, 256]> input_33 = add(x = x_3, y = var_346)[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, 3, 256]> input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor<string, []>("input_35")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_37 = silu(x = inputs_9)[name = tensor<string, []>("input_37")];
+            tensor<fp32, [1, 3, 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_369 = const()[name = tensor<string, []>("op_369"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_370 = mul(x = input_41, y = var_369)[name = tensor<string, []>("op_370")];
+            tensor<fp32, [1, 3, 256]> input_43 = add(x = var_370, 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, 3, 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, 3, 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, 3, 1024]> input_47 = silu(x = inputs_11)[name = tensor<string, []>("input_47")];
-            tensor<fp32, [1, 3, 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_353 = const()[name = tensor<string, []>("op_353"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_354 = mul(x = input_51, y = var_353)[name = tensor<string, []>("op_354")];
-            tensor<fp32, [1, 3, 256]> input_53 = add(x = var_354, y = input_43)[name = tensor<string, []>("input_53")];
+            tensor<fp32, [1, 3, 256]> input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_56, 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, 3, 256]> input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor<string, []>("input_47")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_49 = silu(x = inputs_11)[name = tensor<string, []>("input_49")];
+            tensor<fp32, [1, 3, 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_399 = const()[name = tensor<string, []>("op_399"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_400 = mul(x = input_53, y = var_399)[name = tensor<string, []>("op_400")];
+            tensor<fp32, [1, 3, 256]> input_55 = add(x = var_400, 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, 3, 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, 3, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_56, 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])];
@@ -406,163 +424,163 @@ 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, 3, 256]> var_368 = 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_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_370 = reshape(shape = var_369, x = var_368)[name = tensor<string, []>("op_370")];
+            tensor<fp32, [1, 3, 256]> var_414 = 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_415 = const()[name = tensor<string, []>("op_415"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_416 = reshape(shape = var_415, x = var_414)[name = tensor<string, []>("op_416")];
             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, 3, 256]> var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_13")];
-            tensor<fp32, []> var_375 = const()[name = tensor<string, []>("op_375"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 3, 256]> var_376 = mul(x = var_374, y = var_375)[name = tensor<string, []>("op_376")];
-            tensor<int32, [4]> var_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_378 = reshape(shape = var_377, x = var_376)[name = tensor<string, []>("op_378")];
+            tensor<fp32, [1, 3, 256]> var_420 = 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_421 = const()[name = tensor<string, []>("op_421"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 3, 256]> var_422 = mul(x = var_420, y = var_421)[name = tensor<string, []>("op_422")];
+            tensor<int32, [4]> var_423 = const()[name = tensor<string, []>("op_423"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_424 = reshape(shape = var_423, x = var_422)[name = tensor<string, []>("op_424")];
             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, 3, 256]> var_382 = 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_383 = const()[name = tensor<string, []>("op_383"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_384 = reshape(shape = var_383, x = var_382)[name = tensor<string, []>("op_384")];
+            tensor<fp32, [1, 3, 256]> var_428 = 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_429 = const()[name = tensor<string, []>("op_429"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_430 = reshape(shape = var_429, x = var_428)[name = tensor<string, []>("op_430")];
             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, 3, 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, 3, 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, [3]> s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor<string, []>("s_t_3")];
+            tensor<fp32, [3]> s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_3")];
             tensor<fp32, [3]> 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, 3, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor<string, []>("transpose_50")];
-            tensor<fp32, [1, 4, 3, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor<string, []>("transpose_51")];
+            tensor<fp32, [1, 4, 3, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor<string, []>("transpose_50")];
+            tensor<fp32, [1, 4, 3, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor<string, []>("transpose_51")];
             tensor<fp32, [1, 4, 3, 3]> 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_394 = const()[name = tensor<string, []>("op_394"), val = tensor<int32, [2]>([3, 1])];
-            tensor<fp32, [3, 1]> var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor<string, []>("op_395")];
-            tensor<fp32, [3, 3]> M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor<string, []>("M_3")];
-            tensor<fp32, [1, 4, 3, 3]> var_397 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_397")];
+            tensor<int32, [2]> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor<string, []>("op_441")];
+            tensor<fp32, [3, 3]> M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor<string, []>("M_3")];
+            tensor<fp32, [1, 4, 3, 3]> var_443 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_443")];
             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, 3, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor<string, []>("transpose_49")];
-            tensor<fp32, [1, 4, 3, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor<string, []>("inner_3")];
-            tensor<bool, []> var_399_transpose_x_0 = const()[name = tensor<string, []>("op_399_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_399_transpose_y_0 = const()[name = tensor<string, []>("op_399_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 3, 64]> var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_399")];
-            tensor<fp32, [3]> var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_400")];
-            tensor<int32, [4]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_402 = reshape(shape = var_401, x = var_400)[name = tensor<string, []>("op_402")];
-            tensor<fp32, [1, 4, 3, 64]> cross_3 = mul(x = var_399, y = var_402)[name = tensor<string, []>("cross_3")];
+            tensor<fp32, [1, 4, 3, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor<string, []>("transpose_49")];
+            tensor<fp32, [1, 4, 3, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor<string, []>("inner_3")];
+            tensor<bool, []> var_445_transpose_x_0 = const()[name = tensor<string, []>("op_445_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_445_transpose_y_0 = const()[name = tensor<string, []>("op_445_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 3, 64]> var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_445")];
+            tensor<fp32, [3]> var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_446")];
+            tensor<int32, [4]> var_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_448 = reshape(shape = var_447, x = var_446)[name = tensor<string, []>("op_448")];
+            tensor<fp32, [1, 4, 3, 64]> cross_3 = mul(x = var_445, y = var_448)[name = tensor<string, []>("cross_3")];
             tensor<fp32, [1, 4, 3, 64]> out_7 = add(x = inner_3, y = cross_3)[name = tensor<string, []>("out_7")];
-            tensor<fp32, [1, 4, 64, 64]> var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_405")];
-            tensor<bool, []> var_407_transpose_x_1 = const()[name = tensor<string, []>("op_407_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_407_transpose_y_1 = const()[name = tensor<string, []>("op_407_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_407")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor<string, []>("new_kv_unnorm_3")];
-            tensor<fp32, []> var_409 = const()[name = tensor<string, []>("op_409"), val = tensor<fp32, []>(0x1.8p+1)];
-            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor<string, []>("new_scale_3")];
-            tensor<fp32, [1]> var_411 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_411")];
-            tensor<fp32, [1, 4, 64, 64]> var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor<string, []>("op_412")];
-            tensor<int32, [4]> var_413_perm_0 = const()[name = tensor<string, []>("op_413_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_451")];
+            tensor<bool, []> var_453_transpose_x_1 = const()[name = tensor<string, []>("op_453_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_453_transpose_y_1 = const()[name = tensor<string, []>("op_453_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_453")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor<string, []>("new_kv_unnorm_3")];
+            tensor<fp32, []> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor<string, []>("new_scale_3")];
+            tensor<fp32, [1]> var_457 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_457")];
+            tensor<fp32, [1, 4, 64, 64]> var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor<string, []>("op_458")];
+            tensor<int32, [4]> var_459_perm_0 = const()[name = tensor<string, []>("op_459_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, 3, 4, 64]> var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
-            tensor<fp32, [1, 3, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor<string, []>("out_9")];
-            tensor<int32, [3]> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<fp32, [1, 3, 256]> out_11 = reshape(shape = var_417, x = out_9)[name = tensor<string, []>("out_11")];
-            tensor<fp32, [1, 3, 256]> var_419 = silu(x = input_57)[name = tensor<string, []>("op_419")];
-            tensor<fp32, [1, 3, 256]> input_59 = mul(x = var_419, y = out_11)[name = tensor<string, []>("input_59")];
-            tensor<fp32, [1, 3, 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, 3, 256]> x_9 = add(x = input_53, y = ret_out_3)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 3, 4, 64]> var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
+            tensor<fp32, [1, 3, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_64, x = var_459)[name = tensor<string, []>("out_9")];
+            tensor<int32, [3]> var_463 = const()[name = tensor<string, []>("op_463"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<fp32, [1, 3, 256]> out_11 = reshape(shape = var_463, x = out_9)[name = tensor<string, []>("out_11")];
+            tensor<fp32, [1, 3, 256]> var_465 = silu(x = input_59)[name = tensor<string, []>("op_465")];
+            tensor<fp32, [1, 3, 256]> input_61 = mul(x = var_465, y = out_11)[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 3, 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, 3, 256]> x_9 = add(x = input_55, y = ret_out_3)[name = tensor<string, []>("x_9")];
             tensor<int32, [4]> window_9_begin_0 = const()[name = tensor<string, []>("window_9_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> window_9_end_0 = const()[name = tensor<string, []>("window_9_end_0"), val = tensor<int32, [4]>([2, 1, 16, 256])];
             tensor<bool, [4]> window_9_end_mask_0 = const()[name = tensor<string, []>("window_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_9_squeeze_mask_0 = const()[name = tensor<string, []>("window_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[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, 0, 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, false, 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<int32, [3]> var_473_begin_0 = const()[name = tensor<string, []>("op_473_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_473_end_0 = const()[name = tensor<string, []>("op_473_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_473_end_mask_0 = const()[name = tensor<string, []>("op_473_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor<string, []>("op_473")];
+            tensor<int32, [3]> var_476_begin_0 = const()[name = tensor<string, []>("op_476_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_476_end_0 = const()[name = tensor<string, []>("op_476_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_476_end_mask_0 = const()[name = tensor<string, []>("op_476_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor<string, []>("op_476")];
             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<int32, [3]> var_435_begin_0 = const()[name = tensor<string, []>("op_435_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_435_end_0 = const()[name = tensor<string, []>("op_435_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_435_end_mask_0 = const()[name = tensor<string, []>("op_435_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor<string, []>("op_435")];
-            tensor<int32, [3]> var_438_begin_0 = const()[name = tensor<string, []>("op_438_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_438_end_0 = const()[name = tensor<string, []>("op_438_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_438_end_mask_0 = const()[name = tensor<string, []>("op_438_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor<string, []>("op_438")];
+            tensor<fp32, [1, 16, 256]> window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor<string, []>("window_11")];
+            tensor<int32, [3]> var_481_begin_0 = const()[name = tensor<string, []>("op_481_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_481_end_0 = const()[name = tensor<string, []>("op_481_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_481_end_mask_0 = const()[name = tensor<string, []>("op_481_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor<string, []>("op_481")];
+            tensor<int32, [3]> var_484_begin_0 = const()[name = tensor<string, []>("op_484_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_484_end_0 = const()[name = tensor<string, []>("op_484_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_484_end_mask_0 = const()[name = tensor<string, []>("op_484_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor<string, []>("op_484")];
             tensor<bool, []> window_13_interleave_0 = const()[name = tensor<string, []>("window_13_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor<string, []>("window_13")];
-            tensor<int32, [3]> var_443_begin_0 = const()[name = tensor<string, []>("op_443_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_443_end_0 = const()[name = tensor<string, []>("op_443_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_443_end_mask_0 = const()[name = tensor<string, []>("op_443_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor<string, []>("op_443")];
-            tensor<int32, [3]> var_446_begin_0 = const()[name = tensor<string, []>("op_446_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_446_end_0 = const()[name = tensor<string, []>("op_446_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_446_end_mask_0 = const()[name = tensor<string, []>("op_446_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor<string, []>("op_446")];
+            tensor<fp32, [1, 16, 256]> window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor<string, []>("window_13")];
+            tensor<int32, [3]> var_489_begin_0 = const()[name = tensor<string, []>("op_489_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_489_end_0 = const()[name = tensor<string, []>("op_489_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_489_end_mask_0 = const()[name = tensor<string, []>("op_489_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor<string, []>("op_489")];
+            tensor<int32, [3]> var_492_begin_0 = const()[name = tensor<string, []>("op_492_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_492_end_0 = const()[name = tensor<string, []>("op_492_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_492_end_mask_0 = const()[name = tensor<string, []>("op_492_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor<string, []>("op_492")];
             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_446, var_443))[name = tensor<string, []>("window_15")];
-            tensor<bool, []> input_61_interleave_0 = const()[name = tensor<string, []>("input_61_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [3, 16, 256]> input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 16, 256]> window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor<string, []>("window_15")];
+            tensor<bool, []> input_63_interleave_0 = const()[name = tensor<string, []>("input_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [3, 16, 256]> input_63 = concat(axis = var_59, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[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, [3, 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, [3, 16, 256]> x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_56, 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, [3, 256, 16]> input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
-            tensor<fp32, [3, 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_471_split_sizes_0 = const()[name = tensor<string, []>("op_471_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_471_axis_0 = const()[name = tensor<string, []>("op_471_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [3, 256, 16]> var_471_0, tensor<fp32, [3, 256, 16]> var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_471")];
-            tensor<fp32, [3, 256, 16]> var_473 = sigmoid(x = var_471_1)[name = tensor<string, []>("op_473")];
-            tensor<fp32, [3, 256, 16]> inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor<string, []>("inputs_15")];
+            tensor<fp32, [3, 256, 16]> input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
+            tensor<fp32, [3, 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_517_split_sizes_0 = const()[name = tensor<string, []>("op_517_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_517_axis_0 = const()[name = tensor<string, []>("op_517_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [3, 256, 16]> var_517_0, tensor<fp32, [3, 256, 16]> var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_517")];
+            tensor<fp32, [3, 256, 16]> var_519 = sigmoid(x = var_517_1)[name = tensor<string, []>("op_519")];
+            tensor<fp32, [3, 256, 16]> inputs_15 = mul(x = var_517_0, y = var_519)[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, [3, 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]>([3, 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, [3, 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, [3, 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, [3, 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, [3, 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, [3, 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]>([3, 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, [3, 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, [3, 256, 16]> inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_56, 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, [3, 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, [3, 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_504_begin_0 = const()[name = tensor<string, []>("op_504_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_504_end_0 = const()[name = tensor<string, []>("op_504_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
-            tensor<bool, [3]> var_504_end_mask_0 = const()[name = tensor<string, []>("op_504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [3, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor<string, []>("transpose_46")];
-            tensor<fp32, [3, 1, 256]> var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_504")];
-            tensor<int32, [3]> var_506_perm_0 = const()[name = tensor<string, []>("op_506_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 3, 256]> var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor<string, []>("transpose_45")];
-            tensor<fp32, [1, 3, 256]> input_71 = add(x = x_9, y = var_506)[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, 3, 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, 3, 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, 3, 1024]> input_75 = silu(x = inputs_19)[name = tensor<string, []>("input_75")];
-            tensor<fp32, [1, 3, 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_529 = const()[name = tensor<string, []>("op_529"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_530 = mul(x = input_79, y = var_529)[name = tensor<string, []>("op_530")];
-            tensor<fp32, [1, 3, 256]> input_81 = add(x = var_530, 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, 3, 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_550_begin_0 = const()[name = tensor<string, []>("op_550_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_550_end_0 = const()[name = tensor<string, []>("op_550_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
+            tensor<bool, [3]> var_550_end_mask_0 = const()[name = tensor<string, []>("op_550_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [3, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor<string, []>("transpose_46")];
+            tensor<fp32, [3, 1, 256]> var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_550")];
+            tensor<int32, [3]> var_552_perm_0 = const()[name = tensor<string, []>("op_552_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 3, 256]> var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor<string, []>("transpose_45")];
+            tensor<fp32, [1, 3, 256]> input_73 = add(x = x_9, y = var_552)[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, 3, 256]> input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor<string, []>("input_75")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_77 = silu(x = inputs_19)[name = tensor<string, []>("input_77")];
+            tensor<fp32, [1, 3, 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_575 = const()[name = tensor<string, []>("op_575"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_576 = mul(x = input_81, y = var_575)[name = tensor<string, []>("op_576")];
+            tensor<fp32, [1, 3, 256]> input_83 = add(x = var_576, 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, 3, 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, 3, 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, 3, 1024]> input_87 = silu(x = inputs_21)[name = tensor<string, []>("input_87")];
-            tensor<fp32, [1, 3, 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_559 = const()[name = tensor<string, []>("op_559"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_560 = mul(x = input_91, y = var_559)[name = tensor<string, []>("op_560")];
-            tensor<fp32, [1, 3, 256]> input_93 = add(x = var_560, y = input_83)[name = tensor<string, []>("input_93")];
+            tensor<fp32, [1, 3, 256]> input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_56, 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, 3, 256]> input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor<string, []>("input_87")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_89 = silu(x = inputs_21)[name = tensor<string, []>("input_89")];
+            tensor<fp32, [1, 3, 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_605 = const()[name = tensor<string, []>("op_605"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_606 = mul(x = input_93, y = var_605)[name = tensor<string, []>("op_606")];
+            tensor<fp32, [1, 3, 256]> input_95 = add(x = var_606, 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, 3, 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, 3, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_56, 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])];
@@ -573,163 +591,163 @@ 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, 3, 256]> var_574 = 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_575 = const()[name = tensor<string, []>("op_575"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_576 = reshape(shape = var_575, x = var_574)[name = tensor<string, []>("op_576")];
+            tensor<fp32, [1, 3, 256]> var_620 = 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_621 = const()[name = tensor<string, []>("op_621"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_622 = reshape(shape = var_621, x = var_620)[name = tensor<string, []>("op_622")];
             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, 3, 256]> var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_22")];
-            tensor<fp32, []> var_581 = const()[name = tensor<string, []>("op_581"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 3, 256]> var_582 = mul(x = var_580, y = var_581)[name = tensor<string, []>("op_582")];
-            tensor<int32, [4]> var_583 = const()[name = tensor<string, []>("op_583"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_584 = reshape(shape = var_583, x = var_582)[name = tensor<string, []>("op_584")];
+            tensor<fp32, [1, 3, 256]> var_626 = 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_627 = const()[name = tensor<string, []>("op_627"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 3, 256]> var_628 = mul(x = var_626, y = var_627)[name = tensor<string, []>("op_628")];
+            tensor<int32, [4]> var_629 = const()[name = tensor<string, []>("op_629"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_630 = reshape(shape = var_629, x = var_628)[name = tensor<string, []>("op_630")];
             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, 3, 256]> var_588 = 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_589 = const()[name = tensor<string, []>("op_589"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_590 = reshape(shape = var_589, x = var_588)[name = tensor<string, []>("op_590")];
+            tensor<fp32, [1, 3, 256]> var_634 = 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_635 = const()[name = tensor<string, []>("op_635"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_636 = reshape(shape = var_635, x = var_634)[name = tensor<string, []>("op_636")];
             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, 3, 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, 3, 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, [3]> s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor<string, []>("s_t_5")];
+            tensor<fp32, [3]> s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_5")];
             tensor<fp32, [3]> 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, 3, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor<string, []>("transpose_43")];
-            tensor<fp32, [1, 4, 3, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor<string, []>("transpose_44")];
+            tensor<fp32, [1, 4, 3, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor<string, []>("transpose_43")];
+            tensor<fp32, [1, 4, 3, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor<string, []>("transpose_44")];
             tensor<fp32, [1, 4, 3, 3]> 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_600 = const()[name = tensor<string, []>("op_600"), val = tensor<int32, [2]>([3, 1])];
-            tensor<fp32, [3, 1]> var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor<string, []>("op_601")];
-            tensor<fp32, [3, 3]> M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor<string, []>("M_5")];
-            tensor<fp32, [1, 4, 3, 3]> var_603 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_603")];
+            tensor<int32, [2]> var_646 = const()[name = tensor<string, []>("op_646"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor<string, []>("op_647")];
+            tensor<fp32, [3, 3]> M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor<string, []>("M_5")];
+            tensor<fp32, [1, 4, 3, 3]> var_649 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_649")];
             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, 3, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor<string, []>("transpose_42")];
-            tensor<fp32, [1, 4, 3, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor<string, []>("inner_5")];
-            tensor<bool, []> var_605_transpose_x_0 = const()[name = tensor<string, []>("op_605_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_605_transpose_y_0 = const()[name = tensor<string, []>("op_605_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 3, 64]> var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_605")];
-            tensor<fp32, [3]> var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_606")];
-            tensor<int32, [4]> var_607 = const()[name = tensor<string, []>("op_607"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_608 = reshape(shape = var_607, x = var_606)[name = tensor<string, []>("op_608")];
-            tensor<fp32, [1, 4, 3, 64]> cross_5 = mul(x = var_605, y = var_608)[name = tensor<string, []>("cross_5")];
+            tensor<fp32, [1, 4, 3, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor<string, []>("transpose_42")];
+            tensor<fp32, [1, 4, 3, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor<string, []>("inner_5")];
+            tensor<bool, []> var_651_transpose_x_0 = const()[name = tensor<string, []>("op_651_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_651_transpose_y_0 = const()[name = tensor<string, []>("op_651_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 3, 64]> var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_651")];
+            tensor<fp32, [3]> var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_652")];
+            tensor<int32, [4]> var_653 = const()[name = tensor<string, []>("op_653"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_654 = reshape(shape = var_653, x = var_652)[name = tensor<string, []>("op_654")];
+            tensor<fp32, [1, 4, 3, 64]> cross_5 = mul(x = var_651, y = var_654)[name = tensor<string, []>("cross_5")];
             tensor<fp32, [1, 4, 3, 64]> out_13 = add(x = inner_5, y = cross_5)[name = tensor<string, []>("out_13")];
-            tensor<fp32, [1, 4, 64, 64]> var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_611")];
-            tensor<bool, []> var_613_transpose_x_1 = const()[name = tensor<string, []>("op_613_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_613_transpose_y_1 = const()[name = tensor<string, []>("op_613_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_613")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor<string, []>("new_kv_unnorm_5")];
-            tensor<fp32, []> var_615 = const()[name = tensor<string, []>("op_615"), val = tensor<fp32, []>(0x1.8p+1)];
-            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor<string, []>("new_scale_5")];
-            tensor<fp32, [1]> var_617 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_617")];
-            tensor<fp32, [1, 4, 64, 64]> var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor<string, []>("op_618")];
-            tensor<int32, [4]> var_619_perm_0 = const()[name = tensor<string, []>("op_619_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_657")];
+            tensor<bool, []> var_659_transpose_x_1 = const()[name = tensor<string, []>("op_659_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_659_transpose_y_1 = const()[name = tensor<string, []>("op_659_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_659")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor<string, []>("new_kv_unnorm_5")];
+            tensor<fp32, []> var_661 = const()[name = tensor<string, []>("op_661"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor<string, []>("new_scale_5")];
+            tensor<fp32, [1]> var_663 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_663")];
+            tensor<fp32, [1, 4, 64, 64]> var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor<string, []>("op_664")];
+            tensor<int32, [4]> var_665_perm_0 = const()[name = tensor<string, []>("op_665_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, 3, 4, 64]> var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
-            tensor<fp32, [1, 3, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor<string, []>("out_15")];
-            tensor<int32, [3]> var_623 = const()[name = tensor<string, []>("op_623"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<fp32, [1, 3, 256]> out_17 = reshape(shape = var_623, x = out_15)[name = tensor<string, []>("out_17")];
-            tensor<fp32, [1, 3, 256]> var_625 = silu(x = input_97)[name = tensor<string, []>("op_625")];
-            tensor<fp32, [1, 3, 256]> input_99 = mul(x = var_625, y = out_17)[name = tensor<string, []>("input_99")];
-            tensor<fp32, [1, 3, 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, 3, 256]> x_15 = add(x = input_93, y = ret_out_5)[name = tensor<string, []>("x_15")];
+            tensor<fp32, [1, 3, 4, 64]> var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
+            tensor<fp32, [1, 3, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_64, x = var_665)[name = tensor<string, []>("out_15")];
+            tensor<int32, [3]> var_669 = const()[name = tensor<string, []>("op_669"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<fp32, [1, 3, 256]> out_17 = reshape(shape = var_669, x = out_15)[name = tensor<string, []>("out_17")];
+            tensor<fp32, [1, 3, 256]> var_671 = silu(x = input_99)[name = tensor<string, []>("op_671")];
+            tensor<fp32, [1, 3, 256]> input_101 = mul(x = var_671, y = out_17)[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 3, 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, 3, 256]> x_15 = add(x = input_95, y = ret_out_5)[name = tensor<string, []>("x_15")];
             tensor<int32, [4]> window_17_begin_0 = const()[name = tensor<string, []>("window_17_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> window_17_end_0 = const()[name = tensor<string, []>("window_17_end_0"), val = tensor<int32, [4]>([3, 1, 16, 256])];
             tensor<bool, [4]> window_17_end_mask_0 = const()[name = tensor<string, []>("window_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_17_squeeze_mask_0 = const()[name = tensor<string, []>("window_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_17")];
-            tensor<int32, [3]> var_633_begin_0 = const()[name = tensor<string, []>("op_633_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_633_end_0 = const()[name = tensor<string, []>("op_633_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_633_end_mask_0 = const()[name = tensor<string, []>("op_633_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor<string, []>("op_633")];
-            tensor<int32, [3]> var_636_begin_0 = const()[name = tensor<string, []>("op_636_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_636_end_0 = const()[name = tensor<string, []>("op_636_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_636_end_mask_0 = const()[name = tensor<string, []>("op_636_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor<string, []>("op_636")];
+            tensor<int32, [3]> var_679_begin_0 = const()[name = tensor<string, []>("op_679_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_679_end_0 = const()[name = tensor<string, []>("op_679_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_679_end_mask_0 = const()[name = tensor<string, []>("op_679_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor<string, []>("op_679")];
+            tensor<int32, [3]> var_682_begin_0 = const()[name = tensor<string, []>("op_682_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_682_end_0 = const()[name = tensor<string, []>("op_682_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_682_end_mask_0 = const()[name = tensor<string, []>("op_682_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor<string, []>("op_682")];
             tensor<bool, []> window_19_interleave_0 = const()[name = tensor<string, []>("window_19_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor<string, []>("window_19")];
-            tensor<int32, [3]> var_641_begin_0 = const()[name = tensor<string, []>("op_641_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_641_end_0 = const()[name = tensor<string, []>("op_641_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_641_end_mask_0 = const()[name = tensor<string, []>("op_641_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor<string, []>("op_641")];
-            tensor<int32, [3]> var_644_begin_0 = const()[name = tensor<string, []>("op_644_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_644_end_0 = const()[name = tensor<string, []>("op_644_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_644_end_mask_0 = const()[name = tensor<string, []>("op_644_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor<string, []>("op_644")];
+            tensor<fp32, [1, 16, 256]> window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor<string, []>("window_19")];
+            tensor<int32, [3]> var_687_begin_0 = const()[name = tensor<string, []>("op_687_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_687_end_0 = const()[name = tensor<string, []>("op_687_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_687_end_mask_0 = const()[name = tensor<string, []>("op_687_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor<string, []>("op_687")];
+            tensor<int32, [3]> var_690_begin_0 = const()[name = tensor<string, []>("op_690_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_690_end_0 = const()[name = tensor<string, []>("op_690_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_690_end_mask_0 = const()[name = tensor<string, []>("op_690_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor<string, []>("op_690")];
             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_644, var_641))[name = tensor<string, []>("window_21")];
-            tensor<int32, [3]> var_649_begin_0 = const()[name = tensor<string, []>("op_649_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_649_end_0 = const()[name = tensor<string, []>("op_649_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_649_end_mask_0 = const()[name = tensor<string, []>("op_649_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor<string, []>("op_649")];
-            tensor<int32, [3]> var_652_begin_0 = const()[name = tensor<string, []>("op_652_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_652_end_0 = const()[name = tensor<string, []>("op_652_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_652_end_mask_0 = const()[name = tensor<string, []>("op_652_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor<string, []>("op_652")];
+            tensor<fp32, [1, 16, 256]> window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor<string, []>("window_21")];
+            tensor<int32, [3]> var_695_begin_0 = const()[name = tensor<string, []>("op_695_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_695_end_0 = const()[name = tensor<string, []>("op_695_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_695_end_mask_0 = const()[name = tensor<string, []>("op_695_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor<string, []>("op_695")];
+            tensor<int32, [3]> var_698_begin_0 = const()[name = tensor<string, []>("op_698_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_698_end_0 = const()[name = tensor<string, []>("op_698_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_698_end_mask_0 = const()[name = tensor<string, []>("op_698_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor<string, []>("op_698")];
             tensor<bool, []> window_23_interleave_0 = const()[name = tensor<string, []>("window_23_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor<string, []>("window_23")];
-            tensor<bool, []> input_101_interleave_0 = const()[name = tensor<string, []>("input_101_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [3, 16, 256]> input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 16, 256]> window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor<string, []>("window_23")];
+            tensor<bool, []> input_103_interleave_0 = const()[name = tensor<string, []>("input_103_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [3, 16, 256]> input_103 = concat(axis = var_59, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[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, [3, 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, [3, 16, 256]> x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_56, 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, [3, 256, 16]> input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
-            tensor<fp32, [3, 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_677_split_sizes_0 = const()[name = tensor<string, []>("op_677_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_677_axis_0 = const()[name = tensor<string, []>("op_677_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [3, 256, 16]> var_677_0, tensor<fp32, [3, 256, 16]> var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_677")];
-            tensor<fp32, [3, 256, 16]> var_679 = sigmoid(x = var_677_1)[name = tensor<string, []>("op_679")];
-            tensor<fp32, [3, 256, 16]> inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor<string, []>("inputs_25")];
+            tensor<fp32, [3, 256, 16]> input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
+            tensor<fp32, [3, 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_723_split_sizes_0 = const()[name = tensor<string, []>("op_723_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_723_axis_0 = const()[name = tensor<string, []>("op_723_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [3, 256, 16]> var_723_0, tensor<fp32, [3, 256, 16]> var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_723")];
+            tensor<fp32, [3, 256, 16]> var_725 = sigmoid(x = var_723_1)[name = tensor<string, []>("op_725")];
+            tensor<fp32, [3, 256, 16]> inputs_25 = mul(x = var_723_0, y = var_725)[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, [3, 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]>([3, 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, [3, 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, [3, 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, [3, 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, [3, 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, [3, 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]>([3, 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, [3, 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, [3, 256, 16]> inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_56, 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, [3, 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, [3, 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_710_begin_0 = const()[name = tensor<string, []>("op_710_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_710_end_0 = const()[name = tensor<string, []>("op_710_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
-            tensor<bool, [3]> var_710_end_mask_0 = const()[name = tensor<string, []>("op_710_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [3, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor<string, []>("transpose_39")];
-            tensor<fp32, [3, 1, 256]> var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_710")];
-            tensor<int32, [3]> var_712_perm_0 = const()[name = tensor<string, []>("op_712_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 3, 256]> var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor<string, []>("transpose_38")];
-            tensor<fp32, [1, 3, 256]> input_111 = add(x = x_15, y = var_712)[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, 3, 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, 3, 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, 3, 1024]> input_115 = silu(x = inputs_29)[name = tensor<string, []>("input_115")];
-            tensor<fp32, [1, 3, 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_735 = const()[name = tensor<string, []>("op_735"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_736 = mul(x = input_119, y = var_735)[name = tensor<string, []>("op_736")];
-            tensor<fp32, [1, 3, 256]> input_121 = add(x = var_736, 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, 3, 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_756_begin_0 = const()[name = tensor<string, []>("op_756_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_756_end_0 = const()[name = tensor<string, []>("op_756_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
+            tensor<bool, [3]> var_756_end_mask_0 = const()[name = tensor<string, []>("op_756_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [3, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor<string, []>("transpose_39")];
+            tensor<fp32, [3, 1, 256]> var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_756")];
+            tensor<int32, [3]> var_758_perm_0 = const()[name = tensor<string, []>("op_758_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 3, 256]> var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor<string, []>("transpose_38")];
+            tensor<fp32, [1, 3, 256]> input_113 = add(x = x_15, y = var_758)[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, 3, 256]> input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor<string, []>("input_115")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_117 = silu(x = inputs_29)[name = tensor<string, []>("input_117")];
+            tensor<fp32, [1, 3, 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_781 = const()[name = tensor<string, []>("op_781"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_782 = mul(x = input_121, y = var_781)[name = tensor<string, []>("op_782")];
+            tensor<fp32, [1, 3, 256]> input_123 = add(x = var_782, 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, 3, 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, 3, 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, 3, 1024]> input_127 = silu(x = inputs_31)[name = tensor<string, []>("input_127")];
-            tensor<fp32, [1, 3, 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_765 = const()[name = tensor<string, []>("op_765"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_766 = mul(x = input_131, y = var_765)[name = tensor<string, []>("op_766")];
-            tensor<fp32, [1, 3, 256]> input_133 = add(x = var_766, y = input_123)[name = tensor<string, []>("input_133")];
+            tensor<fp32, [1, 3, 256]> input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_56, 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, 3, 256]> input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor<string, []>("input_127")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_129 = silu(x = inputs_31)[name = tensor<string, []>("input_129")];
+            tensor<fp32, [1, 3, 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_811 = const()[name = tensor<string, []>("op_811"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_812 = mul(x = input_133, y = var_811)[name = tensor<string, []>("op_812")];
+            tensor<fp32, [1, 3, 256]> input_135 = add(x = var_812, 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, 3, 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, 3, 256]> x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_56, 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])];
@@ -740,199 +758,192 @@ 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, 3, 256]> var_780 = 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_781 = const()[name = tensor<string, []>("op_781"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_782 = reshape(shape = var_781, x = var_780)[name = tensor<string, []>("op_782")];
+            tensor<fp32, [1, 3, 256]> var_826 = 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_827 = const()[name = tensor<string, []>("op_827"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_828 = reshape(shape = var_827, x = var_826)[name = tensor<string, []>("op_828")];
             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, 3, 256]> var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_31")];
-            tensor<fp32, []> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 3, 256]> var_788 = mul(x = var_786, y = var_787)[name = tensor<string, []>("op_788")];
-            tensor<int32, [4]> var_789 = const()[name = tensor<string, []>("op_789"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_790 = reshape(shape = var_789, x = var_788)[name = tensor<string, []>("op_790")];
+            tensor<fp32, [1, 3, 256]> var_832 = 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_833 = const()[name = tensor<string, []>("op_833"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 3, 256]> var_834 = mul(x = var_832, y = var_833)[name = tensor<string, []>("op_834")];
+            tensor<int32, [4]> var_835 = const()[name = tensor<string, []>("op_835"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_836 = reshape(shape = var_835, x = var_834)[name = tensor<string, []>("op_836")];
             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, 3, 256]> var_794 = 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_795 = const()[name = tensor<string, []>("op_795"), val = tensor<int32, [4]>([1, 3, 4, 64])];
-            tensor<fp32, [1, 3, 4, 64]> var_796 = reshape(shape = var_795, x = var_794)[name = tensor<string, []>("op_796")];
+            tensor<fp32, [1, 3, 256]> var_840 = 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_841 = const()[name = tensor<string, []>("op_841"), val = tensor<int32, [4]>([1, 3, 4, 64])];
+            tensor<fp32, [1, 3, 4, 64]> var_842 = reshape(shape = var_841, x = var_840)[name = tensor<string, []>("op_842")];
             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, 3, 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, 3, 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, [3]> s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor<string, []>("s_t_7")];
+            tensor<fp32, [3]> s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_7")];
             tensor<fp32, [3]> 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, 3, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor<string, []>("transpose_36")];
-            tensor<fp32, [1, 4, 3, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor<string, []>("transpose_37")];
+            tensor<fp32, [1, 4, 3, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor<string, []>("transpose_36")];
+            tensor<fp32, [1, 4, 3, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor<string, []>("transpose_37")];
             tensor<fp32, [1, 4, 3, 3]> 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_806 = const()[name = tensor<string, []>("op_806"), val = tensor<int32, [2]>([3, 1])];
-            tensor<fp32, [3, 1]> var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor<string, []>("op_807")];
-            tensor<fp32, [3, 3]> M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor<string, []>("M_7")];
-            tensor<fp32, [1, 4, 3, 3]> var_809 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_809")];
+            tensor<int32, [2]> var_852 = const()[name = tensor<string, []>("op_852"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor<string, []>("op_853")];
+            tensor<fp32, [3, 3]> M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor<string, []>("M_7")];
+            tensor<fp32, [1, 4, 3, 3]> var_855 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_855")];
             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, 3, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor<string, []>("transpose_35")];
-            tensor<fp32, [1, 4, 3, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor<string, []>("inner_7")];
-            tensor<bool, []> var_811_transpose_x_0 = const()[name = tensor<string, []>("op_811_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_811_transpose_y_0 = const()[name = tensor<string, []>("op_811_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 3, 64]> var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_811")];
-            tensor<fp32, [3]> var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_812")];
-            tensor<int32, [4]> var_813 = const()[name = tensor<string, []>("op_813"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_814 = reshape(shape = var_813, x = var_812)[name = tensor<string, []>("op_814")];
-            tensor<fp32, [1, 4, 3, 64]> cross_7 = mul(x = var_811, y = var_814)[name = tensor<string, []>("cross_7")];
+            tensor<fp32, [1, 4, 3, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor<string, []>("transpose_35")];
+            tensor<fp32, [1, 4, 3, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor<string, []>("inner_7")];
+            tensor<bool, []> var_857_transpose_x_0 = const()[name = tensor<string, []>("op_857_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_857_transpose_y_0 = const()[name = tensor<string, []>("op_857_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 3, 64]> var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_857")];
+            tensor<fp32, [3]> var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_858")];
+            tensor<int32, [4]> var_859 = const()[name = tensor<string, []>("op_859"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_860 = reshape(shape = var_859, x = var_858)[name = tensor<string, []>("op_860")];
+            tensor<fp32, [1, 4, 3, 64]> cross_7 = mul(x = var_857, y = var_860)[name = tensor<string, []>("cross_7")];
             tensor<fp32, [1, 4, 3, 64]> out_19 = add(x = inner_7, y = cross_7)[name = tensor<string, []>("out_19")];
-            tensor<fp32, [1, 4, 64, 64]> var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_817")];
-            tensor<bool, []> var_819_transpose_x_1 = const()[name = tensor<string, []>("op_819_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_819_transpose_y_1 = const()[name = tensor<string, []>("op_819_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_819")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor<string, []>("new_kv_unnorm_7")];
-            tensor<fp32, []> var_821 = const()[name = tensor<string, []>("op_821"), val = tensor<fp32, []>(0x1.8p+1)];
-            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor<string, []>("new_scale_7")];
-            tensor<fp32, [1]> var_823 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_823")];
-            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor<string, []>("nkv_1")];
-            tensor<int32, [4]> var_825_perm_0 = const()[name = tensor<string, []>("op_825_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_863")];
+            tensor<bool, []> var_865_transpose_x_1 = const()[name = tensor<string, []>("op_865_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_865_transpose_y_1 = const()[name = tensor<string, []>("op_865_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_865")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor<string, []>("new_kv_unnorm_7")];
+            tensor<fp32, []> var_867 = const()[name = tensor<string, []>("op_867"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor<string, []>("new_scale_7")];
+            tensor<fp32, [1]> var_869 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_869")];
+            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor<string, []>("nkv_1")];
+            tensor<int32, [4]> var_871_perm_0 = const()[name = tensor<string, []>("op_871_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, 3, 4, 64]> var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
-            tensor<fp32, [1, 3, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor<string, []>("out_21")];
-            tensor<int32, [3]> var_829 = const()[name = tensor<string, []>("op_829"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<fp32, [1, 3, 256]> out_23 = reshape(shape = var_829, x = out_21)[name = tensor<string, []>("out_23")];
-            tensor<fp32, [1, 3, 256]> var_831 = silu(x = input_137)[name = tensor<string, []>("op_831")];
-            tensor<fp32, [1, 3, 256]> input_139 = mul(x = var_831, y = out_23)[name = tensor<string, []>("input_139")];
-            tensor<fp32, [1, 3, 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, 3, 256]> x_21 = add(x = input_133, y = ret_out_7)[name = tensor<string, []>("x_21")];
+            tensor<fp32, [1, 3, 4, 64]> var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
+            tensor<fp32, [1, 3, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_64, x = var_871)[name = tensor<string, []>("out_21")];
+            tensor<int32, [3]> var_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<fp32, [1, 3, 256]> out_23 = reshape(shape = var_875, x = out_21)[name = tensor<string, []>("out_23")];
+            tensor<fp32, [1, 3, 256]> var_877 = silu(x = input_139)[name = tensor<string, []>("op_877")];
+            tensor<fp32, [1, 3, 256]> input_141 = mul(x = var_877, y = out_23)[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 3, 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, 3, 256]> x_21 = add(x = input_135, y = ret_out_7)[name = tensor<string, []>("x_21")];
             tensor<int32, [4]> window_25_begin_0 = const()[name = tensor<string, []>("window_25_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
             tensor<int32, [4]> window_25_end_0 = const()[name = tensor<string, []>("window_25_end_0"), val = tensor<int32, [4]>([4, 1, 16, 256])];
             tensor<bool, [4]> window_25_end_mask_0 = const()[name = tensor<string, []>("window_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_25_squeeze_mask_0 = const()[name = tensor<string, []>("window_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_25")];
-            tensor<int32, [3]> var_839_begin_0 = const()[name = tensor<string, []>("op_839_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_839_end_0 = const()[name = tensor<string, []>("op_839_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_839_end_mask_0 = const()[name = tensor<string, []>("op_839_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor<string, []>("op_839")];
-            tensor<int32, [3]> var_842_begin_0 = const()[name = tensor<string, []>("op_842_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_842_end_0 = const()[name = tensor<string, []>("op_842_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_842_end_mask_0 = const()[name = tensor<string, []>("op_842_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor<string, []>("op_842")];
+            tensor<int32, [3]> var_885_begin_0 = const()[name = tensor<string, []>("op_885_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_885_end_0 = const()[name = tensor<string, []>("op_885_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_885_end_mask_0 = const()[name = tensor<string, []>("op_885_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor<string, []>("op_885")];
+            tensor<int32, [3]> var_888_begin_0 = const()[name = tensor<string, []>("op_888_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_888_end_0 = const()[name = tensor<string, []>("op_888_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_888_end_mask_0 = const()[name = tensor<string, []>("op_888_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor<string, []>("op_888")];
             tensor<bool, []> window_27_interleave_0 = const()[name = tensor<string, []>("window_27_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor<string, []>("window_27")];
-            tensor<int32, [3]> var_847_begin_0 = const()[name = tensor<string, []>("op_847_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_847_end_0 = const()[name = tensor<string, []>("op_847_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_847_end_mask_0 = const()[name = tensor<string, []>("op_847_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor<string, []>("op_847")];
-            tensor<int32, [3]> var_850_begin_0 = const()[name = tensor<string, []>("op_850_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_850_end_0 = const()[name = tensor<string, []>("op_850_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_850_end_mask_0 = const()[name = tensor<string, []>("op_850_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor<string, []>("op_850")];
+            tensor<fp32, [1, 16, 256]> window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor<string, []>("window_27")];
+            tensor<int32, [3]> var_893_begin_0 = const()[name = tensor<string, []>("op_893_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_893_end_0 = const()[name = tensor<string, []>("op_893_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_893_end_mask_0 = const()[name = tensor<string, []>("op_893_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor<string, []>("op_893")];
+            tensor<int32, [3]> var_896_begin_0 = const()[name = tensor<string, []>("op_896_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_896_end_0 = const()[name = tensor<string, []>("op_896_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_896_end_mask_0 = const()[name = tensor<string, []>("op_896_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor<string, []>("op_896")];
             tensor<bool, []> window_29_interleave_0 = const()[name = tensor<string, []>("window_29_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor<string, []>("window_29")];
-            tensor<int32, [3]> var_855_begin_0 = const()[name = tensor<string, []>("op_855_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_855_end_0 = const()[name = tensor<string, []>("op_855_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_855_end_mask_0 = const()[name = tensor<string, []>("op_855_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor<string, []>("op_855")];
-            tensor<int32, [3]> var_858_begin_0 = const()[name = tensor<string, []>("op_858_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_858_end_0 = const()[name = tensor<string, []>("op_858_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_858_end_mask_0 = const()[name = tensor<string, []>("op_858_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor<string, []>("op_858")];
+            tensor<fp32, [1, 16, 256]> window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor<string, []>("window_29")];
+            tensor<int32, [3]> var_901_begin_0 = const()[name = tensor<string, []>("op_901_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_901_end_0 = const()[name = tensor<string, []>("op_901_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_901_end_mask_0 = const()[name = tensor<string, []>("op_901_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor<string, []>("op_901")];
+            tensor<int32, [3]> var_904_begin_0 = const()[name = tensor<string, []>("op_904_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_904_end_0 = const()[name = tensor<string, []>("op_904_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_904_end_mask_0 = const()[name = tensor<string, []>("op_904_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor<string, []>("op_904")];
             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_858, var_855))[name = tensor<string, []>("window")];
-            tensor<bool, []> input_141_interleave_0 = const()[name = tensor<string, []>("input_141_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [3, 16, 256]> input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 16, 256]> window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor<string, []>("window")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [3, 16, 256]> input_143 = concat(axis = var_59, interleave = input_143_interleave_0, values = (window_27, window_29, 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, [3, 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, [3, 16, 256]> x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_56, 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, [3, 256, 16]> input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
-            tensor<fp32, [3, 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_883_split_sizes_0 = const()[name = tensor<string, []>("op_883_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_883_axis_0 = const()[name = tensor<string, []>("op_883_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [3, 256, 16]> var_883_0, tensor<fp32, [3, 256, 16]> var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_883")];
-            tensor<fp32, [3, 256, 16]> var_885 = sigmoid(x = var_883_1)[name = tensor<string, []>("op_885")];
-            tensor<fp32, [3, 256, 16]> inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor<string, []>("inputs_35")];
+            tensor<fp32, [3, 256, 16]> input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
+            tensor<fp32, [3, 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_929_split_sizes_0 = const()[name = tensor<string, []>("op_929_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_929_axis_0 = const()[name = tensor<string, []>("op_929_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [3, 256, 16]> var_929_0, tensor<fp32, [3, 256, 16]> var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_929")];
+            tensor<fp32, [3, 256, 16]> var_931 = sigmoid(x = var_929_1)[name = tensor<string, []>("op_931")];
+            tensor<fp32, [3, 256, 16]> inputs_35 = mul(x = var_929_0, y = var_931)[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, [3, 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]>([3, 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, [3, 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, [3, 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, [3, 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, [3, 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, [3, 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]>([3, 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, [3, 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, [3, 256, 16]> inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_56, 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, [3, 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, [3, 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_916_begin_0 = const()[name = tensor<string, []>("op_916_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_916_end_0 = const()[name = tensor<string, []>("op_916_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
-            tensor<bool, [3]> var_916_end_mask_0 = const()[name = tensor<string, []>("op_916_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [3, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor<string, []>("transpose_32")];
-            tensor<fp32, [3, 1, 256]> var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_916")];
-            tensor<int32, [3]> var_918_perm_0 = const()[name = tensor<string, []>("op_918_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 3, 256]> var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor<string, []>("transpose_31")];
-            tensor<fp32, [1, 3, 256]> input_151 = add(x = x_21, y = var_918)[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, 3, 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, 3, 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, 3, 1024]> input_155 = silu(x = inputs)[name = tensor<string, []>("input_155")];
-            tensor<fp32, [1, 3, 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_941 = const()[name = tensor<string, []>("op_941"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 3, 256]> var_942 = mul(x = input_159, y = var_941)[name = tensor<string, []>("op_942")];
-            tensor<fp32, [1, 3, 256]> input_161 = add(x = var_942, y = input_151)[name = tensor<string, []>("input_161")];
+            tensor<int32, [3]> var_962_begin_0 = const()[name = tensor<string, []>("op_962_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_962_end_0 = const()[name = tensor<string, []>("op_962_end_0"), val = tensor<int32, [3]>([3, 16, 256])];
+            tensor<bool, [3]> var_962_end_mask_0 = const()[name = tensor<string, []>("op_962_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [3, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor<string, []>("transpose_32")];
+            tensor<fp32, [3, 1, 256]> var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_962")];
+            tensor<int32, [3]> var_964_perm_0 = const()[name = tensor<string, []>("op_964_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 3, 256]> var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor<string, []>("transpose_31")];
+            tensor<fp32, [1, 3, 256]> input_153 = add(x = x_21, y = var_964)[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, 3, 256]> input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor<string, []>("input_155")];
+            tensor<fp32, [1, 3, 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, 3, 1024]> input_157 = silu(x = inputs)[name = tensor<string, []>("input_157")];
+            tensor<fp32, [1, 3, 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_987 = const()[name = tensor<string, []>("op_987"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 3, 256]> var_988 = mul(x = input_161, y = var_987)[name = tensor<string, []>("op_988")];
+            tensor<fp32, [1, 3, 256]> input_163 = add(x = var_988, 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, 3, 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, 3, 256]> x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_56, 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, 3]> x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor<string, []>("transpose_30")];
-            tensor<fp32, [1, 256, 21]> cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor<string, []>("cat")];
+            tensor<fp32, [1, 256, 21]> cat = concat(axis = var_61, 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, 3]> 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_960_begin_0 = const()[name = tensor<string, []>("op_960_begin_0"), val = tensor<int32, [3]>([0, 0, 3])];
-            tensor<int32, [3]> var_960_end_0 = const()[name = tensor<string, []>("op_960_end_0"), val = tensor<int32, [3]>([1, 256, 21])];
-            tensor<bool, [3]> var_960_end_mask_0 = const()[name = tensor<string, []>("op_960_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor<string, []>("op_960")];
-            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_962 = const()[name = tensor<string, []>("op_962"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 3, 256]> input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
-            tensor<fp32, [1, 3, 1]> var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor<string, []>("op_963")];
+            tensor<fp32, [1, 256, 3]> 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_1006_begin_0 = const()[name = tensor<string, []>("op_1006_begin_0"), val = tensor<int32, [3]>([0, 0, 3])];
+            tensor<int32, [3]> var_1006_end_0 = const()[name = tensor<string, []>("op_1006_end_0"), val = tensor<int32, [3]>([1, 256, 21])];
+            tensor<bool, [3]> var_1006_end_mask_0 = const()[name = tensor<string, []>("op_1006_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor<string, []>("op_1006")];
+            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_1008 = const()[name = tensor<string, []>("op_1008"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 3, 256]> input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
+            tensor<fp32, [1, 3, 1]> var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_55, x = input_165)[name = tensor<string, []>("op_1009")];
             tensor<fp32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 3, 1]> clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor<string, []>("clip_0")];
-            tensor<fp32, [1, 3, 256]> emb = real_div(x = input_163, y = clip_0)[name = tensor<string, []>("emb")];
-            tensor<int32, []> var_967_axis_0 = const()[name = tensor<string, []>("op_967_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor<string, []>("op_967")];
-            tensor<int32, []> var_969_axis_0 = const()[name = tensor<string, []>("op_969_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_969")];
-            tensor<int32, []> var_971_axis_0 = const()[name = tensor<string, []>("op_971_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor<string, []>("op_971")];
-            tensor<fp32, []> var_980 = const()[name = tensor<string, []>("op_980"), val = tensor<fp32, []>(0x1.5798eep-27)];
-            tensor<fp32, []> var_985 = const()[name = tensor<string, []>("op_985"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
-            tensor<fp32, []> var_987 = const()[name = tensor<string, []>("op_987"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
-            tensor<bool, []> var_988 = const()[name = tensor<string, []>("op_988"), val = tensor<bool, []>(true)];
-            tensor<fp32, []> var_990 = const()[name = tensor<string, []>("op_990"), val = tensor<fp32, []>(0x1p+0)];
-            tensor<int32, []> var_994 = const()[name = tensor<string, []>("op_994"), val = tensor<int32, []>(-1)];
-            tensor<int32, []> var_1000 = const()[name = tensor<string, []>("op_1000"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 3, 1]> clip_0 = clip(alpha = var_69, beta = const_12, x = var_1009)[name = tensor<string, []>("clip_0")];
+            tensor<fp32, [1, 3, 256]> emb = real_div(x = input_165, y = clip_0)[name = tensor<string, []>("emb")];
+            tensor<int32, []> var_1013_axis_0 = const()[name = tensor<string, []>("op_1013_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor<string, []>("op_1013")];
+            tensor<int32, []> var_1015_axis_0 = const()[name = tensor<string, []>("op_1015_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_1015")];
+            tensor<int32, []> var_1017_axis_0 = const()[name = tensor<string, []>("op_1017_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor<string, []>("op_1017")];
             tensor<fp32, [1, 3, 9, 256]> pos = const()[name = tensor<string, []>("pos"), val = tensor<fp32, [1, 3, 9, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44395328)))];
-            tensor<int32, [1]> var_1062_axes_0 = const()[name = tensor<string, []>("op_1062_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<fp32, [1, 3, 1, 256]> var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor<string, []>("op_1062")];
+            tensor<int32, [1]> var_1085_axes_0 = const()[name = tensor<string, []>("op_1085_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 3, 1, 256]> var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor<string, []>("op_1085")];
             tensor<int32, [4]> emb_exp_reps_0 = const()[name = tensor<string, []>("emb_exp_reps_0"), val = tensor<int32, [4]>([1, 1, 9, 1])];
-            tensor<fp32, [1, 3, 9, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[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, 3, 9, 512]> input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_165")];
-            tensor<fp32, [1, 3, 9, 256]> x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor<string, []>("linear_37")];
-            tensor<int32, [4]> var_1070_perm_0 = const()[name = tensor<string, []>("op_1070_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1074 = const()[name = tensor<string, []>("op_1074"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [1, 9, 3, 256]> var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
-            tensor<fp32, [9, 3, 256]> x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor<string, []>("x_29")];
+            tensor<fp32, [1, 3, 9, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[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, 3, 9, 512]> input_167 = concat(axis = var_62, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_167")];
+            tensor<fp32, [1, 3, 9, 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_1093_perm_0 = const()[name = tensor<string, []>("op_1093_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1097 = const()[name = tensor<string, []>("op_1097"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [1, 9, 3, 256]> var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
+            tensor<fp32, [9, 3, 256]> x_29 = reshape(shape = var_1097, x = var_1093)[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, 9, 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])];
@@ -943,132 +954,132 @@ 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, [9, 3, 256]> var_1082 = 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_1083 = const()[name = tensor<string, []>("op_1083"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor<string, []>("op_1084")];
+            tensor<fp32, [9, 3, 256]> var_1105 = 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_1106 = const()[name = tensor<string, []>("op_1106"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor<string, []>("op_1107")];
             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, [9, 3, 256]> var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_39")];
-            tensor<fp32, []> var_1089 = const()[name = tensor<string, []>("op_1089"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [9, 3, 256]> var_1090 = mul(x = var_1088, y = var_1089)[name = tensor<string, []>("op_1090")];
-            tensor<int32, [4]> var_1091 = const()[name = tensor<string, []>("op_1091"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor<string, []>("op_1092")];
+            tensor<fp32, [9, 3, 256]> var_1111 = 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_1112 = const()[name = tensor<string, []>("op_1112"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [9, 3, 256]> var_1113 = mul(x = var_1111, y = var_1112)[name = tensor<string, []>("op_1113")];
+            tensor<int32, [4]> var_1114 = const()[name = tensor<string, []>("op_1114"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor<string, []>("op_1115")];
             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, [9, 3, 256]> var_1096 = 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_1097 = const()[name = tensor<string, []>("op_1097"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor<string, []>("op_1098")];
+            tensor<fp32, [9, 3, 256]> var_1119 = 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_1120 = const()[name = tensor<string, []>("op_1120"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor<string, []>("op_1121")];
             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, [9, 3, 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, [9, 3, 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, [3]> cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor<string, []>("cumsum_mask_1")];
+            tensor<fp32, [3]> cumsum_mask_1 = cumsum(axis = var_59, 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, [3]> 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, [3]> clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
+            tensor<fp32, [3]> clip_1 = clip(alpha = var_49, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
             tensor<fp32, [3]> 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, [9, 4, 3, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor<string, []>("transpose_26")];
-            tensor<fp32, [9, 4, 3, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor<string, []>("transpose_27")];
+            tensor<fp32, [9, 4, 3, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor<string, []>("transpose_26")];
+            tensor<fp32, [9, 4, 3, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor<string, []>("transpose_27")];
             tensor<fp32, [9, 4, 3, 3]> 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_1110 = const()[name = tensor<string, []>("op_1110"), val = tensor<int32, [2]>([1, 3])];
-            tensor<fp32, [1, 3]> var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor<string, []>("op_1111")];
-            tensor<fp32, [3, 3]> causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor<string, []>("causal_with_valid_1")];
-            tensor<int32, [2]> var_1113 = const()[name = tensor<string, []>("op_1113"), val = tensor<int32, [2]>([3, 1])];
-            tensor<fp32, [3, 1]> var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor<string, []>("op_1114")];
-            tensor<fp32, [3, 3]> M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor<string, []>("M_9")];
-            tensor<fp32, [9, 4, 3, 3]> var_1116 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1116")];
+            tensor<int32, [2]> var_1133 = const()[name = tensor<string, []>("op_1133"), val = tensor<int32, [2]>([1, 3])];
+            tensor<fp32, [1, 3]> var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor<string, []>("op_1134")];
+            tensor<fp32, [3, 3]> causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor<string, []>("causal_with_valid_1")];
+            tensor<int32, [2]> var_1136 = const()[name = tensor<string, []>("op_1136"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor<string, []>("op_1137")];
+            tensor<fp32, [3, 3]> M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor<string, []>("M_9")];
+            tensor<fp32, [9, 4, 3, 3]> var_1139 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1139")];
             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, [9, 4, 3, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor<string, []>("transpose_25")];
-            tensor<fp32, [9, 4, 3, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor<string, []>("inner_9")];
-            tensor<bool, []> var_1118_transpose_x_0 = const()[name = tensor<string, []>("op_1118_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1118_transpose_y_0 = const()[name = tensor<string, []>("op_1118_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [9, 4, 3, 64]> var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1118")];
-            tensor<fp32, [3]> var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1119")];
-            tensor<int32, [4]> var_1120 = const()[name = tensor<string, []>("op_1120"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor<string, []>("op_1121")];
-            tensor<fp32, [9, 4, 3, 64]> cross_9 = mul(x = var_1118, y = var_1121)[name = tensor<string, []>("cross_9")];
+            tensor<fp32, [9, 4, 3, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor<string, []>("transpose_25")];
+            tensor<fp32, [9, 4, 3, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor<string, []>("inner_9")];
+            tensor<bool, []> var_1141_transpose_x_0 = const()[name = tensor<string, []>("op_1141_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1141_transpose_y_0 = const()[name = tensor<string, []>("op_1141_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [9, 4, 3, 64]> var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1141")];
+            tensor<fp32, [3]> var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1142")];
+            tensor<int32, [4]> var_1143 = const()[name = tensor<string, []>("op_1143"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor<string, []>("op_1144")];
+            tensor<fp32, [9, 4, 3, 64]> cross_9 = mul(x = var_1141, y = var_1144)[name = tensor<string, []>("cross_9")];
             tensor<fp32, [9, 4, 3, 64]> out_25 = add(x = inner_9, y = cross_9)[name = tensor<string, []>("out_25")];
-            tensor<int32, [4]> var_1124 = const()[name = tensor<string, []>("op_1124"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor<string, []>("op_1125")];
-            tensor<fp32, [9, 4, 3, 64]> v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor<string, []>("v_masked_1")];
-            tensor<fp32, [9, 4, 64, 64]> var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1127")];
-            tensor<bool, []> var_1129_transpose_x_1 = const()[name = tensor<string, []>("op_1129_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1129_transpose_y_1 = const()[name = tensor<string, []>("op_1129_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [9, 4, 64, 64]> var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1129")];
-            tensor<fp32, [9, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor<string, []>("new_kv_unnorm_9")];
-            tensor<bool, []> var_1131_keep_dims_0 = const()[name = tensor<string, []>("op_1131_keep_dims_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, []> var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1131")];
-            tensor<int32, [1]> var_1132 = const()[name = tensor<string, []>("op_1132"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [1]> var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor<string, []>("op_1133")];
-            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor<string, []>("new_scale_9")];
+            tensor<int32, [4]> var_1147 = const()[name = tensor<string, []>("op_1147"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor<string, []>("op_1148")];
+            tensor<fp32, [9, 4, 3, 64]> v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor<string, []>("v_masked_1")];
+            tensor<fp32, [9, 4, 64, 64]> var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1150")];
+            tensor<bool, []> var_1152_transpose_x_1 = const()[name = tensor<string, []>("op_1152_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1152_transpose_y_1 = const()[name = tensor<string, []>("op_1152_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [9, 4, 64, 64]> var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1152")];
+            tensor<fp32, [9, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor<string, []>("new_kv_unnorm_9")];
+            tensor<bool, []> var_1154_keep_dims_0 = const()[name = tensor<string, []>("op_1154_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, []> var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1154")];
+            tensor<int32, [1]> var_1155 = const()[name = tensor<string, []>("op_1155"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1]> var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor<string, []>("op_1156")];
+            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1156)[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_990, beta = const_21, x = new_scale_9)[name = tensor<string, []>("clip_2")];
+            tensor<fp32, [1]> clip_2 = clip(alpha = var_49, 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, [9, 4, 64, 64]> var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1137")];
-            tensor<int32, [4]> var_1138_perm_0 = const()[name = tensor<string, []>("op_1138_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [9, 4, 64, 64]> var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1160")];
+            tensor<int32, [4]> var_1161_perm_0 = const()[name = tensor<string, []>("op_1161_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, [9, 3, 4, 64]> var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
-            tensor<fp32, [9, 3, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor<string, []>("out_27")];
-            tensor<int32, [3]> var_1142 = const()[name = tensor<string, []>("op_1142"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [9, 3, 256]> out_29 = reshape(shape = var_1142, x = out_27)[name = tensor<string, []>("out_29")];
-            tensor<fp32, [9, 3, 256]> var_1144 = silu(x = input_169)[name = tensor<string, []>("op_1144")];
-            tensor<fp32, [9, 3, 256]> input_171 = mul(x = var_1144, y = out_29)[name = tensor<string, []>("input_171")];
-            tensor<fp32, [9, 3, 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, [9, 3, 256]> input_173 = add(x = x_29, y = ret_out_9)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [9, 3, 4, 64]> var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
+            tensor<fp32, [9, 3, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_64, x = var_1161)[name = tensor<string, []>("out_27")];
+            tensor<int32, [3]> var_1165 = const()[name = tensor<string, []>("op_1165"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [9, 3, 256]> out_29 = reshape(shape = var_1165, x = out_27)[name = tensor<string, []>("out_29")];
+            tensor<fp32, [9, 3, 256]> var_1167 = silu(x = input_171)[name = tensor<string, []>("op_1167")];
+            tensor<fp32, [9, 3, 256]> input_173 = mul(x = var_1167, y = out_29)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [9, 3, 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, [9, 3, 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, [9, 3, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor<string, []>("xt_1")];
-            tensor<int32, [4]> var_1154 = const()[name = tensor<string, []>("op_1154"), val = tensor<int32, [4]>([1, 9, 3, 256])];
-            tensor<fp32, [1, 9, 3, 256]> var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor<string, []>("op_1155")];
-            tensor<int32, [4]> var_1156_perm_0 = const()[name = tensor<string, []>("op_1156_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1159 = const()[name = tensor<string, []>("op_1159"), val = tensor<int32, [3]>([3, 9, 256])];
-            tensor<fp32, [1, 3, 9, 256]> var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor<string, []>("transpose_23")];
-            tensor<fp32, [3, 9, 256]> query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor<string, []>("query_1")];
+            tensor<fp32, [9, 3, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_56, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor<string, []>("xt_1")];
+            tensor<int32, [4]> var_1177 = const()[name = tensor<string, []>("op_1177"), val = tensor<int32, [4]>([1, 9, 3, 256])];
+            tensor<fp32, [1, 9, 3, 256]> var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor<string, []>("op_1178")];
+            tensor<int32, [4]> var_1179_perm_0 = const()[name = tensor<string, []>("op_1179_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1182 = const()[name = tensor<string, []>("op_1182"), val = tensor<int32, [3]>([3, 9, 256])];
+            tensor<fp32, [1, 3, 9, 256]> var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor<string, []>("transpose_23")];
+            tensor<fp32, [3, 9, 256]> query_1 = reshape(shape = var_1182, x = var_1179)[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, [9, 3, 256]> query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor<string, []>("transpose_22")];
-            tensor<fp32, [9, 3, 768]> var_1182 = 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, [9, 3, 768]> var_1205 = 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]>([9, 3, 3, 256])];
-            tensor<fp32, [9, 3, 3, 256]> var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor<string, []>("op_1184")];
-            tensor<int32, [1]> var_1185_axes_0 = const()[name = tensor<string, []>("op_1185_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 9, 3, 3, 256]> var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor<string, []>("op_1185")];
-            tensor<int32, [5]> var_1186_perm_0 = const()[name = tensor<string, []>("op_1186_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1187_axes_0 = const()[name = tensor<string, []>("op_1187_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 9, 3, 1, 256]> var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor<string, []>("transpose_21")];
-            tensor<fp32, [3, 9, 3, 256]> var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor<string, []>("op_1187")];
+            tensor<fp32, [9, 3, 3, 256]> var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor<string, []>("op_1207")];
+            tensor<int32, [1]> var_1208_axes_0 = const()[name = tensor<string, []>("op_1208_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 9, 3, 3, 256]> var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor<string, []>("op_1208")];
+            tensor<int32, [5]> var_1209_perm_0 = const()[name = tensor<string, []>("op_1209_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1210_axes_0 = const()[name = tensor<string, []>("op_1210_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 9, 3, 1, 256]> var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor<string, []>("transpose_21")];
+            tensor<fp32, [3, 9, 3, 256]> var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor<string, []>("op_1210")];
             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, 9, 3, 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, [9, 3, 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_1187)[name = tensor<string, []>("q_11")];
+            tensor<fp32, [9, 3, 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_1210)[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, 9, 3, 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, [9, 3, 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_1187)[name = tensor<string, []>("k_11")];
+            tensor<fp32, [9, 3, 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_1210)[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, 9, 3, 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, [9, 3, 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_1187)[name = tensor<string, []>("v_11")];
-            tensor<int32, [3]> var_1195 = const()[name = tensor<string, []>("op_1195"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor<string, []>("op_1196")];
+            tensor<fp32, [9, 3, 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_1210)[name = tensor<string, []>("v_11")];
+            tensor<int32, [3]> var_1218 = const()[name = tensor<string, []>("op_1218"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor<string, []>("op_1219")];
             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_1202 = const()[name = tensor<string, []>("op_1202"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor<string, []>("op_1203")];
+            tensor<int32, [3]> var_1225 = const()[name = tensor<string, []>("op_1225"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor<string, []>("op_1226")];
             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_1209 = const()[name = tensor<string, []>("op_1209"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor<string, []>("op_1210")];
+            tensor<int32, [3]> var_1232 = const()[name = tensor<string, []>("op_1232"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor<string, []>("op_1233")];
             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_1213 = const()[name = tensor<string, []>("op_1213"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor<string, []>("transpose_20")];
-            tensor<fp32, [3, 4, 9, 64]> q_15 = reshape(shape = var_1213, x = q_13)[name = tensor<string, []>("q_15")];
-            tensor<int32, [4]> var_1215 = const()[name = tensor<string, []>("op_1215"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor<string, []>("transpose_19")];
-            tensor<fp32, [3, 4, 9, 64]> k_15 = reshape(shape = var_1215, x = k_13)[name = tensor<string, []>("k_15")];
-            tensor<int32, [4]> var_1217 = const()[name = tensor<string, []>("op_1217"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor<string, []>("transpose_18")];
-            tensor<fp32, [3, 4, 9, 64]> v_15 = reshape(shape = var_1217, x = v_13)[name = tensor<string, []>("v_15")];
+            tensor<int32, [4]> var_1236 = const()[name = tensor<string, []>("op_1236"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor<string, []>("transpose_20")];
+            tensor<fp32, [3, 4, 9, 64]> q_15 = reshape(shape = var_1236, x = q_13)[name = tensor<string, []>("q_15")];
+            tensor<int32, [4]> var_1238 = const()[name = tensor<string, []>("op_1238"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor<string, []>("transpose_19")];
+            tensor<fp32, [3, 4, 9, 64]> k_15 = reshape(shape = var_1238, x = k_13)[name = tensor<string, []>("k_15")];
+            tensor<int32, [4]> var_1240 = const()[name = tensor<string, []>("op_1240"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor<string, []>("transpose_18")];
+            tensor<fp32, [3, 4, 9, 64]> v_15 = reshape(shape = var_1240, 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, [3, 4, 9, 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)];
@@ -1079,30 +1090,30 @@ 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, [3, 4, 9, 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_1220 = const()[name = tensor<string, []>("op_1220"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1225 = const()[name = tensor<string, []>("op_1225"), val = tensor<int32, [2]>([27, 256])];
-            tensor<fp32, [9, 3, 4, 64]> var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
-            tensor<fp32, [27, 256]> attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor<string, []>("attn_output_3")];
-            tensor<fp32, [27, 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_1229 = const()[name = tensor<string, []>("op_1229"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [9, 3, 256]> attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor<string, []>("attn_output_7")];
+            tensor<int32, [4]> var_1243 = const()[name = tensor<string, []>("op_1243"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1248 = const()[name = tensor<string, []>("op_1248"), val = tensor<int32, [2]>([27, 256])];
+            tensor<fp32, [9, 3, 4, 64]> var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
+            tensor<fp32, [27, 256]> attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor<string, []>("attn_output_3")];
+            tensor<fp32, [27, 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_1252 = const()[name = tensor<string, []>("op_1252"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [9, 3, 256]> attn_output_7 = reshape(shape = var_1252, 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, [3, 9, 256]> sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor<string, []>("transpose_16")];
-            tensor<fp32, [3, 9, 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, [3, 9, 256]> input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor<string, []>("input_177")];
-            tensor<fp32, [3, 9, 2048]> input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor<string, []>("linear_45")];
-            tensor<fp32, [3, 9, 2048]> input_181 = relu(x = input_179)[name = tensor<string, []>("input_181")];
-            tensor<fp32, [3, 9, 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, [3, 9, 256]> input_183 = add(x = input_177, y = ffn_out_1)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [3, 9, 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, [3, 9, 256]> input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_56, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor<string, []>("input_179")];
+            tensor<fp32, [3, 9, 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, [3, 9, 2048]> input_183 = relu(x = input_181)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [3, 9, 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, [3, 9, 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, [3, 9, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor<string, []>("xt_3")];
-            tensor<int32, [4]> var_1249 = const()[name = tensor<string, []>("op_1249"), val = tensor<int32, [4]>([1, 3, 9, 256])];
-            tensor<fp32, [1, 3, 9, 256]> x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor<string, []>("x_31")];
-            tensor<int32, [4]> var_1251_perm_0 = const()[name = tensor<string, []>("op_1251_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1255 = const()[name = tensor<string, []>("op_1255"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [1, 9, 3, 256]> var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
-            tensor<fp32, [9, 3, 256]> x = reshape(shape = var_1255, x = var_1251)[name = tensor<string, []>("x")];
+            tensor<fp32, [3, 9, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_56, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor<string, []>("xt_3")];
+            tensor<int32, [4]> var_1272 = const()[name = tensor<string, []>("op_1272"), val = tensor<int32, [4]>([1, 3, 9, 256])];
+            tensor<fp32, [1, 3, 9, 256]> x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor<string, []>("x_31")];
+            tensor<int32, [4]> var_1274_perm_0 = const()[name = tensor<string, []>("op_1274_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [1, 9, 3, 256]> var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
+            tensor<fp32, [9, 3, 256]> x = reshape(shape = var_1278, x = var_1274)[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, 9, 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])];
@@ -1113,120 +1124,120 @@ 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, [9, 3, 256]> var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor<string, []>("linear_47")];
-            tensor<int32, [4]> var_1264 = const()[name = tensor<string, []>("op_1264"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor<string, []>("op_1265")];
+            tensor<fp32, [9, 3, 256]> var_1286 = 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_1287 = const()[name = tensor<string, []>("op_1287"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor<string, []>("op_1288")];
             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, [9, 3, 256]> var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor<string, []>("linear_48")];
-            tensor<fp32, []> var_1270 = const()[name = tensor<string, []>("op_1270"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [9, 3, 256]> var_1271 = mul(x = var_1269, y = var_1270)[name = tensor<string, []>("op_1271")];
-            tensor<int32, [4]> var_1272 = const()[name = tensor<string, []>("op_1272"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor<string, []>("op_1273")];
+            tensor<fp32, [9, 3, 256]> var_1292 = 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_1293 = const()[name = tensor<string, []>("op_1293"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [9, 3, 256]> var_1294 = mul(x = var_1292, y = var_1293)[name = tensor<string, []>("op_1294")];
+            tensor<int32, [4]> var_1295 = const()[name = tensor<string, []>("op_1295"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor<string, []>("op_1296")];
             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, [9, 3, 256]> var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor<string, []>("linear_49")];
-            tensor<int32, [4]> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<int32, [4]>([9, 3, 4, 64])];
-            tensor<fp32, [9, 3, 4, 64]> var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor<string, []>("op_1279")];
+            tensor<fp32, [9, 3, 256]> var_1300 = 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_1301 = const()[name = tensor<string, []>("op_1301"), val = tensor<int32, [4]>([9, 3, 4, 64])];
+            tensor<fp32, [9, 3, 4, 64]> var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor<string, []>("op_1302")];
             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, [9, 3, 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, [9, 3, 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, [3]> 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, [3]> clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
+            tensor<fp32, [3]> clip_3 = clip(alpha = var_49, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
             tensor<fp32, [3]> 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, [9, 4, 3, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor<string, []>("transpose_13")];
-            tensor<fp32, [9, 4, 3, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor<string, []>("transpose_14")];
+            tensor<fp32, [9, 4, 3, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor<string, []>("transpose_13")];
+            tensor<fp32, [9, 4, 3, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor<string, []>("transpose_14")];
             tensor<fp32, [9, 4, 3, 3]> 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_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [2]>([3, 1])];
-            tensor<fp32, [3, 1]> var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor<string, []>("op_1295")];
-            tensor<fp32, [3, 3]> M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor<string, []>("M")];
-            tensor<fp32, [9, 4, 3, 3]> var_1297 = mul(x = qk, y = M)[name = tensor<string, []>("op_1297")];
-            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, [9, 4, 3, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor<string, []>("transpose_12")];
-            tensor<fp32, [9, 4, 3, 64]> inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor<string, []>("inner")];
-            tensor<bool, []> var_1299_transpose_x_0 = const()[name = tensor<string, []>("op_1299_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1299_transpose_y_0 = const()[name = tensor<string, []>("op_1299_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [9, 4, 3, 64]> var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1299")];
-            tensor<fp32, [3]> var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1300")];
-            tensor<int32, [4]> var_1301 = const()[name = tensor<string, []>("op_1301"), val = tensor<int32, [4]>([1, 1, 3, 1])];
-            tensor<fp32, [1, 1, 3, 1]> var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor<string, []>("op_1302")];
-            tensor<fp32, [9, 4, 3, 64]> cross = mul(x = var_1299, y = var_1302)[name = tensor<string, []>("cross")];
-            tensor<fp32, [9, 4, 3, 64]> out_31 = add(x = inner, y = cross)[name = tensor<string, []>("out_31")];
-            tensor<fp32, [9, 4, 3, 64]> v_masked = mul(x = v_17, y = var_1125)[name = tensor<string, []>("v_masked")];
-            tensor<fp32, [9, 4, 64, 64]> var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1308")];
-            tensor<bool, []> var_1310_transpose_x_1 = const()[name = tensor<string, []>("op_1310_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1310_transpose_y_1 = const()[name = tensor<string, []>("op_1310_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [9, 4, 64, 64]> var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1310")];
-            tensor<fp32, [9, 4, 64, 64]> new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor<string, []>("new_kv_unnorm")];
-            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1133)[name = tensor<string, []>("new_scale")];
+            tensor<int32, [2]> var_1317 = const()[name = tensor<string, []>("op_1317"), val = tensor<int32, [2]>([3, 1])];
+            tensor<fp32, [3, 1]> var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor<string, []>("op_1318")];
+            tensor<fp32, [3, 3]> M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor<string, []>("M")];
+            tensor<fp32, [9, 4, 3, 3]> var_1320 = mul(x = qk, y = M)[name = tensor<string, []>("op_1320")];
+            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, [9, 4, 3, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor<string, []>("transpose_12")];
+            tensor<fp32, [9, 4, 3, 64]> inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor<string, []>("inner_11")];
+            tensor<bool, []> var_1322_transpose_x_0 = const()[name = tensor<string, []>("op_1322_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1322_transpose_y_0 = const()[name = tensor<string, []>("op_1322_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [9, 4, 3, 64]> var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1322")];
+            tensor<fp32, [3]> var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1323")];
+            tensor<int32, [4]> var_1324 = const()[name = tensor<string, []>("op_1324"), val = tensor<int32, [4]>([1, 1, 3, 1])];
+            tensor<fp32, [1, 1, 3, 1]> var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor<string, []>("op_1325")];
+            tensor<fp32, [9, 4, 3, 64]> cross = mul(x = var_1322, y = var_1325)[name = tensor<string, []>("cross")];
+            tensor<fp32, [9, 4, 3, 64]> out_31 = add(x = inner_11, y = cross)[name = tensor<string, []>("out_31")];
+            tensor<fp32, [9, 4, 3, 64]> v_masked = mul(x = v_17, y = var_1148)[name = tensor<string, []>("v_masked")];
+            tensor<fp32, [9, 4, 64, 64]> var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1331")];
+            tensor<bool, []> var_1333_transpose_x_1 = const()[name = tensor<string, []>("op_1333_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1333_transpose_y_1 = const()[name = tensor<string, []>("op_1333_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [9, 4, 64, 64]> var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1333")];
+            tensor<fp32, [9, 4, 64, 64]> new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor<string, []>("new_kv_unnorm")];
+            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1156)[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_990, beta = const_33, x = new_scale)[name = tensor<string, []>("clip_4")];
+            tensor<fp32, [1]> clip_4 = clip(alpha = var_49, 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, [9, 4, 64, 64]> nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor<string, []>("nkv")];
-            tensor<int32, [4]> var_1319_perm_0 = const()[name = tensor<string, []>("op_1319_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_1342_perm_0 = const()[name = tensor<string, []>("op_1342_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, [9, 3, 4, 64]> var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
-            tensor<fp32, [9, 3, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor<string, []>("out_33")];
-            tensor<int32, [3]> var_1323 = const()[name = tensor<string, []>("op_1323"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [9, 3, 256]> out = reshape(shape = var_1323, x = out_33)[name = tensor<string, []>("out")];
-            tensor<fp32, [9, 3, 256]> var_1325 = silu(x = input_187)[name = tensor<string, []>("op_1325")];
-            tensor<fp32, [9, 3, 256]> input_189 = mul(x = var_1325, y = out)[name = tensor<string, []>("input_189")];
-            tensor<fp32, [9, 3, 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, [9, 3, 256]> input_191 = add(x = x, y = ret_out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [9, 3, 4, 64]> var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
+            tensor<fp32, [9, 3, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_64, x = var_1342)[name = tensor<string, []>("out_33")];
+            tensor<int32, [3]> var_1346 = const()[name = tensor<string, []>("op_1346"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [9, 3, 256]> out = reshape(shape = var_1346, x = out_33)[name = tensor<string, []>("out")];
+            tensor<fp32, [9, 3, 256]> var_1348 = silu(x = input_189)[name = tensor<string, []>("op_1348")];
+            tensor<fp32, [9, 3, 256]> input_191 = mul(x = var_1348, y = out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [9, 3, 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, [9, 3, 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, [9, 3, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor<string, []>("xt_5")];
-            tensor<int32, [4]> var_1335 = const()[name = tensor<string, []>("op_1335"), val = tensor<int32, [4]>([1, 9, 3, 256])];
-            tensor<fp32, [1, 9, 3, 256]> var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor<string, []>("op_1336")];
-            tensor<int32, [4]> var_1337_perm_0 = const()[name = tensor<string, []>("op_1337_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1340 = const()[name = tensor<string, []>("op_1340"), val = tensor<int32, [3]>([3, 9, 256])];
-            tensor<fp32, [1, 3, 9, 256]> var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor<string, []>("transpose_10")];
-            tensor<fp32, [3, 9, 256]> query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor<string, []>("query_5")];
+            tensor<fp32, [9, 3, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_56, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor<string, []>("xt_5")];
+            tensor<int32, [4]> var_1358 = const()[name = tensor<string, []>("op_1358"), val = tensor<int32, [4]>([1, 9, 3, 256])];
+            tensor<fp32, [1, 9, 3, 256]> var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor<string, []>("op_1359")];
+            tensor<int32, [4]> var_1360_perm_0 = const()[name = tensor<string, []>("op_1360_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1363 = const()[name = tensor<string, []>("op_1363"), val = tensor<int32, [3]>([3, 9, 256])];
+            tensor<fp32, [1, 3, 9, 256]> var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor<string, []>("transpose_10")];
+            tensor<fp32, [3, 9, 256]> query_5 = reshape(shape = var_1363, x = var_1360)[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, [9, 3, 256]> query = transpose(perm = query_perm_0, x = query_5)[name = tensor<string, []>("transpose_9")];
-            tensor<fp32, [9, 3, 768]> var_1363 = 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, [9, 3, 768]> var_1386 = 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]>([9, 3, 3, 256])];
-            tensor<fp32, [9, 3, 3, 256]> var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor<string, []>("op_1365")];
-            tensor<int32, [1]> var_1366_axes_0 = const()[name = tensor<string, []>("op_1366_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 9, 3, 3, 256]> var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor<string, []>("op_1366")];
-            tensor<int32, [5]> var_1367_perm_0 = const()[name = tensor<string, []>("op_1367_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1368_axes_0 = const()[name = tensor<string, []>("op_1368_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 9, 3, 1, 256]> var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor<string, []>("transpose_8")];
-            tensor<fp32, [3, 9, 3, 256]> var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor<string, []>("op_1368")];
+            tensor<fp32, [9, 3, 3, 256]> var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor<string, []>("op_1388")];
+            tensor<int32, [1]> var_1389_axes_0 = const()[name = tensor<string, []>("op_1389_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 9, 3, 3, 256]> var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor<string, []>("op_1389")];
+            tensor<int32, [5]> var_1390_perm_0 = const()[name = tensor<string, []>("op_1390_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1391_axes_0 = const()[name = tensor<string, []>("op_1391_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 9, 3, 1, 256]> var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor<string, []>("transpose_8")];
+            tensor<fp32, [3, 9, 3, 256]> var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor<string, []>("op_1391")];
             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, 9, 3, 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, [9, 3, 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_1368)[name = tensor<string, []>("q_19")];
+            tensor<fp32, [9, 3, 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_1391)[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, 9, 3, 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, [9, 3, 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_1368)[name = tensor<string, []>("k_19")];
+            tensor<fp32, [9, 3, 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_1391)[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, 9, 3, 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, [9, 3, 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_1368)[name = tensor<string, []>("v_19")];
-            tensor<int32, [3]> var_1376 = const()[name = tensor<string, []>("op_1376"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor<string, []>("op_1377")];
+            tensor<fp32, [9, 3, 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_1391)[name = tensor<string, []>("v_19")];
+            tensor<int32, [3]> var_1399 = const()[name = tensor<string, []>("op_1399"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor<string, []>("op_1400")];
             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_1383 = const()[name = tensor<string, []>("op_1383"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor<string, []>("op_1384")];
+            tensor<int32, [3]> var_1406 = const()[name = tensor<string, []>("op_1406"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor<string, []>("op_1407")];
             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_1390 = const()[name = tensor<string, []>("op_1390"), val = tensor<int32, [3]>([9, 12, 64])];
-            tensor<fp32, [9, 12, 64]> var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor<string, []>("op_1391")];
+            tensor<int32, [3]> var_1413 = const()[name = tensor<string, []>("op_1413"), val = tensor<int32, [3]>([9, 12, 64])];
+            tensor<fp32, [9, 12, 64]> var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor<string, []>("op_1414")];
             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_1394 = const()[name = tensor<string, []>("op_1394"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor<string, []>("transpose_7")];
-            tensor<fp32, [3, 4, 9, 64]> q = reshape(shape = var_1394, x = q_21)[name = tensor<string, []>("q")];
-            tensor<int32, [4]> var_1396 = const()[name = tensor<string, []>("op_1396"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor<string, []>("transpose_6")];
-            tensor<fp32, [3, 4, 9, 64]> k = reshape(shape = var_1396, x = k_21)[name = tensor<string, []>("k")];
-            tensor<int32, [4]> var_1398 = const()[name = tensor<string, []>("op_1398"), val = tensor<int32, [4]>([3, 4, 9, 64])];
-            tensor<fp32, [12, 9, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor<string, []>("transpose_5")];
-            tensor<fp32, [3, 4, 9, 64]> v = reshape(shape = var_1398, x = v_21)[name = tensor<string, []>("v")];
+            tensor<int32, [4]> var_1417 = const()[name = tensor<string, []>("op_1417"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor<string, []>("transpose_7")];
+            tensor<fp32, [3, 4, 9, 64]> q = reshape(shape = var_1417, x = q_21)[name = tensor<string, []>("q")];
+            tensor<int32, [4]> var_1419 = const()[name = tensor<string, []>("op_1419"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor<string, []>("transpose_6")];
+            tensor<fp32, [3, 4, 9, 64]> k = reshape(shape = var_1419, x = k_21)[name = tensor<string, []>("k")];
+            tensor<int32, [4]> var_1421 = const()[name = tensor<string, []>("op_1421"), val = tensor<int32, [4]>([3, 4, 9, 64])];
+            tensor<fp32, [12, 9, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor<string, []>("transpose_5")];
+            tensor<fp32, [3, 4, 9, 64]> v = reshape(shape = var_1421, 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, [3, 4, 9, 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)];
@@ -1237,36 +1248,36 @@ 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, [3, 4, 9, 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_1401 = const()[name = tensor<string, []>("op_1401"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1406 = const()[name = tensor<string, []>("op_1406"), val = tensor<int32, [2]>([27, 256])];
-            tensor<fp32, [9, 3, 4, 64]> var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
-            tensor<fp32, [27, 256]> attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor<string, []>("attn_output_11")];
-            tensor<fp32, [27, 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_1410 = const()[name = tensor<string, []>("op_1410"), val = tensor<int32, [3]>([9, 3, 256])];
-            tensor<fp32, [9, 3, 256]> attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor<string, []>("attn_output")];
+            tensor<int32, [4]> var_1424 = const()[name = tensor<string, []>("op_1424"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1429 = const()[name = tensor<string, []>("op_1429"), val = tensor<int32, [2]>([27, 256])];
+            tensor<fp32, [9, 3, 4, 64]> var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
+            tensor<fp32, [27, 256]> attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor<string, []>("attn_output_11")];
+            tensor<fp32, [27, 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_1433 = const()[name = tensor<string, []>("op_1433"), val = tensor<int32, [3]>([9, 3, 256])];
+            tensor<fp32, [9, 3, 256]> attn_output = reshape(shape = var_1433, 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, [3, 9, 256]> sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor<string, []>("transpose_3")];
-            tensor<fp32, [3, 9, 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, [3, 9, 256]> input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor<string, []>("input_195")];
-            tensor<fp32, [3, 9, 2048]> input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor<string, []>("linear_54")];
-            tensor<fp32, [3, 9, 2048]> input_199 = relu(x = input_197)[name = tensor<string, []>("input_199")];
-            tensor<fp32, [3, 9, 256]> ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor<string, []>("linear_55")];
-            tensor<fp32, [3, 9, 256]> input_201 = add(x = input_195, y = ffn_out)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [3, 9, 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, [3, 9, 256]> input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_56, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor<string, []>("input_197")];
+            tensor<fp32, [3, 9, 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, [3, 9, 2048]> input_201 = relu(x = input_199)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [3, 9, 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, [3, 9, 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, [3, 9, 256]> xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor<string, []>("xt")];
-            tensor<int32, [4]> var_1430 = const()[name = tensor<string, []>("op_1430"), val = tensor<int32, [4]>([1, 3, 9, 256])];
-            tensor<fp32, [1, 3, 9, 256]> input = reshape(shape = var_1430, x = xt)[name = tensor<string, []>("input")];
-            tensor<int32, [1]> var_1432 = const()[name = tensor<string, []>("op_1432"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 3, 9, 1]> var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor<string, []>("op_1433")];
+            tensor<fp32, [3, 9, 256]> xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_56, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor<string, []>("xt")];
+            tensor<int32, [4]> var_1453 = const()[name = tensor<string, []>("op_1453"), val = tensor<int32, [4]>([1, 3, 9, 256])];
+            tensor<fp32, [1, 3, 9, 256]> input = reshape(shape = var_1453, x = xt)[name = tensor<string, []>("input")];
+            tensor<int32, [1]> var_1455 = const()[name = tensor<string, []>("op_1455"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 3, 9, 1]> var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_55, x = input)[name = tensor<string, []>("op_1456")];
             tensor<fp32, []> const_42 = const()[name = tensor<string, []>("const_42"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 3, 9, 1]> clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor<string, []>("clip_5")];
-            tensor<fp32, [1, 3, 9, 256]> var_1435 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1435")];
+            tensor<fp32, [1, 3, 9, 1]> clip_5 = clip(alpha = var_69, beta = const_42, x = var_1456)[name = tensor<string, []>("clip_5")];
+            tensor<fp32, [1, 3, 9, 256]> var_1458 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1458")];
             tensor<int32, [3]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<int32, [3]>([3, 1, 256])];
             tensor<fp32, [3, 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]>([3, 256, 9])];
-            tensor<fp32, [1, 3, 256, 9]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 3, 256, 9]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor<string, []>("transpose_2")];
             tensor<fp32, [3, 256, 9]> 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)];
@@ -1277,10 +1288,10 @@ program(1.0)
             tensor<int32, [3]> output_end_0 = const()[name = tensor<string, []>("output_end_0"), val = tensor<int32, [3]>([1, 3, 8])];
             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, 3, 7]> 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, 3, 7]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1439")];
-            tensor<int32, []> var_1441_axis_0 = const()[name = tensor<string, []>("op_1441_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 9, 4, 64, 64]> dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor<string, []>("op_1441")];
-            tensor<int32, []> var_1443_axis_0 = const()[name = tensor<string, []>("op_1443_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1443")];
+            tensor<fp32, [1, 3, 7]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1462")];
+            tensor<int32, []> var_1464_axis_0 = const()[name = tensor<string, []>("op_1464_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 9, 4, 64, 64]> dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor<string, []>("op_1464")];
+            tensor<int32, []> var_1466_axis_0 = const()[name = tensor<string, []>("op_1466_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1466")];
         } -> (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