diff --git "a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil" "b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil"
--- "a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil"
+++ "b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil"
@@ -1,234 +1,256 @@
 program(1.0)
 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
 {
-    func main<ios17>(tensor<fp32, [1, 256, 18]> cnn_window, tensor<fp32, [2, 6, 4, 64, 64]> dec_kv, tensor<fp32, [2, 1]> dec_scale, tensor<fp32, [4, 1, 16, 256]> enc_conv_cache, tensor<fp32, [4, 1, 4, 64, 64]> enc_kv, tensor<fp32, [4, 1]> enc_scale, tensor<fp32, [1, 4, 345]> features, tensor<fp32, [4]> 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, [4, 4]> encoder__causal_mask = const()[name = tensor<string, []>("encoder__causal_mask"), val = tensor<fp32, [4, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4981952)))];
-            tensor<fp32, [4]> encoder__t_index = const()[name = tensor<string, []>("encoder__t_index"), val = tensor<fp32, [4]>([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])];
-            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, []>(4982080)))];
-            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, []>(4983168)))];
-            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, []>(5336512)))];
-            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, []>(5337600)))];
-            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, []>(5338688)))];
-            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, []>(5339776)))];
-            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, []>(5340864)))];
-            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, []>(5345024)))];
-            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, []>(6393664)))];
-            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, []>(6394752)))];
-            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, []>(7443392)))];
-            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, []>(7444480)))];
-            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, []>(7445568)))];
-            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, []>(7446656)))];
-            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, []>(7708864)))];
-            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, []>(7709952)))];
-            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, []>(7972160)))];
-            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, []>(7973248)))];
-            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, []>(8235456)))];
-            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, []>(8236544)))];
-            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, []>(8498752)))];
-            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, []>(8499840)))];
-            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, []>(8762048)))];
-            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, []>(8763136)))];
-            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, []>(8764224)))];
-            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, []>(8766336)))];
-            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, []>(9290688)))];
-            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, []>(9307136)))];
-            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, []>(9308224)))];
-            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, []>(9309312)))];
-            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, []>(9310400)))];
-            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, []>(9311488)))];
-            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, []>(9312576)))];
-            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, []>(9574784)))];
-            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, []>(9575872)))];
-            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, []>(9576960)))];
-            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, []>(9581120)))];
-            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, []>(10629760)))];
-            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, []>(10630848)))];
-            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, []>(11679488)))];
-            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, []>(11680576)))];
-            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, []>(11681664)))];
-            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, []>(11682752)))];
-            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, []>(11683840)))];
-            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, []>(11688000)))];
-            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, []>(12736640)))];
-            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, []>(12737728)))];
-            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, []>(13786368)))];
-            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, []>(13787456)))];
-            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, []>(13788544)))];
-            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, []>(13789632)))];
-            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, []>(14051840)))];
-            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, []>(14052928)))];
-            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, []>(14315136)))];
-            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, []>(14316224)))];
-            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, []>(14578432)))];
-            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, []>(14579520)))];
-            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, []>(14841728)))];
-            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, []>(14842816)))];
-            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, []>(15105024)))];
-            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, []>(15106112)))];
-            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, []>(15107200)))];
-            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, []>(15109312)))];
-            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, []>(15633664)))];
-            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, []>(15650112)))];
-            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, []>(15651200)))];
-            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, []>(15652288)))];
-            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, []>(15653376)))];
-            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, []>(15654464)))];
-            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, []>(15655552)))];
-            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, []>(15917760)))];
-            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, []>(15918848)))];
-            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, []>(15919936)))];
-            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, []>(15924096)))];
-            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, []>(16972736)))];
-            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, []>(16973824)))];
-            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, []>(18022464)))];
-            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, []>(18023552)))];
-            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, []>(18024640)))];
-            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, []>(18025728)))];
-            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, []>(18026816)))];
-            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, []>(18030976)))];
-            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, []>(19079616)))];
-            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, []>(19080704)))];
-            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, []>(20129344)))];
-            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, []>(20130432)))];
-            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, []>(20131520)))];
-            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, []>(20132608)))];
-            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, []>(20394816)))];
-            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, []>(20395904)))];
-            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, []>(20658112)))];
-            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, []>(20659200)))];
-            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, []>(20921408)))];
-            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, []>(20922496)))];
-            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, []>(21184704)))];
-            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, []>(21185792)))];
-            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, []>(21448000)))];
-            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, []>(21449088)))];
-            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, []>(21450176)))];
-            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, []>(21452288)))];
-            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, []>(21976640)))];
-            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, []>(21993088)))];
-            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, []>(21994176)))];
-            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, []>(21995264)))];
-            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, []>(21996352)))];
-            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, []>(21997440)))];
-            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, []>(21998528)))];
-            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, []>(22260736)))];
-            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, []>(22261824)))];
-            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, []>(22262912)))];
-            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, []>(22267072)))];
-            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, []>(23315712)))];
-            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, []>(23316800)))];
-            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, []>(24365440)))];
-            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, []>(24366528)))];
-            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, []>(24367616)))];
-            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, []>(24368704)))];
-            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, []>(24369792)))];
-            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, []>(24373952)))];
-            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, []>(25422592)))];
-            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, []>(25423680)))];
-            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, []>(26472320)))];
-            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, []>(26473408)))];
-            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, []>(26474496)))];
-            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, []>(26475584)))];
-            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, []>(26737792)))];
-            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, []>(26738880)))];
-            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, []>(27001088)))];
-            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, []>(27002176)))];
-            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, []>(27264384)))];
-            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, []>(27265472)))];
-            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, []>(27527680)))];
-            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, []>(27528768)))];
-            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, []>(27790976)))];
-            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, []>(27792064)))];
-            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, []>(27793152)))];
-            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, []>(27795264)))];
-            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, []>(28319616)))];
-            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, []>(28336064)))];
-            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, []>(28337152)))];
-            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, []>(28338240)))];
-            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, []>(28339328)))];
-            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, []>(28340416)))];
-            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, []>(28341504)))];
-            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, []>(28603712)))];
-            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, []>(28604800)))];
-            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, []>(28605888)))];
-            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, []>(28610048)))];
-            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, []>(29658688)))];
-            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, []>(29659776)))];
-            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, []>(30708416)))];
-            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, []>(30709504)))];
-            tensor<fp32, [4, 4]> decoder__causal_mask = const()[name = tensor<string, []>("decoder__causal_mask"), val = tensor<fp32, [4, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30710592)))];
-            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, []>(30710720)))];
-            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, []>(30711808)))];
-            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, []>(31236160)))];
-            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, []>(31237248)))];
-            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, []>(31499456)))];
-            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, []>(31500544)))];
-            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, []>(31762752)))];
-            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, []>(31763840)))];
-            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, []>(32026048)))];
-            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, []>(32027136)))];
-            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, []>(32289344)))];
-            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, []>(32290432)))];
-            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, []>(32552640)))];
-            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, []>(32553728)))];
-            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, []>(32554816)))];
-            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, []>(32555904)))];
-            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, []>(32818112)))];
-            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, []>(32821248)))];
-            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, []>(33607744)))];
-            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, []>(33608832)))];
-            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, []>(33609920)))];
-            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, []>(33618176)))];
-            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, []>(35715392)))];
-            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, []>(35716480)))];
-            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, []>(37813696)))];
-            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, []>(37814784)))];
-            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, []>(37815872)))];
-            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, []>(37816960)))];
-            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, []>(38079168)))];
-            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, []>(38080256)))];
-            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, []>(38342464)))];
-            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, []>(38343552)))];
-            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, []>(38605760)))];
-            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, []>(38606848)))];
-            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, []>(38869056)))];
-            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, []>(38870144)))];
-            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, []>(39132352)))];
-            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, []>(39133440)))];
-            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, []>(39134528)))];
-            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, []>(39135616)))];
-            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, []>(39397824)))];
-            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, []>(39400960)))];
-            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, []>(40187456)))];
-            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, []>(40188544)))];
-            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, []>(40189632)))];
-            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, []>(40197888)))];
-            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, []>(42295104)))];
-            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, []>(42296192)))];
-            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, []>(44393408)))];
-            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, []>(44394496)))];
-            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, 4, 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, 4, 256]> input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor<string, []>("input_3")];
+    func main<ios17>(tensor<fp32, [1, 256, 18]> cnn_window, tensor<fp32, [2, 6, 4, 64, 64]> dec_kv, tensor<fp32, [2, 1]> dec_scale, tensor<fp32, [4, 1, 16, 256]> enc_conv_cache, tensor<fp32, [4, 1, 4, 64, 64]> enc_kv, tensor<fp32, [4, 1]> enc_scale, tensor<fp32, [1, 45, 23]> features, tensor<fp32, [4]> 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, [4, 4]> inner_encoder__causal_mask = const()[name = tensor<string, []>("inner_encoder__causal_mask"), val = tensor<fp32, [4, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4981952)))];
+            tensor<fp32, [4]> inner_encoder__t_index = const()[name = tensor<string, []>("inner_encoder__t_index"), val = tensor<fp32, [4]>([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])];
+            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, []>(4982080)))];
+            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, []>(4983168)))];
+            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, []>(5336512)))];
+            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, []>(5337600)))];
+            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, []>(5338688)))];
+            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, []>(5339776)))];
+            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, []>(5340864)))];
+            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, []>(5345024)))];
+            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, []>(6393664)))];
+            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, []>(6394752)))];
+            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, []>(7443392)))];
+            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, []>(7444480)))];
+            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, []>(7445568)))];
+            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, []>(7446656)))];
+            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, []>(7708864)))];
+            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, []>(7709952)))];
+            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, []>(7972160)))];
+            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, []>(7973248)))];
+            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, []>(8235456)))];
+            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, []>(8236544)))];
+            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, []>(8498752)))];
+            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, []>(8499840)))];
+            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, []>(8762048)))];
+            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, []>(8763136)))];
+            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, []>(8764224)))];
+            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, []>(8766336)))];
+            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, []>(9290688)))];
+            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, []>(9307136)))];
+            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, []>(9308224)))];
+            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, []>(9309312)))];
+            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, []>(9310400)))];
+            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, []>(9311488)))];
+            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, []>(9312576)))];
+            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, []>(9574784)))];
+            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, []>(9575872)))];
+            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, []>(9576960)))];
+            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, []>(9581120)))];
+            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, []>(10629760)))];
+            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, []>(10630848)))];
+            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, []>(11679488)))];
+            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, []>(11680576)))];
+            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, []>(11681664)))];
+            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, []>(11682752)))];
+            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, []>(11683840)))];
+            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, []>(11688000)))];
+            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, []>(12736640)))];
+            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, []>(12737728)))];
+            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, []>(13786368)))];
+            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, []>(13787456)))];
+            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, []>(13788544)))];
+            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, []>(13789632)))];
+            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, []>(14051840)))];
+            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, []>(14052928)))];
+            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, []>(14315136)))];
+            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, []>(14316224)))];
+            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, []>(14578432)))];
+            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, []>(14579520)))];
+            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, []>(14841728)))];
+            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, []>(14842816)))];
+            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, []>(15105024)))];
+            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, []>(15106112)))];
+            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, []>(15107200)))];
+            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, []>(15109312)))];
+            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, []>(15633664)))];
+            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, []>(15650112)))];
+            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, []>(15651200)))];
+            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, []>(15652288)))];
+            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, []>(15653376)))];
+            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, []>(15654464)))];
+            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, []>(15655552)))];
+            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, []>(15917760)))];
+            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, []>(15918848)))];
+            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, []>(15919936)))];
+            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, []>(15924096)))];
+            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, []>(16972736)))];
+            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, []>(16973824)))];
+            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, []>(18022464)))];
+            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, []>(18023552)))];
+            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, []>(18024640)))];
+            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, []>(18025728)))];
+            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, []>(18026816)))];
+            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, []>(18030976)))];
+            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, []>(19079616)))];
+            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, []>(19080704)))];
+            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, []>(20129344)))];
+            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, []>(20130432)))];
+            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, []>(20131520)))];
+            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, []>(20132608)))];
+            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, []>(20394816)))];
+            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, []>(20395904)))];
+            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, []>(20658112)))];
+            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, []>(20659200)))];
+            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, []>(20921408)))];
+            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, []>(20922496)))];
+            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, []>(21184704)))];
+            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, []>(21185792)))];
+            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, []>(21448000)))];
+            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, []>(21449088)))];
+            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, []>(21450176)))];
+            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, []>(21452288)))];
+            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, []>(21976640)))];
+            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, []>(21993088)))];
+            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, []>(21994176)))];
+            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, []>(21995264)))];
+            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, []>(21996352)))];
+            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, []>(21997440)))];
+            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, []>(21998528)))];
+            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, []>(22260736)))];
+            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, []>(22261824)))];
+            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, []>(22262912)))];
+            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, []>(22267072)))];
+            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, []>(23315712)))];
+            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, []>(23316800)))];
+            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, []>(24365440)))];
+            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, []>(24366528)))];
+            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, []>(24367616)))];
+            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, []>(24368704)))];
+            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, []>(24369792)))];
+            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, []>(24373952)))];
+            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, []>(25422592)))];
+            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, []>(25423680)))];
+            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, []>(26472320)))];
+            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, []>(26473408)))];
+            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, []>(26474496)))];
+            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, []>(26475584)))];
+            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, []>(26737792)))];
+            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, []>(26738880)))];
+            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, []>(27001088)))];
+            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, []>(27002176)))];
+            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, []>(27264384)))];
+            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, []>(27265472)))];
+            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, []>(27527680)))];
+            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, []>(27528768)))];
+            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, []>(27790976)))];
+            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, []>(27792064)))];
+            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, []>(27793152)))];
+            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, []>(27795264)))];
+            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, []>(28319616)))];
+            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, []>(28336064)))];
+            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, []>(28337152)))];
+            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, []>(28338240)))];
+            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, []>(28339328)))];
+            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, []>(28340416)))];
+            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, []>(28341504)))];
+            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, []>(28603712)))];
+            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, []>(28604800)))];
+            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, []>(28605888)))];
+            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, []>(28610048)))];
+            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, []>(29658688)))];
+            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, []>(29659776)))];
+            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, []>(30708416)))];
+            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, []>(30709504)))];
+            tensor<fp32, [4, 4]> inner_decoder__causal_mask = const()[name = tensor<string, []>("inner_decoder__causal_mask"), val = tensor<fp32, [4, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30710592)))];
+            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, []>(30710720)))];
+            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, []>(30711808)))];
+            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, []>(31236160)))];
+            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, []>(31237248)))];
+            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, []>(31499456)))];
+            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, []>(31500544)))];
+            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, []>(31762752)))];
+            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, []>(31763840)))];
+            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, []>(32026048)))];
+            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, []>(32027136)))];
+            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, []>(32289344)))];
+            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, []>(32290432)))];
+            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, []>(32552640)))];
+            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, []>(32553728)))];
+            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, []>(32554816)))];
+            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, []>(32555904)))];
+            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, []>(32818112)))];
+            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, []>(32821248)))];
+            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, []>(33607744)))];
+            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, []>(33608832)))];
+            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, []>(33609920)))];
+            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, []>(33618176)))];
+            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, []>(35715392)))];
+            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, []>(35716480)))];
+            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, []>(37813696)))];
+            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, []>(37814784)))];
+            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, []>(37815872)))];
+            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, []>(37816960)))];
+            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, []>(38079168)))];
+            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, []>(38080256)))];
+            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, []>(38342464)))];
+            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, []>(38343552)))];
+            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, []>(38605760)))];
+            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, []>(38606848)))];
+            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, []>(38869056)))];
+            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, []>(38870144)))];
+            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, []>(39132352)))];
+            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, []>(39133440)))];
+            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, []>(39134528)))];
+            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, []>(39135616)))];
+            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, []>(39397824)))];
+            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, []>(39400960)))];
+            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, []>(40187456)))];
+            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, []>(40188544)))];
+            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, []>(40189632)))];
+            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, []>(40197888)))];
+            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, []>(42295104)))];
+            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, []>(42296192)))];
+            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, []>(44393408)))];
+            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, []>(44394496)))];
+            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, 35, 23])];
+            tensor<bool, [3]> var_39_end_mask_0 = const()[name = tensor<string, []>("op_39_end_mask_0"), val = tensor<bool, [3]>([true, false, 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, [3]> var_49_begin_0 = const()[name = tensor<string, []>("op_49_begin_0"), val = tensor<int32, [3]>([0, 30, 0])];
+            tensor<int32, [3]> var_49_end_0 = const()[name = tensor<string, []>("op_49_end_0"), val = tensor<int32, [3]>([1, 1, 23])];
+            tensor<bool, [3]> var_49_end_mask_0 = const()[name = tensor<string, []>("op_49_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 23]> var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor<string, []>("op_49")];
+            tensor<int32, []> stacked_axis_0 = const()[name = tensor<string, []>("stacked_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 4, 15, 23]> stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor<string, []>("stacked")];
+            tensor<int32, [3]> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, [3]>([1, 4, 345])];
+            tensor<fp32, [1, 4, 345]> input_1 = reshape(shape = var_56, x = stacked)[name = tensor<string, []>("input_1")];
+            tensor<fp32, []> var_58 = const()[name = tensor<string, []>("op_58"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<bool, []> var_64 = const()[name = tensor<string, []>("op_64"), val = tensor<bool, []>(true)];
+            tensor<fp32, []> var_65 = const()[name = tensor<string, []>("op_65"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
+            tensor<int32, []> var_68 = const()[name = tensor<string, []>("op_68"), val = tensor<int32, []>(0)];
+            tensor<int32, []> var_70 = const()[name = tensor<string, []>("op_70"), val = tensor<int32, []>(2)];
+            tensor<int32, []> var_71 = const()[name = tensor<string, []>("op_71"), val = tensor<int32, []>(-1)];
+            tensor<fp32, []> var_73 = const()[name = tensor<string, []>("op_73"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, []> var_78 = const()[name = tensor<string, []>("op_78"), val = tensor<fp32, []>(0x1.5798eep-27)];
+            tensor<int32, []> var_82 = const()[name = tensor<string, []>("op_82"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 4, 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, 4, 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, 4, 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, 4, 1024]> input_7 = silu(x = inputs_1)[name = tensor<string, []>("input_7")];
-            tensor<fp32, [1, 4, 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, 4, 256]> var_148 = mul(x = input_11, y = var_147)[name = tensor<string, []>("op_148")];
-            tensor<fp32, [1, 4, 256]> input_13 = add(x = var_148, y = input_3)[name = tensor<string, []>("input_13")];
+            tensor<fp32, [1, 4, 256]> input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_65, 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, 4, 256]> input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_9 = silu(x = inputs_1)[name = tensor<string, []>("input_9")];
+            tensor<fp32, [1, 4, 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_203 = const()[name = tensor<string, []>("op_203"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_204 = mul(x = input_13, y = var_203)[name = tensor<string, []>("op_204")];
+            tensor<fp32, [1, 4, 256]> input_15 = add(x = var_204, 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, 4, 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, 4, 256]> x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_65, 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])];
@@ -239,173 +261,173 @@ 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, 4, 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, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_164 = reshape(shape = var_163, x = var_162)[name = tensor<string, []>("op_164")];
+            tensor<fp32, [1, 4, 256]> var_218 = 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_219 = const()[name = tensor<string, []>("op_219"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_220 = reshape(shape = var_219, x = var_218)[name = tensor<string, []>("op_220")];
             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, 4, 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, 4, 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, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_172 = reshape(shape = var_171, x = var_170)[name = tensor<string, []>("op_172")];
+            tensor<fp32, [1, 4, 256]> var_224 = 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_225 = const()[name = tensor<string, []>("op_225"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 4, 256]> var_226 = mul(x = var_224, y = var_225)[name = tensor<string, []>("op_226")];
+            tensor<int32, [4]> var_227 = const()[name = tensor<string, []>("op_227"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_228 = reshape(shape = var_227, x = var_226)[name = tensor<string, []>("op_228")];
             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, 4, 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, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_178 = reshape(shape = var_177, x = var_176)[name = tensor<string, []>("op_178")];
+            tensor<fp32, [1, 4, 256]> var_232 = 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_233 = const()[name = tensor<string, []>("op_233"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_234 = reshape(shape = var_233, x = var_232)[name = tensor<string, []>("op_234")];
             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, 4, 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, 4, 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, [4]> s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor<string, []>("s_t_1")];
+            tensor<fp32, [4]> s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_1")];
             tensor<fp32, [4]> 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, 4, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor<string, []>("transpose_57")];
-            tensor<fp32, [1, 4, 4, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor<string, []>("transpose_58")];
+            tensor<fp32, [1, 4, 4, 64]> k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor<string, []>("transpose_57")];
+            tensor<fp32, [1, 4, 4, 64]> q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor<string, []>("transpose_58")];
             tensor<fp32, [1, 4, 4, 4]> 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]>([4, 1])];
-            tensor<fp32, [4, 1]> var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor<string, []>("op_189")];
-            tensor<fp32, [4, 4]> M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor<string, []>("M_1")];
-            tensor<fp32, [1, 4, 4, 4]> var_191 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_191")];
+            tensor<int32, [2]> var_244 = const()[name = tensor<string, []>("op_244"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor<string, []>("op_245")];
+            tensor<fp32, [4, 4]> M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor<string, []>("M_1")];
+            tensor<fp32, [1, 4, 4, 4]> var_247 = mul(x = qk_1, y = M_1)[name = tensor<string, []>("op_247")];
             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, 4, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor<string, []>("transpose_56")];
-            tensor<fp32, [1, 4, 4, 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, 4, 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, [4]> 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, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_196 = reshape(shape = var_195, x = var_194)[name = tensor<string, []>("op_196")];
-            tensor<fp32, [1, 4, 4, 64]> cross_1 = mul(x = var_193, y = var_196)[name = tensor<string, []>("cross_1")];
+            tensor<fp32, [1, 4, 4, 64]> v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor<string, []>("transpose_56")];
+            tensor<fp32, [1, 4, 4, 64]> inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_247, y = v_1)[name = tensor<string, []>("inner_1")];
+            tensor<bool, []> var_249_transpose_x_0 = const()[name = tensor<string, []>("op_249_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_249_transpose_y_0 = const()[name = tensor<string, []>("op_249_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 4, 64]> var_249 = matmul(transpose_x = var_249_transpose_x_0, transpose_y = var_249_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor<string, []>("op_249")];
+            tensor<fp32, [4]> var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor<string, []>("op_250")];
+            tensor<int32, [4]> var_251 = const()[name = tensor<string, []>("op_251"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_252 = reshape(shape = var_251, x = var_250)[name = tensor<string, []>("op_252")];
+            tensor<fp32, [1, 4, 4, 64]> cross_1 = mul(x = var_249, y = var_252)[name = tensor<string, []>("cross_1")];
             tensor<fp32, [1, 4, 4, 64]> out_1 = add(x = inner_1, y = cross_1)[name = tensor<string, []>("out_1")];
-            tensor<fp32, [1, 4, 64, 64]> var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor<string, []>("op_199")];
-            tensor<bool, []> var_201_transpose_x_1 = const()[name = tensor<string, []>("op_201_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_201_transpose_y_1 = const()[name = tensor<string, []>("op_201_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor<string, []>("op_201")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor<string, []>("new_kv_unnorm_1")];
-            tensor<fp32, []> var_203 = const()[name = tensor<string, []>("op_203"), val = tensor<fp32, []>(0x1p+2)];
-            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_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor<string, []>("op_255")];
+            tensor<bool, []> var_257_transpose_x_1 = const()[name = tensor<string, []>("op_257_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_257_transpose_y_1 = const()[name = tensor<string, []>("op_257_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_257 = matmul(transpose_x = var_257_transpose_x_1, transpose_y = var_257_transpose_y_1, x = k_1, y = v_1)[name = tensor<string, []>("op_257")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor<string, []>("new_kv_unnorm_1")];
+            tensor<fp32, []> var_259 = const()[name = tensor<string, []>("op_259"), val = tensor<fp32, []>(0x1p+2)];
+            tensor<fp32, [1]> new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor<string, []>("new_scale_1")];
+            tensor<fp32, [1]> var_261 = sqrt(x = new_scale_1)[name = tensor<string, []>("op_261")];
+            tensor<fp32, [1, 4, 64, 64]> var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor<string, []>("op_262")];
+            tensor<int32, [4]> var_263_perm_0 = const()[name = tensor<string, []>("op_263_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, 4, 4, 64]> var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
-            tensor<fp32, [1, 4, 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, 4, 256])];
-            tensor<fp32, [1, 4, 256]> out_5 = reshape(shape = var_211, x = out_3)[name = tensor<string, []>("out_5")];
-            tensor<fp32, [1, 4, 256]> var_213 = silu(x = input_17)[name = tensor<string, []>("op_213")];
-            tensor<fp32, [1, 4, 256]> input_19 = mul(x = var_213, y = out_5)[name = tensor<string, []>("input_19")];
-            tensor<fp32, [1, 4, 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, 4, 256]> x_3 = add(x = input_13, y = ret_out_1)[name = tensor<string, []>("x_3")];
+            tensor<fp32, [1, 4, 4, 64]> var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor<string, []>("transpose_55")];
+            tensor<fp32, [1, 4, 4, 64]> out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_73, x = var_263)[name = tensor<string, []>("out_3")];
+            tensor<int32, [3]> var_267 = const()[name = tensor<string, []>("op_267"), val = tensor<int32, [3]>([1, 4, 256])];
+            tensor<fp32, [1, 4, 256]> out_5 = reshape(shape = var_267, x = out_3)[name = tensor<string, []>("out_5")];
+            tensor<fp32, [1, 4, 256]> var_269 = silu(x = input_19)[name = tensor<string, []>("op_269")];
+            tensor<fp32, [1, 4, 256]> input_21 = mul(x = var_269, y = out_5)[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 4, 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, 4, 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_277_begin_0 = const()[name = tensor<string, []>("op_277_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_277_end_0 = const()[name = tensor<string, []>("op_277_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_277_end_mask_0 = const()[name = tensor<string, []>("op_277_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_277 = slice_by_index(begin = var_277_begin_0, end = var_277_end_0, end_mask = var_277_end_mask_0, x = x_3)[name = tensor<string, []>("op_277")];
+            tensor<int32, [3]> var_280_begin_0 = const()[name = tensor<string, []>("op_280_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_280_end_0 = const()[name = tensor<string, []>("op_280_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_280_end_mask_0 = const()[name = tensor<string, []>("op_280_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_280 = slice_by_index(begin = var_280_begin_0, end = var_280_end_0, end_mask = var_280_end_mask_0, x = window_1)[name = tensor<string, []>("op_280")];
             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_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor<string, []>("window_3")];
+            tensor<int32, [3]> var_285_begin_0 = const()[name = tensor<string, []>("op_285_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_285_end_0 = const()[name = tensor<string, []>("op_285_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_285_end_mask_0 = const()[name = tensor<string, []>("op_285_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_285 = slice_by_index(begin = var_285_begin_0, end = var_285_end_0, end_mask = var_285_end_mask_0, x = x_3)[name = tensor<string, []>("op_285")];
+            tensor<int32, [3]> var_288_begin_0 = const()[name = tensor<string, []>("op_288_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_288_end_0 = const()[name = tensor<string, []>("op_288_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_288_end_mask_0 = const()[name = tensor<string, []>("op_288_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = window_3)[name = tensor<string, []>("op_288")];
             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, 3, 256])];
-            tensor<bool, [3]> var_237_end_mask_0 = const()[name = tensor<string, []>("op_237_end_mask_0"), val = tensor<bool, [3]>([true, false, 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_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor<string, []>("window_5")];
+            tensor<int32, [3]> var_293_begin_0 = const()[name = tensor<string, []>("op_293_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_293_end_0 = const()[name = tensor<string, []>("op_293_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<bool, [3]> var_293_end_mask_0 = const()[name = tensor<string, []>("op_293_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_293 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, x = x_3)[name = tensor<string, []>("op_293")];
+            tensor<int32, [3]> var_296_begin_0 = const()[name = tensor<string, []>("op_296_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_296_end_0 = const()[name = tensor<string, []>("op_296_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_296_end_mask_0 = const()[name = tensor<string, []>("op_296_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = window_5)[name = tensor<string, []>("op_296")];
             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<int32, [3]> var_245_begin_0 = const()[name = tensor<string, []>("op_245_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
-            tensor<int32, [3]> var_245_end_0 = const()[name = tensor<string, []>("op_245_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_245_end_mask_0 = const()[name = tensor<string, []>("op_245_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor<string, []>("op_245")];
-            tensor<int32, [3]> var_248_begin_0 = const()[name = tensor<string, []>("op_248_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_248_end_0 = const()[name = tensor<string, []>("op_248_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_248_end_mask_0 = const()[name = tensor<string, []>("op_248_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor<string, []>("op_248")];
+            tensor<fp32, [1, 16, 256]> window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor<string, []>("window_7")];
+            tensor<int32, [3]> var_301_begin_0 = const()[name = tensor<string, []>("op_301_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
+            tensor<int32, [3]> var_301_end_0 = const()[name = tensor<string, []>("op_301_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_301_end_mask_0 = const()[name = tensor<string, []>("op_301_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_301 = slice_by_index(begin = var_301_begin_0, end = var_301_end_0, end_mask = var_301_end_mask_0, x = x_3)[name = tensor<string, []>("op_301")];
+            tensor<int32, [3]> var_304_begin_0 = const()[name = tensor<string, []>("op_304_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_304_end_0 = const()[name = tensor<string, []>("op_304_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_304_end_mask_0 = const()[name = tensor<string, []>("op_304_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = window_7)[name = tensor<string, []>("op_304")];
             tensor<bool, []> window_9_interleave_0 = const()[name = tensor<string, []>("window_9_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor<string, []>("window_9")];
-            tensor<bool, []> input_21_interleave_0 = const()[name = tensor<string, []>("input_21_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [4, 16, 256]> input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 16, 256]> window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor<string, []>("window_9")];
+            tensor<bool, []> input_23_interleave_0 = const()[name = tensor<string, []>("input_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [4, 16, 256]> input_23 = concat(axis = var_68, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[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, [4, 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, [4, 16, 256]> x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_65, 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, [4, 256, 16]> input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
-            tensor<fp32, [4, 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_273_split_sizes_0 = const()[name = tensor<string, []>("op_273_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_273_axis_0 = const()[name = tensor<string, []>("op_273_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [4, 256, 16]> var_273_0, tensor<fp32, [4, 256, 16]> var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_273")];
-            tensor<fp32, [4, 256, 16]> var_275 = sigmoid(x = var_273_1)[name = tensor<string, []>("op_275")];
-            tensor<fp32, [4, 256, 16]> inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor<string, []>("inputs_5")];
+            tensor<fp32, [4, 256, 16]> input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor<string, []>("transpose_54")];
+            tensor<fp32, [4, 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_329_split_sizes_0 = const()[name = tensor<string, []>("op_329_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_329_axis_0 = const()[name = tensor<string, []>("op_329_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [4, 256, 16]> var_329_0, tensor<fp32, [4, 256, 16]> var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor<string, []>("op_329")];
+            tensor<fp32, [4, 256, 16]> var_331 = sigmoid(x = var_329_1)[name = tensor<string, []>("op_331")];
+            tensor<fp32, [4, 256, 16]> inputs_5 = mul(x = var_329_0, y = var_331)[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, [4, 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]>([4, 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, [4, 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, [4, 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, [4, 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, [4, 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, [4, 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]>([4, 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, [4, 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, [4, 256, 16]> inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_65, 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, [4, 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, [4, 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_306_begin_0 = const()[name = tensor<string, []>("op_306_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_306_end_0 = const()[name = tensor<string, []>("op_306_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
-            tensor<bool, [3]> var_306_end_mask_0 = const()[name = tensor<string, []>("op_306_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [4, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor<string, []>("transpose_53")];
-            tensor<fp32, [4, 1, 256]> var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_306")];
-            tensor<int32, [3]> var_308_perm_0 = const()[name = tensor<string, []>("op_308_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 4, 256]> var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor<string, []>("transpose_52")];
-            tensor<fp32, [1, 4, 256]> input_31 = add(x = x_3, y = var_308)[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, 4, 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, 4, 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, 4, 1024]> input_35 = silu(x = inputs_9)[name = tensor<string, []>("input_35")];
-            tensor<fp32, [1, 4, 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_331 = const()[name = tensor<string, []>("op_331"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_332 = mul(x = input_39, y = var_331)[name = tensor<string, []>("op_332")];
-            tensor<fp32, [1, 4, 256]> input_41 = add(x = var_332, 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, 4, 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_362_begin_0 = const()[name = tensor<string, []>("op_362_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_362_end_0 = const()[name = tensor<string, []>("op_362_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
+            tensor<bool, [3]> var_362_end_mask_0 = const()[name = tensor<string, []>("op_362_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [4, 16, 256]> conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor<string, []>("transpose_53")];
+            tensor<fp32, [4, 1, 256]> var_362 = slice_by_index(begin = var_362_begin_0, end = var_362_end_0, end_mask = var_362_end_mask_0, x = conv_out_1)[name = tensor<string, []>("op_362")];
+            tensor<int32, [3]> var_364_perm_0 = const()[name = tensor<string, []>("op_364_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 4, 256]> var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor<string, []>("transpose_52")];
+            tensor<fp32, [1, 4, 256]> input_33 = add(x = x_3, y = var_364)[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, 4, 256]> input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor<string, []>("input_35")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_37 = silu(x = inputs_9)[name = tensor<string, []>("input_37")];
+            tensor<fp32, [1, 4, 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_387 = const()[name = tensor<string, []>("op_387"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_388 = mul(x = input_41, y = var_387)[name = tensor<string, []>("op_388")];
+            tensor<fp32, [1, 4, 256]> input_43 = add(x = var_388, 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, 4, 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, 4, 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, 4, 1024]> input_47 = silu(x = inputs_11)[name = tensor<string, []>("input_47")];
-            tensor<fp32, [1, 4, 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_361 = const()[name = tensor<string, []>("op_361"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_362 = mul(x = input_51, y = var_361)[name = tensor<string, []>("op_362")];
-            tensor<fp32, [1, 4, 256]> input_53 = add(x = var_362, y = input_43)[name = tensor<string, []>("input_53")];
+            tensor<fp32, [1, 4, 256]> input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_65, 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, 4, 256]> input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor<string, []>("input_47")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_49 = silu(x = inputs_11)[name = tensor<string, []>("input_49")];
+            tensor<fp32, [1, 4, 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_417 = const()[name = tensor<string, []>("op_417"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_418 = mul(x = input_53, y = var_417)[name = tensor<string, []>("op_418")];
+            tensor<fp32, [1, 4, 256]> input_55 = add(x = var_418, 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, 4, 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, 4, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_65, 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])];
@@ -416,173 +438,173 @@ 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, 4, 256]> var_376 = 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_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_378 = reshape(shape = var_377, x = var_376)[name = tensor<string, []>("op_378")];
+            tensor<fp32, [1, 4, 256]> var_432 = 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_433 = const()[name = tensor<string, []>("op_433"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_434 = reshape(shape = var_433, x = var_432)[name = tensor<string, []>("op_434")];
             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, 4, 256]> var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor<string, []>("linear_13")];
-            tensor<fp32, []> var_383 = const()[name = tensor<string, []>("op_383"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 4, 256]> var_384 = mul(x = var_382, y = var_383)[name = tensor<string, []>("op_384")];
-            tensor<int32, [4]> var_385 = const()[name = tensor<string, []>("op_385"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_386 = reshape(shape = var_385, x = var_384)[name = tensor<string, []>("op_386")];
+            tensor<fp32, [1, 4, 256]> var_438 = 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_439 = const()[name = tensor<string, []>("op_439"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 4, 256]> var_440 = mul(x = var_438, y = var_439)[name = tensor<string, []>("op_440")];
+            tensor<int32, [4]> var_441 = const()[name = tensor<string, []>("op_441"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_442 = reshape(shape = var_441, x = var_440)[name = tensor<string, []>("op_442")];
             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, 4, 256]> var_390 = 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_391 = const()[name = tensor<string, []>("op_391"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_392 = reshape(shape = var_391, x = var_390)[name = tensor<string, []>("op_392")];
+            tensor<fp32, [1, 4, 256]> var_446 = 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_447 = const()[name = tensor<string, []>("op_447"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_448 = reshape(shape = var_447, x = var_446)[name = tensor<string, []>("op_448")];
             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, 4, 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, 4, 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, [4]> s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor<string, []>("s_t_3")];
+            tensor<fp32, [4]> s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_3")];
             tensor<fp32, [4]> 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, 4, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor<string, []>("transpose_50")];
-            tensor<fp32, [1, 4, 4, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor<string, []>("transpose_51")];
+            tensor<fp32, [1, 4, 4, 64]> k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor<string, []>("transpose_50")];
+            tensor<fp32, [1, 4, 4, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor<string, []>("transpose_51")];
             tensor<fp32, [1, 4, 4, 4]> 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_402 = const()[name = tensor<string, []>("op_402"), val = tensor<int32, [2]>([4, 1])];
-            tensor<fp32, [4, 1]> var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor<string, []>("op_403")];
-            tensor<fp32, [4, 4]> M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor<string, []>("M_3")];
-            tensor<fp32, [1, 4, 4, 4]> var_405 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_405")];
+            tensor<int32, [2]> var_458 = const()[name = tensor<string, []>("op_458"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor<string, []>("op_459")];
+            tensor<fp32, [4, 4]> M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor<string, []>("M_3")];
+            tensor<fp32, [1, 4, 4, 4]> var_461 = mul(x = qk_3, y = M_3)[name = tensor<string, []>("op_461")];
             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, 4, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor<string, []>("transpose_49")];
-            tensor<fp32, [1, 4, 4, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor<string, []>("inner_3")];
-            tensor<bool, []> var_407_transpose_x_0 = const()[name = tensor<string, []>("op_407_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_407_transpose_y_0 = const()[name = tensor<string, []>("op_407_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 4, 64]> var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_407")];
-            tensor<fp32, [4]> var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_408")];
-            tensor<int32, [4]> var_409 = const()[name = tensor<string, []>("op_409"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_410 = reshape(shape = var_409, x = var_408)[name = tensor<string, []>("op_410")];
-            tensor<fp32, [1, 4, 4, 64]> cross_3 = mul(x = var_407, y = var_410)[name = tensor<string, []>("cross_3")];
+            tensor<fp32, [1, 4, 4, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor<string, []>("transpose_49")];
+            tensor<fp32, [1, 4, 4, 64]> inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_461, y = v_3)[name = tensor<string, []>("inner_3")];
+            tensor<bool, []> var_463_transpose_x_0 = const()[name = tensor<string, []>("op_463_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_463_transpose_y_0 = const()[name = tensor<string, []>("op_463_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 4, 64]> var_463 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor<string, []>("op_463")];
+            tensor<fp32, [4]> var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor<string, []>("op_464")];
+            tensor<int32, [4]> var_465 = const()[name = tensor<string, []>("op_465"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_466 = reshape(shape = var_465, x = var_464)[name = tensor<string, []>("op_466")];
+            tensor<fp32, [1, 4, 4, 64]> cross_3 = mul(x = var_463, y = var_466)[name = tensor<string, []>("cross_3")];
             tensor<fp32, [1, 4, 4, 64]> out_7 = add(x = inner_3, y = cross_3)[name = tensor<string, []>("out_7")];
-            tensor<fp32, [1, 4, 64, 64]> var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_413")];
-            tensor<bool, []> var_415_transpose_x_1 = const()[name = tensor<string, []>("op_415_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_415_transpose_y_1 = const()[name = tensor<string, []>("op_415_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_415")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor<string, []>("new_kv_unnorm_3")];
-            tensor<fp32, []> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<fp32, []>(0x1p+2)];
-            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor<string, []>("new_scale_3")];
-            tensor<fp32, [1]> var_419 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_419")];
-            tensor<fp32, [1, 4, 64, 64]> var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor<string, []>("op_420")];
-            tensor<int32, [4]> var_421_perm_0 = const()[name = tensor<string, []>("op_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor<string, []>("op_469")];
+            tensor<bool, []> var_471_transpose_x_1 = const()[name = tensor<string, []>("op_471_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_471_transpose_y_1 = const()[name = tensor<string, []>("op_471_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_471 = matmul(transpose_x = var_471_transpose_x_1, transpose_y = var_471_transpose_y_1, x = k_3, y = v_3)[name = tensor<string, []>("op_471")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor<string, []>("new_kv_unnorm_3")];
+            tensor<fp32, []> var_473 = const()[name = tensor<string, []>("op_473"), val = tensor<fp32, []>(0x1p+2)];
+            tensor<fp32, [1]> new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor<string, []>("new_scale_3")];
+            tensor<fp32, [1]> var_475 = sqrt(x = new_scale_3)[name = tensor<string, []>("op_475")];
+            tensor<fp32, [1, 4, 64, 64]> var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor<string, []>("op_476")];
+            tensor<int32, [4]> var_477_perm_0 = const()[name = tensor<string, []>("op_477_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, 4, 4, 64]> var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
-            tensor<fp32, [1, 4, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor<string, []>("out_9")];
-            tensor<int32, [3]> var_425 = const()[name = tensor<string, []>("op_425"), val = tensor<int32, [3]>([1, 4, 256])];
-            tensor<fp32, [1, 4, 256]> out_11 = reshape(shape = var_425, x = out_9)[name = tensor<string, []>("out_11")];
-            tensor<fp32, [1, 4, 256]> var_427 = silu(x = input_57)[name = tensor<string, []>("op_427")];
-            tensor<fp32, [1, 4, 256]> input_59 = mul(x = var_427, y = out_11)[name = tensor<string, []>("input_59")];
-            tensor<fp32, [1, 4, 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, 4, 256]> x_9 = add(x = input_53, y = ret_out_3)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 4, 4, 64]> var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor<string, []>("transpose_48")];
+            tensor<fp32, [1, 4, 4, 64]> out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_73, x = var_477)[name = tensor<string, []>("out_9")];
+            tensor<int32, [3]> var_481 = const()[name = tensor<string, []>("op_481"), val = tensor<int32, [3]>([1, 4, 256])];
+            tensor<fp32, [1, 4, 256]> out_11 = reshape(shape = var_481, x = out_9)[name = tensor<string, []>("out_11")];
+            tensor<fp32, [1, 4, 256]> var_483 = silu(x = input_59)[name = tensor<string, []>("op_483")];
+            tensor<fp32, [1, 4, 256]> input_61 = mul(x = var_483, y = out_11)[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 4, 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, 4, 256]> x_9 = add(x = input_55, y = ret_out_3)[name = tensor<string, []>("x_9")];
             tensor<int32, [4]> window_11_begin_0 = const()[name = tensor<string, []>("window_11_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> window_11_end_0 = const()[name = tensor<string, []>("window_11_end_0"), val = tensor<int32, [4]>([2, 1, 16, 256])];
             tensor<bool, [4]> window_11_end_mask_0 = const()[name = tensor<string, []>("window_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_11_squeeze_mask_0 = const()[name = tensor<string, []>("window_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[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, 0, 0])];
-            tensor<int32, [3]> var_435_end_0 = const()[name = tensor<string, []>("op_435_end_0"), val = tensor<int32, [3]>([1, 1, 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<int32, [3]> var_491_begin_0 = const()[name = tensor<string, []>("op_491_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_491_end_0 = const()[name = tensor<string, []>("op_491_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_491_end_mask_0 = const()[name = tensor<string, []>("op_491_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_491 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = x_9)[name = tensor<string, []>("op_491")];
+            tensor<int32, [3]> var_494_begin_0 = const()[name = tensor<string, []>("op_494_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_494_end_0 = const()[name = tensor<string, []>("op_494_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_494_end_mask_0 = const()[name = tensor<string, []>("op_494_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_494 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = window_11)[name = tensor<string, []>("op_494")];
             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, 1, 0])];
-            tensor<int32, [3]> var_443_end_0 = const()[name = tensor<string, []>("op_443_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_443_end_mask_0 = const()[name = tensor<string, []>("op_443_end_mask_0"), val = tensor<bool, [3]>([true, false, 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_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor<string, []>("window_13")];
+            tensor<int32, [3]> var_499_begin_0 = const()[name = tensor<string, []>("op_499_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_499_end_0 = const()[name = tensor<string, []>("op_499_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_499_end_mask_0 = const()[name = tensor<string, []>("op_499_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_499 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = x_9)[name = tensor<string, []>("op_499")];
+            tensor<int32, [3]> var_502_begin_0 = const()[name = tensor<string, []>("op_502_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_502_end_0 = const()[name = tensor<string, []>("op_502_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_502_end_mask_0 = const()[name = tensor<string, []>("op_502_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_502 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = window_13)[name = tensor<string, []>("op_502")];
             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<int32, [3]> var_451_begin_0 = const()[name = tensor<string, []>("op_451_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_451_end_0 = const()[name = tensor<string, []>("op_451_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<bool, [3]> var_451_end_mask_0 = const()[name = tensor<string, []>("op_451_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor<string, []>("op_451")];
-            tensor<int32, [3]> var_454_begin_0 = const()[name = tensor<string, []>("op_454_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_454_end_0 = const()[name = tensor<string, []>("op_454_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_454_end_mask_0 = const()[name = tensor<string, []>("op_454_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor<string, []>("op_454")];
+            tensor<fp32, [1, 16, 256]> window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor<string, []>("window_15")];
+            tensor<int32, [3]> var_507_begin_0 = const()[name = tensor<string, []>("op_507_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_507_end_0 = const()[name = tensor<string, []>("op_507_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<bool, [3]> var_507_end_mask_0 = const()[name = tensor<string, []>("op_507_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_507 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = x_9)[name = tensor<string, []>("op_507")];
+            tensor<int32, [3]> var_510_begin_0 = const()[name = tensor<string, []>("op_510_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_510_end_0 = const()[name = tensor<string, []>("op_510_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_510_end_mask_0 = const()[name = tensor<string, []>("op_510_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = window_15)[name = tensor<string, []>("op_510")];
             tensor<bool, []> window_17_interleave_0 = const()[name = tensor<string, []>("window_17_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor<string, []>("window_17")];
-            tensor<int32, [3]> var_459_begin_0 = const()[name = tensor<string, []>("op_459_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
-            tensor<int32, [3]> var_459_end_0 = const()[name = tensor<string, []>("op_459_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_459_end_mask_0 = const()[name = tensor<string, []>("op_459_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor<string, []>("op_459")];
-            tensor<int32, [3]> var_462_begin_0 = const()[name = tensor<string, []>("op_462_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_462_end_0 = const()[name = tensor<string, []>("op_462_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_462_end_mask_0 = const()[name = tensor<string, []>("op_462_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor<string, []>("op_462")];
+            tensor<fp32, [1, 16, 256]> window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor<string, []>("window_17")];
+            tensor<int32, [3]> var_515_begin_0 = const()[name = tensor<string, []>("op_515_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
+            tensor<int32, [3]> var_515_end_0 = const()[name = tensor<string, []>("op_515_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_515_end_mask_0 = const()[name = tensor<string, []>("op_515_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_515 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = x_9)[name = tensor<string, []>("op_515")];
+            tensor<int32, [3]> var_518_begin_0 = const()[name = tensor<string, []>("op_518_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_518_end_0 = const()[name = tensor<string, []>("op_518_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_518_end_mask_0 = const()[name = tensor<string, []>("op_518_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = window_17)[name = tensor<string, []>("op_518")];
             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_462, var_459))[name = tensor<string, []>("window_19")];
-            tensor<bool, []> input_61_interleave_0 = const()[name = tensor<string, []>("input_61_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [4, 16, 256]> input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor<string, []>("input_61")];
+            tensor<fp32, [1, 16, 256]> window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor<string, []>("window_19")];
+            tensor<bool, []> input_63_interleave_0 = const()[name = tensor<string, []>("input_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [4, 16, 256]> input_63 = concat(axis = var_68, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[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, [4, 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, [4, 16, 256]> x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_65, 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, [4, 256, 16]> input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
-            tensor<fp32, [4, 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_487_split_sizes_0 = const()[name = tensor<string, []>("op_487_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_487_axis_0 = const()[name = tensor<string, []>("op_487_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [4, 256, 16]> var_487_0, tensor<fp32, [4, 256, 16]> var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_487")];
-            tensor<fp32, [4, 256, 16]> var_489 = sigmoid(x = var_487_1)[name = tensor<string, []>("op_489")];
-            tensor<fp32, [4, 256, 16]> inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor<string, []>("inputs_15")];
+            tensor<fp32, [4, 256, 16]> input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor<string, []>("transpose_47")];
+            tensor<fp32, [4, 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_543_split_sizes_0 = const()[name = tensor<string, []>("op_543_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_543_axis_0 = const()[name = tensor<string, []>("op_543_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [4, 256, 16]> var_543_0, tensor<fp32, [4, 256, 16]> var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor<string, []>("op_543")];
+            tensor<fp32, [4, 256, 16]> var_545 = sigmoid(x = var_543_1)[name = tensor<string, []>("op_545")];
+            tensor<fp32, [4, 256, 16]> inputs_15 = mul(x = var_543_0, y = var_545)[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, [4, 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]>([4, 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, [4, 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, [4, 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, [4, 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, [4, 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, [4, 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]>([4, 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, [4, 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, [4, 256, 16]> inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_65, 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, [4, 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, [4, 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_520_begin_0 = const()[name = tensor<string, []>("op_520_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_520_end_0 = const()[name = tensor<string, []>("op_520_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
-            tensor<bool, [3]> var_520_end_mask_0 = const()[name = tensor<string, []>("op_520_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [4, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor<string, []>("transpose_46")];
-            tensor<fp32, [4, 1, 256]> var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_520")];
-            tensor<int32, [3]> var_522_perm_0 = const()[name = tensor<string, []>("op_522_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 4, 256]> var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor<string, []>("transpose_45")];
-            tensor<fp32, [1, 4, 256]> input_71 = add(x = x_9, y = var_522)[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, 4, 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, 4, 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, 4, 1024]> input_75 = silu(x = inputs_19)[name = tensor<string, []>("input_75")];
-            tensor<fp32, [1, 4, 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_545 = const()[name = tensor<string, []>("op_545"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_546 = mul(x = input_79, y = var_545)[name = tensor<string, []>("op_546")];
-            tensor<fp32, [1, 4, 256]> input_81 = add(x = var_546, 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, 4, 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_576_begin_0 = const()[name = tensor<string, []>("op_576_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_576_end_0 = const()[name = tensor<string, []>("op_576_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
+            tensor<bool, [3]> var_576_end_mask_0 = const()[name = tensor<string, []>("op_576_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [4, 16, 256]> conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor<string, []>("transpose_46")];
+            tensor<fp32, [4, 1, 256]> var_576 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = conv_out_3)[name = tensor<string, []>("op_576")];
+            tensor<int32, [3]> var_578_perm_0 = const()[name = tensor<string, []>("op_578_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 4, 256]> var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor<string, []>("transpose_45")];
+            tensor<fp32, [1, 4, 256]> input_73 = add(x = x_9, y = var_578)[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, 4, 256]> input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor<string, []>("input_75")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_77 = silu(x = inputs_19)[name = tensor<string, []>("input_77")];
+            tensor<fp32, [1, 4, 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_601 = const()[name = tensor<string, []>("op_601"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_602 = mul(x = input_81, y = var_601)[name = tensor<string, []>("op_602")];
+            tensor<fp32, [1, 4, 256]> input_83 = add(x = var_602, 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, 4, 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, 4, 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, 4, 1024]> input_87 = silu(x = inputs_21)[name = tensor<string, []>("input_87")];
-            tensor<fp32, [1, 4, 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_575 = const()[name = tensor<string, []>("op_575"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_576 = mul(x = input_91, y = var_575)[name = tensor<string, []>("op_576")];
-            tensor<fp32, [1, 4, 256]> input_93 = add(x = var_576, y = input_83)[name = tensor<string, []>("input_93")];
+            tensor<fp32, [1, 4, 256]> input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_65, 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, 4, 256]> input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor<string, []>("input_87")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_89 = silu(x = inputs_21)[name = tensor<string, []>("input_89")];
+            tensor<fp32, [1, 4, 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_631 = const()[name = tensor<string, []>("op_631"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_632 = mul(x = input_93, y = var_631)[name = tensor<string, []>("op_632")];
+            tensor<fp32, [1, 4, 256]> input_95 = add(x = var_632, 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, 4, 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, 4, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_65, 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])];
@@ -593,173 +615,173 @@ 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, 4, 256]> var_590 = 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_591 = const()[name = tensor<string, []>("op_591"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_592 = reshape(shape = var_591, x = var_590)[name = tensor<string, []>("op_592")];
+            tensor<fp32, [1, 4, 256]> var_646 = 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_647 = const()[name = tensor<string, []>("op_647"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_648 = reshape(shape = var_647, x = var_646)[name = tensor<string, []>("op_648")];
             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, 4, 256]> var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor<string, []>("linear_22")];
-            tensor<fp32, []> var_597 = const()[name = tensor<string, []>("op_597"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 4, 256]> var_598 = mul(x = var_596, y = var_597)[name = tensor<string, []>("op_598")];
-            tensor<int32, [4]> var_599 = const()[name = tensor<string, []>("op_599"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_600 = reshape(shape = var_599, x = var_598)[name = tensor<string, []>("op_600")];
+            tensor<fp32, [1, 4, 256]> var_652 = 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_653 = const()[name = tensor<string, []>("op_653"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 4, 256]> var_654 = mul(x = var_652, y = var_653)[name = tensor<string, []>("op_654")];
+            tensor<int32, [4]> var_655 = const()[name = tensor<string, []>("op_655"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_656 = reshape(shape = var_655, x = var_654)[name = tensor<string, []>("op_656")];
             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, 4, 256]> var_604 = 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_605 = const()[name = tensor<string, []>("op_605"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_606 = reshape(shape = var_605, x = var_604)[name = tensor<string, []>("op_606")];
+            tensor<fp32, [1, 4, 256]> var_660 = 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_661 = const()[name = tensor<string, []>("op_661"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_662 = reshape(shape = var_661, x = var_660)[name = tensor<string, []>("op_662")];
             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, 4, 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, 4, 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, [4]> s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor<string, []>("s_t_5")];
+            tensor<fp32, [4]> s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_5")];
             tensor<fp32, [4]> 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, 4, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor<string, []>("transpose_43")];
-            tensor<fp32, [1, 4, 4, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor<string, []>("transpose_44")];
+            tensor<fp32, [1, 4, 4, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor<string, []>("transpose_43")];
+            tensor<fp32, [1, 4, 4, 64]> q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor<string, []>("transpose_44")];
             tensor<fp32, [1, 4, 4, 4]> 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_616 = const()[name = tensor<string, []>("op_616"), val = tensor<int32, [2]>([4, 1])];
-            tensor<fp32, [4, 1]> var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor<string, []>("op_617")];
-            tensor<fp32, [4, 4]> M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor<string, []>("M_5")];
-            tensor<fp32, [1, 4, 4, 4]> var_619 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_619")];
+            tensor<int32, [2]> var_672 = const()[name = tensor<string, []>("op_672"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor<string, []>("op_673")];
+            tensor<fp32, [4, 4]> M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor<string, []>("M_5")];
+            tensor<fp32, [1, 4, 4, 4]> var_675 = mul(x = qk_5, y = M_5)[name = tensor<string, []>("op_675")];
             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, 4, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor<string, []>("transpose_42")];
-            tensor<fp32, [1, 4, 4, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor<string, []>("inner_5")];
-            tensor<bool, []> var_621_transpose_x_0 = const()[name = tensor<string, []>("op_621_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_621_transpose_y_0 = const()[name = tensor<string, []>("op_621_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 4, 64]> var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_621")];
-            tensor<fp32, [4]> var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_622")];
-            tensor<int32, [4]> var_623 = const()[name = tensor<string, []>("op_623"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_624 = reshape(shape = var_623, x = var_622)[name = tensor<string, []>("op_624")];
-            tensor<fp32, [1, 4, 4, 64]> cross_5 = mul(x = var_621, y = var_624)[name = tensor<string, []>("cross_5")];
+            tensor<fp32, [1, 4, 4, 64]> v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor<string, []>("transpose_42")];
+            tensor<fp32, [1, 4, 4, 64]> inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_675, y = v_5)[name = tensor<string, []>("inner_5")];
+            tensor<bool, []> var_677_transpose_x_0 = const()[name = tensor<string, []>("op_677_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_677_transpose_y_0 = const()[name = tensor<string, []>("op_677_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 4, 64]> var_677 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor<string, []>("op_677")];
+            tensor<fp32, [4]> var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor<string, []>("op_678")];
+            tensor<int32, [4]> var_679 = const()[name = tensor<string, []>("op_679"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_680 = reshape(shape = var_679, x = var_678)[name = tensor<string, []>("op_680")];
+            tensor<fp32, [1, 4, 4, 64]> cross_5 = mul(x = var_677, y = var_680)[name = tensor<string, []>("cross_5")];
             tensor<fp32, [1, 4, 4, 64]> out_13 = add(x = inner_5, y = cross_5)[name = tensor<string, []>("out_13")];
-            tensor<fp32, [1, 4, 64, 64]> var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_627")];
-            tensor<bool, []> var_629_transpose_x_1 = const()[name = tensor<string, []>("op_629_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_629_transpose_y_1 = const()[name = tensor<string, []>("op_629_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_629")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor<string, []>("new_kv_unnorm_5")];
-            tensor<fp32, []> var_631 = const()[name = tensor<string, []>("op_631"), val = tensor<fp32, []>(0x1p+2)];
-            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor<string, []>("new_scale_5")];
-            tensor<fp32, [1]> var_633 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_633")];
-            tensor<fp32, [1, 4, 64, 64]> var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor<string, []>("op_634")];
-            tensor<int32, [4]> var_635_perm_0 = const()[name = tensor<string, []>("op_635_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor<string, []>("op_683")];
+            tensor<bool, []> var_685_transpose_x_1 = const()[name = tensor<string, []>("op_685_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_685_transpose_y_1 = const()[name = tensor<string, []>("op_685_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_685 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = k_5, y = v_5)[name = tensor<string, []>("op_685")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor<string, []>("new_kv_unnorm_5")];
+            tensor<fp32, []> var_687 = const()[name = tensor<string, []>("op_687"), val = tensor<fp32, []>(0x1p+2)];
+            tensor<fp32, [1]> new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor<string, []>("new_scale_5")];
+            tensor<fp32, [1]> var_689 = sqrt(x = new_scale_5)[name = tensor<string, []>("op_689")];
+            tensor<fp32, [1, 4, 64, 64]> var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor<string, []>("op_690")];
+            tensor<int32, [4]> var_691_perm_0 = const()[name = tensor<string, []>("op_691_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, 4, 4, 64]> var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
-            tensor<fp32, [1, 4, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor<string, []>("out_15")];
-            tensor<int32, [3]> var_639 = const()[name = tensor<string, []>("op_639"), val = tensor<int32, [3]>([1, 4, 256])];
-            tensor<fp32, [1, 4, 256]> out_17 = reshape(shape = var_639, x = out_15)[name = tensor<string, []>("out_17")];
-            tensor<fp32, [1, 4, 256]> var_641 = silu(x = input_97)[name = tensor<string, []>("op_641")];
-            tensor<fp32, [1, 4, 256]> input_99 = mul(x = var_641, y = out_17)[name = tensor<string, []>("input_99")];
-            tensor<fp32, [1, 4, 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, 4, 256]> x_15 = add(x = input_93, y = ret_out_5)[name = tensor<string, []>("x_15")];
+            tensor<fp32, [1, 4, 4, 64]> var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor<string, []>("transpose_41")];
+            tensor<fp32, [1, 4, 4, 64]> out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_73, x = var_691)[name = tensor<string, []>("out_15")];
+            tensor<int32, [3]> var_695 = const()[name = tensor<string, []>("op_695"), val = tensor<int32, [3]>([1, 4, 256])];
+            tensor<fp32, [1, 4, 256]> out_17 = reshape(shape = var_695, x = out_15)[name = tensor<string, []>("out_17")];
+            tensor<fp32, [1, 4, 256]> var_697 = silu(x = input_99)[name = tensor<string, []>("op_697")];
+            tensor<fp32, [1, 4, 256]> input_101 = mul(x = var_697, y = out_17)[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 4, 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, 4, 256]> x_15 = add(x = input_95, y = ret_out_5)[name = tensor<string, []>("x_15")];
             tensor<int32, [4]> window_21_begin_0 = const()[name = tensor<string, []>("window_21_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> window_21_end_0 = const()[name = tensor<string, []>("window_21_end_0"), val = tensor<int32, [4]>([3, 1, 16, 256])];
             tensor<bool, [4]> window_21_end_mask_0 = const()[name = tensor<string, []>("window_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_21_squeeze_mask_0 = const()[name = tensor<string, []>("window_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[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, 0, 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, false, 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<int32, [3]> var_705_begin_0 = const()[name = tensor<string, []>("op_705_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_705_end_0 = const()[name = tensor<string, []>("op_705_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_705_end_mask_0 = const()[name = tensor<string, []>("op_705_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_705 = slice_by_index(begin = var_705_begin_0, end = var_705_end_0, end_mask = var_705_end_mask_0, x = x_15)[name = tensor<string, []>("op_705")];
+            tensor<int32, [3]> var_708_begin_0 = const()[name = tensor<string, []>("op_708_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_708_end_0 = const()[name = tensor<string, []>("op_708_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_708_end_mask_0 = const()[name = tensor<string, []>("op_708_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_708 = slice_by_index(begin = var_708_begin_0, end = var_708_end_0, end_mask = var_708_end_mask_0, x = window_21)[name = tensor<string, []>("op_708")];
             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<int32, [3]> var_657_begin_0 = const()[name = tensor<string, []>("op_657_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_657_end_0 = const()[name = tensor<string, []>("op_657_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_657_end_mask_0 = const()[name = tensor<string, []>("op_657_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor<string, []>("op_657")];
-            tensor<int32, [3]> var_660_begin_0 = const()[name = tensor<string, []>("op_660_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_660_end_0 = const()[name = tensor<string, []>("op_660_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_660_end_mask_0 = const()[name = tensor<string, []>("op_660_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor<string, []>("op_660")];
+            tensor<fp32, [1, 16, 256]> window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor<string, []>("window_23")];
+            tensor<int32, [3]> var_713_begin_0 = const()[name = tensor<string, []>("op_713_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_713_end_0 = const()[name = tensor<string, []>("op_713_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_713_end_mask_0 = const()[name = tensor<string, []>("op_713_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_713 = slice_by_index(begin = var_713_begin_0, end = var_713_end_0, end_mask = var_713_end_mask_0, x = x_15)[name = tensor<string, []>("op_713")];
+            tensor<int32, [3]> var_716_begin_0 = const()[name = tensor<string, []>("op_716_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_716_end_0 = const()[name = tensor<string, []>("op_716_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_716_end_mask_0 = const()[name = tensor<string, []>("op_716_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_716 = slice_by_index(begin = var_716_begin_0, end = var_716_end_0, end_mask = var_716_end_mask_0, x = window_23)[name = tensor<string, []>("op_716")];
             tensor<bool, []> window_25_interleave_0 = const()[name = tensor<string, []>("window_25_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor<string, []>("window_25")];
-            tensor<int32, [3]> var_665_begin_0 = const()[name = tensor<string, []>("op_665_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_665_end_0 = const()[name = tensor<string, []>("op_665_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<bool, [3]> var_665_end_mask_0 = const()[name = tensor<string, []>("op_665_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor<string, []>("op_665")];
-            tensor<int32, [3]> var_668_begin_0 = const()[name = tensor<string, []>("op_668_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_668_end_0 = const()[name = tensor<string, []>("op_668_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_668_end_mask_0 = const()[name = tensor<string, []>("op_668_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor<string, []>("op_668")];
+            tensor<fp32, [1, 16, 256]> window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor<string, []>("window_25")];
+            tensor<int32, [3]> var_721_begin_0 = const()[name = tensor<string, []>("op_721_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_721_end_0 = const()[name = tensor<string, []>("op_721_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<bool, [3]> var_721_end_mask_0 = const()[name = tensor<string, []>("op_721_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_721 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = x_15)[name = tensor<string, []>("op_721")];
+            tensor<int32, [3]> var_724_begin_0 = const()[name = tensor<string, []>("op_724_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_724_end_0 = const()[name = tensor<string, []>("op_724_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_724_end_mask_0 = const()[name = tensor<string, []>("op_724_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_724 = slice_by_index(begin = var_724_begin_0, end = var_724_end_0, end_mask = var_724_end_mask_0, x = window_25)[name = tensor<string, []>("op_724")];
             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_668, var_665))[name = tensor<string, []>("window_27")];
-            tensor<int32, [3]> var_673_begin_0 = const()[name = tensor<string, []>("op_673_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
-            tensor<int32, [3]> var_673_end_0 = const()[name = tensor<string, []>("op_673_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_673_end_mask_0 = const()[name = tensor<string, []>("op_673_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor<string, []>("op_673")];
-            tensor<int32, [3]> var_676_begin_0 = const()[name = tensor<string, []>("op_676_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_676_end_0 = const()[name = tensor<string, []>("op_676_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_676_end_mask_0 = const()[name = tensor<string, []>("op_676_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor<string, []>("op_676")];
+            tensor<fp32, [1, 16, 256]> window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor<string, []>("window_27")];
+            tensor<int32, [3]> var_729_begin_0 = const()[name = tensor<string, []>("op_729_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
+            tensor<int32, [3]> var_729_end_0 = const()[name = tensor<string, []>("op_729_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_729_end_mask_0 = const()[name = tensor<string, []>("op_729_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_729 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = x_15)[name = tensor<string, []>("op_729")];
+            tensor<int32, [3]> var_732_begin_0 = const()[name = tensor<string, []>("op_732_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_732_end_0 = const()[name = tensor<string, []>("op_732_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_732_end_mask_0 = const()[name = tensor<string, []>("op_732_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = window_27)[name = tensor<string, []>("op_732")];
             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_676, var_673))[name = tensor<string, []>("window_29")];
-            tensor<bool, []> input_101_interleave_0 = const()[name = tensor<string, []>("input_101_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [4, 16, 256]> input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, 16, 256]> window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor<string, []>("window_29")];
+            tensor<bool, []> input_103_interleave_0 = const()[name = tensor<string, []>("input_103_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [4, 16, 256]> input_103 = concat(axis = var_68, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[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, [4, 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, [4, 16, 256]> x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_65, 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, [4, 256, 16]> input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
-            tensor<fp32, [4, 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_701_split_sizes_0 = const()[name = tensor<string, []>("op_701_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_701_axis_0 = const()[name = tensor<string, []>("op_701_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [4, 256, 16]> var_701_0, tensor<fp32, [4, 256, 16]> var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_701")];
-            tensor<fp32, [4, 256, 16]> var_703 = sigmoid(x = var_701_1)[name = tensor<string, []>("op_703")];
-            tensor<fp32, [4, 256, 16]> inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor<string, []>("inputs_25")];
+            tensor<fp32, [4, 256, 16]> input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor<string, []>("transpose_40")];
+            tensor<fp32, [4, 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_757_split_sizes_0 = const()[name = tensor<string, []>("op_757_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_757_axis_0 = const()[name = tensor<string, []>("op_757_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [4, 256, 16]> var_757_0, tensor<fp32, [4, 256, 16]> var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor<string, []>("op_757")];
+            tensor<fp32, [4, 256, 16]> var_759 = sigmoid(x = var_757_1)[name = tensor<string, []>("op_759")];
+            tensor<fp32, [4, 256, 16]> inputs_25 = mul(x = var_757_0, y = var_759)[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, [4, 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]>([4, 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, [4, 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, [4, 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, [4, 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, [4, 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, [4, 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]>([4, 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, [4, 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, [4, 256, 16]> inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_65, 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, [4, 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, [4, 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_734_begin_0 = const()[name = tensor<string, []>("op_734_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_734_end_0 = const()[name = tensor<string, []>("op_734_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
-            tensor<bool, [3]> var_734_end_mask_0 = const()[name = tensor<string, []>("op_734_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [4, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor<string, []>("transpose_39")];
-            tensor<fp32, [4, 1, 256]> var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_734")];
-            tensor<int32, [3]> var_736_perm_0 = const()[name = tensor<string, []>("op_736_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 4, 256]> var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor<string, []>("transpose_38")];
-            tensor<fp32, [1, 4, 256]> input_111 = add(x = x_15, y = var_736)[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, 4, 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, 4, 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, 4, 1024]> input_115 = silu(x = inputs_29)[name = tensor<string, []>("input_115")];
-            tensor<fp32, [1, 4, 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_759 = const()[name = tensor<string, []>("op_759"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_760 = mul(x = input_119, y = var_759)[name = tensor<string, []>("op_760")];
-            tensor<fp32, [1, 4, 256]> input_121 = add(x = var_760, 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, 4, 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_790_begin_0 = const()[name = tensor<string, []>("op_790_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_790_end_0 = const()[name = tensor<string, []>("op_790_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
+            tensor<bool, [3]> var_790_end_mask_0 = const()[name = tensor<string, []>("op_790_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [4, 16, 256]> conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor<string, []>("transpose_39")];
+            tensor<fp32, [4, 1, 256]> var_790 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = conv_out_5)[name = tensor<string, []>("op_790")];
+            tensor<int32, [3]> var_792_perm_0 = const()[name = tensor<string, []>("op_792_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 4, 256]> var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor<string, []>("transpose_38")];
+            tensor<fp32, [1, 4, 256]> input_113 = add(x = x_15, y = var_792)[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, 4, 256]> input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor<string, []>("input_115")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_117 = silu(x = inputs_29)[name = tensor<string, []>("input_117")];
+            tensor<fp32, [1, 4, 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_815 = const()[name = tensor<string, []>("op_815"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_816 = mul(x = input_121, y = var_815)[name = tensor<string, []>("op_816")];
+            tensor<fp32, [1, 4, 256]> input_123 = add(x = var_816, 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, 4, 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, 4, 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, 4, 1024]> input_127 = silu(x = inputs_31)[name = tensor<string, []>("input_127")];
-            tensor<fp32, [1, 4, 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_789 = const()[name = tensor<string, []>("op_789"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_790 = mul(x = input_131, y = var_789)[name = tensor<string, []>("op_790")];
-            tensor<fp32, [1, 4, 256]> input_133 = add(x = var_790, y = input_123)[name = tensor<string, []>("input_133")];
+            tensor<fp32, [1, 4, 256]> input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_65, 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, 4, 256]> input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor<string, []>("input_127")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_129 = silu(x = inputs_31)[name = tensor<string, []>("input_129")];
+            tensor<fp32, [1, 4, 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_845 = const()[name = tensor<string, []>("op_845"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_846 = mul(x = input_133, y = var_845)[name = tensor<string, []>("op_846")];
+            tensor<fp32, [1, 4, 256]> input_135 = add(x = var_846, 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, 4, 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, 4, 256]> x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_65, 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])];
@@ -770,209 +792,202 @@ 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, 4, 256]> var_804 = 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_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_806 = reshape(shape = var_805, x = var_804)[name = tensor<string, []>("op_806")];
+            tensor<fp32, [1, 4, 256]> var_860 = 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_861 = const()[name = tensor<string, []>("op_861"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_862 = reshape(shape = var_861, x = var_860)[name = tensor<string, []>("op_862")];
             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, 4, 256]> var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor<string, []>("linear_31")];
-            tensor<fp32, []> var_811 = const()[name = tensor<string, []>("op_811"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [1, 4, 256]> var_812 = mul(x = var_810, y = var_811)[name = tensor<string, []>("op_812")];
-            tensor<int32, [4]> var_813 = const()[name = tensor<string, []>("op_813"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_814 = reshape(shape = var_813, x = var_812)[name = tensor<string, []>("op_814")];
+            tensor<fp32, [1, 4, 256]> var_866 = 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_867 = const()[name = tensor<string, []>("op_867"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [1, 4, 256]> var_868 = mul(x = var_866, y = var_867)[name = tensor<string, []>("op_868")];
+            tensor<int32, [4]> var_869 = const()[name = tensor<string, []>("op_869"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_870 = reshape(shape = var_869, x = var_868)[name = tensor<string, []>("op_870")];
             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, 4, 256]> var_818 = 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_819 = const()[name = tensor<string, []>("op_819"), val = tensor<int32, [4]>([1, 4, 4, 64])];
-            tensor<fp32, [1, 4, 4, 64]> var_820 = reshape(shape = var_819, x = var_818)[name = tensor<string, []>("op_820")];
+            tensor<fp32, [1, 4, 256]> var_874 = 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_875 = const()[name = tensor<string, []>("op_875"), val = tensor<int32, [4]>([1, 4, 4, 64])];
+            tensor<fp32, [1, 4, 4, 64]> var_876 = reshape(shape = var_875, x = var_874)[name = tensor<string, []>("op_876")];
             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, 4, 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, 4, 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, [4]> s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor<string, []>("s_t_7")];
+            tensor<fp32, [4]> s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor<string, []>("s_t_7")];
             tensor<fp32, [4]> 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, 4, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor<string, []>("transpose_36")];
-            tensor<fp32, [1, 4, 4, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor<string, []>("transpose_37")];
+            tensor<fp32, [1, 4, 4, 64]> k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor<string, []>("transpose_36")];
+            tensor<fp32, [1, 4, 4, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor<string, []>("transpose_37")];
             tensor<fp32, [1, 4, 4, 4]> 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_830 = const()[name = tensor<string, []>("op_830"), val = tensor<int32, [2]>([4, 1])];
-            tensor<fp32, [4, 1]> var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor<string, []>("op_831")];
-            tensor<fp32, [4, 4]> M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor<string, []>("M_7")];
-            tensor<fp32, [1, 4, 4, 4]> var_833 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_833")];
+            tensor<int32, [2]> var_886 = const()[name = tensor<string, []>("op_886"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor<string, []>("op_887")];
+            tensor<fp32, [4, 4]> M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor<string, []>("M_7")];
+            tensor<fp32, [1, 4, 4, 4]> var_889 = mul(x = qk_7, y = M_7)[name = tensor<string, []>("op_889")];
             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, 4, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor<string, []>("transpose_35")];
-            tensor<fp32, [1, 4, 4, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor<string, []>("inner_7")];
-            tensor<bool, []> var_835_transpose_x_0 = const()[name = tensor<string, []>("op_835_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_835_transpose_y_0 = const()[name = tensor<string, []>("op_835_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 4, 64]> var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_835")];
-            tensor<fp32, [4]> var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_836")];
-            tensor<int32, [4]> var_837 = const()[name = tensor<string, []>("op_837"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_838 = reshape(shape = var_837, x = var_836)[name = tensor<string, []>("op_838")];
-            tensor<fp32, [1, 4, 4, 64]> cross_7 = mul(x = var_835, y = var_838)[name = tensor<string, []>("cross_7")];
+            tensor<fp32, [1, 4, 4, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor<string, []>("transpose_35")];
+            tensor<fp32, [1, 4, 4, 64]> inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_889, y = v_7)[name = tensor<string, []>("inner_7")];
+            tensor<bool, []> var_891_transpose_x_0 = const()[name = tensor<string, []>("op_891_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_891_transpose_y_0 = const()[name = tensor<string, []>("op_891_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 4, 64]> var_891 = matmul(transpose_x = var_891_transpose_x_0, transpose_y = var_891_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor<string, []>("op_891")];
+            tensor<fp32, [4]> var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor<string, []>("op_892")];
+            tensor<int32, [4]> var_893 = const()[name = tensor<string, []>("op_893"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_894 = reshape(shape = var_893, x = var_892)[name = tensor<string, []>("op_894")];
+            tensor<fp32, [1, 4, 4, 64]> cross_7 = mul(x = var_891, y = var_894)[name = tensor<string, []>("cross_7")];
             tensor<fp32, [1, 4, 4, 64]> out_19 = add(x = inner_7, y = cross_7)[name = tensor<string, []>("out_19")];
-            tensor<fp32, [1, 4, 64, 64]> var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_841")];
-            tensor<bool, []> var_843_transpose_x_1 = const()[name = tensor<string, []>("op_843_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_843_transpose_y_1 = const()[name = tensor<string, []>("op_843_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 4, 64, 64]> var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_843")];
-            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor<string, []>("new_kv_unnorm_7")];
-            tensor<fp32, []> var_845 = const()[name = tensor<string, []>("op_845"), val = tensor<fp32, []>(0x1p+2)];
-            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor<string, []>("new_scale_7")];
-            tensor<fp32, [1]> var_847 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_847")];
-            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor<string, []>("nkv_1")];
-            tensor<int32, [4]> var_849_perm_0 = const()[name = tensor<string, []>("op_849_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [1, 4, 64, 64]> var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor<string, []>("op_897")];
+            tensor<bool, []> var_899_transpose_x_1 = const()[name = tensor<string, []>("op_899_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_899_transpose_y_1 = const()[name = tensor<string, []>("op_899_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 4, 64, 64]> var_899 = matmul(transpose_x = var_899_transpose_x_1, transpose_y = var_899_transpose_y_1, x = k_7, y = v_7)[name = tensor<string, []>("op_899")];
+            tensor<fp32, [1, 4, 64, 64]> new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor<string, []>("new_kv_unnorm_7")];
+            tensor<fp32, []> var_901 = const()[name = tensor<string, []>("op_901"), val = tensor<fp32, []>(0x1p+2)];
+            tensor<fp32, [1]> new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor<string, []>("new_scale_7")];
+            tensor<fp32, [1]> var_903 = sqrt(x = new_scale_7)[name = tensor<string, []>("op_903")];
+            tensor<fp32, [1, 4, 64, 64]> nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor<string, []>("nkv_1")];
+            tensor<int32, [4]> var_905_perm_0 = const()[name = tensor<string, []>("op_905_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, 4, 4, 64]> var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
-            tensor<fp32, [1, 4, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor<string, []>("out_21")];
-            tensor<int32, [3]> var_853 = const()[name = tensor<string, []>("op_853"), val = tensor<int32, [3]>([1, 4, 256])];
-            tensor<fp32, [1, 4, 256]> out_23 = reshape(shape = var_853, x = out_21)[name = tensor<string, []>("out_23")];
-            tensor<fp32, [1, 4, 256]> var_855 = silu(x = input_137)[name = tensor<string, []>("op_855")];
-            tensor<fp32, [1, 4, 256]> input_139 = mul(x = var_855, y = out_23)[name = tensor<string, []>("input_139")];
-            tensor<fp32, [1, 4, 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, 4, 256]> x_21 = add(x = input_133, y = ret_out_7)[name = tensor<string, []>("x_21")];
+            tensor<fp32, [1, 4, 4, 64]> var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor<string, []>("transpose_34")];
+            tensor<fp32, [1, 4, 4, 64]> out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_73, x = var_905)[name = tensor<string, []>("out_21")];
+            tensor<int32, [3]> var_909 = const()[name = tensor<string, []>("op_909"), val = tensor<int32, [3]>([1, 4, 256])];
+            tensor<fp32, [1, 4, 256]> out_23 = reshape(shape = var_909, x = out_21)[name = tensor<string, []>("out_23")];
+            tensor<fp32, [1, 4, 256]> var_911 = silu(x = input_139)[name = tensor<string, []>("op_911")];
+            tensor<fp32, [1, 4, 256]> input_141 = mul(x = var_911, y = out_23)[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 4, 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, 4, 256]> x_21 = add(x = input_135, y = ret_out_7)[name = tensor<string, []>("x_21")];
             tensor<int32, [4]> window_31_begin_0 = const()[name = tensor<string, []>("window_31_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
             tensor<int32, [4]> window_31_end_0 = const()[name = tensor<string, []>("window_31_end_0"), val = tensor<int32, [4]>([4, 1, 16, 256])];
             tensor<bool, [4]> window_31_end_mask_0 = const()[name = tensor<string, []>("window_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> window_31_squeeze_mask_0 = const()[name = tensor<string, []>("window_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
             tensor<fp32, [1, 16, 256]> window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor<string, []>("window_31")];
-            tensor<int32, [3]> var_863_begin_0 = const()[name = tensor<string, []>("op_863_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
-            tensor<int32, [3]> var_863_end_0 = const()[name = tensor<string, []>("op_863_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_863_end_mask_0 = const()[name = tensor<string, []>("op_863_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor<string, []>("op_863")];
-            tensor<int32, [3]> var_866_begin_0 = const()[name = tensor<string, []>("op_866_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_866_end_0 = const()[name = tensor<string, []>("op_866_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_866_end_mask_0 = const()[name = tensor<string, []>("op_866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor<string, []>("op_866")];
+            tensor<int32, [3]> var_919_begin_0 = const()[name = tensor<string, []>("op_919_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_919_end_0 = const()[name = tensor<string, []>("op_919_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_919_end_mask_0 = const()[name = tensor<string, []>("op_919_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_919 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_21)[name = tensor<string, []>("op_919")];
+            tensor<int32, [3]> var_922_begin_0 = const()[name = tensor<string, []>("op_922_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_922_end_0 = const()[name = tensor<string, []>("op_922_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_922_end_mask_0 = const()[name = tensor<string, []>("op_922_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_922 = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = window_31)[name = tensor<string, []>("op_922")];
             tensor<bool, []> window_33_interleave_0 = const()[name = tensor<string, []>("window_33_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor<string, []>("window_33")];
-            tensor<int32, [3]> var_871_begin_0 = const()[name = tensor<string, []>("op_871_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_871_end_0 = const()[name = tensor<string, []>("op_871_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
-            tensor<bool, [3]> var_871_end_mask_0 = const()[name = tensor<string, []>("op_871_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor<string, []>("op_871")];
-            tensor<int32, [3]> var_874_begin_0 = const()[name = tensor<string, []>("op_874_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_874_end_0 = const()[name = tensor<string, []>("op_874_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_874_end_mask_0 = const()[name = tensor<string, []>("op_874_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor<string, []>("op_874")];
+            tensor<fp32, [1, 16, 256]> window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor<string, []>("window_33")];
+            tensor<int32, [3]> var_927_begin_0 = const()[name = tensor<string, []>("op_927_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_927_end_0 = const()[name = tensor<string, []>("op_927_end_0"), val = tensor<int32, [3]>([1, 2, 256])];
+            tensor<bool, [3]> var_927_end_mask_0 = const()[name = tensor<string, []>("op_927_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_21)[name = tensor<string, []>("op_927")];
+            tensor<int32, [3]> var_930_begin_0 = const()[name = tensor<string, []>("op_930_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_930_end_0 = const()[name = tensor<string, []>("op_930_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_930_end_mask_0 = const()[name = tensor<string, []>("op_930_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_930 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = window_33)[name = tensor<string, []>("op_930")];
             tensor<bool, []> window_35_interleave_0 = const()[name = tensor<string, []>("window_35_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor<string, []>("window_35")];
-            tensor<int32, [3]> var_879_begin_0 = const()[name = tensor<string, []>("op_879_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
-            tensor<int32, [3]> var_879_end_0 = const()[name = tensor<string, []>("op_879_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
-            tensor<bool, [3]> var_879_end_mask_0 = const()[name = tensor<string, []>("op_879_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
-            tensor<fp32, [1, 1, 256]> var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor<string, []>("op_879")];
-            tensor<int32, [3]> var_882_begin_0 = const()[name = tensor<string, []>("op_882_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_882_end_0 = const()[name = tensor<string, []>("op_882_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_882_end_mask_0 = const()[name = tensor<string, []>("op_882_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor<string, []>("op_882")];
+            tensor<fp32, [1, 16, 256]> window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor<string, []>("window_35")];
+            tensor<int32, [3]> var_935_begin_0 = const()[name = tensor<string, []>("op_935_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_935_end_0 = const()[name = tensor<string, []>("op_935_end_0"), val = tensor<int32, [3]>([1, 3, 256])];
+            tensor<bool, [3]> var_935_end_mask_0 = const()[name = tensor<string, []>("op_935_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 1, 256]> var_935 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, x = x_21)[name = tensor<string, []>("op_935")];
+            tensor<int32, [3]> var_938_begin_0 = const()[name = tensor<string, []>("op_938_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_938_end_0 = const()[name = tensor<string, []>("op_938_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_938_end_mask_0 = const()[name = tensor<string, []>("op_938_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_938 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = window_35)[name = tensor<string, []>("op_938")];
             tensor<bool, []> window_37_interleave_0 = const()[name = tensor<string, []>("window_37_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, 16, 256]> window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor<string, []>("window_37")];
-            tensor<int32, [3]> var_887_begin_0 = const()[name = tensor<string, []>("op_887_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
-            tensor<int32, [3]> var_887_end_0 = const()[name = tensor<string, []>("op_887_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<bool, [3]> var_887_end_mask_0 = const()[name = tensor<string, []>("op_887_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 1, 256]> var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor<string, []>("op_887")];
-            tensor<int32, [3]> var_890_begin_0 = const()[name = tensor<string, []>("op_890_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
-            tensor<int32, [3]> var_890_end_0 = const()[name = tensor<string, []>("op_890_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
-            tensor<bool, [3]> var_890_end_mask_0 = const()[name = tensor<string, []>("op_890_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 15, 256]> var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor<string, []>("op_890")];
+            tensor<fp32, [1, 16, 256]> window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor<string, []>("window_37")];
+            tensor<int32, [3]> var_943_begin_0 = const()[name = tensor<string, []>("op_943_begin_0"), val = tensor<int32, [3]>([0, 3, 0])];
+            tensor<int32, [3]> var_943_end_0 = const()[name = tensor<string, []>("op_943_end_0"), val = tensor<int32, [3]>([1, 1, 256])];
+            tensor<bool, [3]> var_943_end_mask_0 = const()[name = tensor<string, []>("op_943_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1, 256]> var_943 = slice_by_index(begin = var_943_begin_0, end = var_943_end_0, end_mask = var_943_end_mask_0, x = x_21)[name = tensor<string, []>("op_943")];
+            tensor<int32, [3]> var_946_begin_0 = const()[name = tensor<string, []>("op_946_begin_0"), val = tensor<int32, [3]>([0, 1, 0])];
+            tensor<int32, [3]> var_946_end_0 = const()[name = tensor<string, []>("op_946_end_0"), val = tensor<int32, [3]>([1, 16, 256])];
+            tensor<bool, [3]> var_946_end_mask_0 = const()[name = tensor<string, []>("op_946_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 15, 256]> var_946 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = window_37)[name = tensor<string, []>("op_946")];
             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_890, var_887))[name = tensor<string, []>("window")];
-            tensor<bool, []> input_141_interleave_0 = const()[name = tensor<string, []>("input_141_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [4, 16, 256]> input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor<string, []>("input_141")];
+            tensor<fp32, [1, 16, 256]> window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor<string, []>("window")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [4, 16, 256]> input_143 = concat(axis = var_68, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, 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, [4, 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, [4, 16, 256]> x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_65, 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, [4, 256, 16]> input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
-            tensor<fp32, [4, 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_915_split_sizes_0 = const()[name = tensor<string, []>("op_915_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
-            tensor<int32, []> var_915_axis_0 = const()[name = tensor<string, []>("op_915_axis_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [4, 256, 16]> var_915_0, tensor<fp32, [4, 256, 16]> var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_915")];
-            tensor<fp32, [4, 256, 16]> var_917 = sigmoid(x = var_915_1)[name = tensor<string, []>("op_917")];
-            tensor<fp32, [4, 256, 16]> inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor<string, []>("inputs_35")];
+            tensor<fp32, [4, 256, 16]> input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor<string, []>("transpose_33")];
+            tensor<fp32, [4, 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_971_split_sizes_0 = const()[name = tensor<string, []>("op_971_split_sizes_0"), val = tensor<int32, [2]>([256, 256])];
+            tensor<int32, []> var_971_axis_0 = const()[name = tensor<string, []>("op_971_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [4, 256, 16]> var_971_0, tensor<fp32, [4, 256, 16]> var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor<string, []>("op_971")];
+            tensor<fp32, [4, 256, 16]> var_973 = sigmoid(x = var_971_1)[name = tensor<string, []>("op_973")];
+            tensor<fp32, [4, 256, 16]> inputs_35 = mul(x = var_971_0, y = var_973)[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, [4, 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]>([4, 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, [4, 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, [4, 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, [4, 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, [4, 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, [4, 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]>([4, 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, [4, 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, [4, 256, 16]> inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_65, 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, [4, 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, [4, 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_948_begin_0 = const()[name = tensor<string, []>("op_948_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
-            tensor<int32, [3]> var_948_end_0 = const()[name = tensor<string, []>("op_948_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
-            tensor<bool, [3]> var_948_end_mask_0 = const()[name = tensor<string, []>("op_948_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [4, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor<string, []>("transpose_32")];
-            tensor<fp32, [4, 1, 256]> var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_948")];
-            tensor<int32, [3]> var_950_perm_0 = const()[name = tensor<string, []>("op_950_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [1, 4, 256]> var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor<string, []>("transpose_31")];
-            tensor<fp32, [1, 4, 256]> input_151 = add(x = x_21, y = var_950)[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, 4, 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, 4, 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, 4, 1024]> input_155 = silu(x = inputs)[name = tensor<string, []>("input_155")];
-            tensor<fp32, [1, 4, 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_973 = const()[name = tensor<string, []>("op_973"), val = tensor<fp32, []>(0x1p-1)];
-            tensor<fp32, [1, 4, 256]> var_974 = mul(x = input_159, y = var_973)[name = tensor<string, []>("op_974")];
-            tensor<fp32, [1, 4, 256]> input_161 = add(x = var_974, y = input_151)[name = tensor<string, []>("input_161")];
+            tensor<int32, [3]> var_1004_begin_0 = const()[name = tensor<string, []>("op_1004_begin_0"), val = tensor<int32, [3]>([0, -1, 0])];
+            tensor<int32, [3]> var_1004_end_0 = const()[name = tensor<string, []>("op_1004_end_0"), val = tensor<int32, [3]>([4, 16, 256])];
+            tensor<bool, [3]> var_1004_end_mask_0 = const()[name = tensor<string, []>("op_1004_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [4, 16, 256]> conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor<string, []>("transpose_32")];
+            tensor<fp32, [4, 1, 256]> var_1004 = slice_by_index(begin = var_1004_begin_0, end = var_1004_end_0, end_mask = var_1004_end_mask_0, x = conv_out_7)[name = tensor<string, []>("op_1004")];
+            tensor<int32, [3]> var_1006_perm_0 = const()[name = tensor<string, []>("op_1006_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp32, [1, 4, 256]> var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor<string, []>("transpose_31")];
+            tensor<fp32, [1, 4, 256]> input_153 = add(x = x_21, y = var_1006)[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, 4, 256]> input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor<string, []>("input_155")];
+            tensor<fp32, [1, 4, 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, 4, 1024]> input_157 = silu(x = inputs)[name = tensor<string, []>("input_157")];
+            tensor<fp32, [1, 4, 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_1029 = const()[name = tensor<string, []>("op_1029"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 4, 256]> var_1030 = mul(x = input_161, y = var_1029)[name = tensor<string, []>("op_1030")];
+            tensor<fp32, [1, 4, 256]> input_163 = add(x = var_1030, 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, 4, 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, 4, 256]> x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_65, 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, 4]> x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor<string, []>("transpose_30")];
-            tensor<fp32, [1, 256, 22]> cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor<string, []>("cat")];
+            tensor<fp32, [1, 256, 22]> cat = concat(axis = var_70, 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, 4]> 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_992_begin_0 = const()[name = tensor<string, []>("op_992_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
-            tensor<int32, [3]> var_992_end_0 = const()[name = tensor<string, []>("op_992_end_0"), val = tensor<int32, [3]>([1, 256, 22])];
-            tensor<bool, [3]> var_992_end_mask_0 = const()[name = tensor<string, []>("op_992_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
-            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor<string, []>("op_992")];
-            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_994 = const()[name = tensor<string, []>("op_994"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 4, 256]> input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
-            tensor<fp32, [1, 4, 1]> var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor<string, []>("op_995")];
+            tensor<fp32, [1, 256, 4]> 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_1048_begin_0 = const()[name = tensor<string, []>("op_1048_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
+            tensor<int32, [3]> var_1048_end_0 = const()[name = tensor<string, []>("op_1048_end_0"), val = tensor<int32, [3]>([1, 256, 22])];
+            tensor<bool, [3]> var_1048_end_mask_0 = const()[name = tensor<string, []>("op_1048_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 256, 18]> cnn_window_new = slice_by_index(begin = var_1048_begin_0, end = var_1048_end_0, end_mask = var_1048_end_mask_0, x = cat)[name = tensor<string, []>("op_1048")];
+            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_1050 = const()[name = tensor<string, []>("op_1050"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 4, 256]> input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor<string, []>("transpose_29")];
+            tensor<fp32, [1, 4, 1]> var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_64, x = input_165)[name = tensor<string, []>("op_1051")];
             tensor<fp32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 4, 1]> clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor<string, []>("clip_0")];
-            tensor<fp32, [1, 4, 256]> emb = real_div(x = input_163, y = clip_0)[name = tensor<string, []>("emb")];
-            tensor<int32, []> var_999_axis_0 = const()[name = tensor<string, []>("op_999_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor<string, []>("op_999")];
-            tensor<int32, []> var_1001_axis_0 = const()[name = tensor<string, []>("op_1001_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_1001")];
-            tensor<int32, []> var_1003_axis_0 = const()[name = tensor<string, []>("op_1003_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor<string, []>("op_1003")];
-            tensor<fp32, []> var_1012 = const()[name = tensor<string, []>("op_1012"), val = tensor<fp32, []>(0x1.5798eep-27)];
-            tensor<fp32, []> var_1017 = const()[name = tensor<string, []>("op_1017"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
-            tensor<fp32, []> var_1019 = const()[name = tensor<string, []>("op_1019"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
-            tensor<bool, []> var_1020 = const()[name = tensor<string, []>("op_1020"), val = tensor<bool, []>(true)];
-            tensor<fp32, []> var_1022 = const()[name = tensor<string, []>("op_1022"), val = tensor<fp32, []>(0x1p+0)];
-            tensor<int32, []> var_1026 = const()[name = tensor<string, []>("op_1026"), val = tensor<int32, []>(-1)];
-            tensor<int32, []> var_1032 = const()[name = tensor<string, []>("op_1032"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 4, 1]> clip_0 = clip(alpha = var_78, beta = const_12, x = var_1051)[name = tensor<string, []>("clip_0")];
+            tensor<fp32, [1, 4, 256]> emb = real_div(x = input_165, y = clip_0)[name = tensor<string, []>("emb")];
+            tensor<int32, []> var_1055_axis_0 = const()[name = tensor<string, []>("op_1055_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 4, 64, 64]> enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor<string, []>("op_1055")];
+            tensor<int32, []> var_1057_axis_0 = const()[name = tensor<string, []>("op_1057_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1]> enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor<string, []>("op_1057")];
+            tensor<int32, []> var_1059_axis_0 = const()[name = tensor<string, []>("op_1059_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [4, 1, 16, 256]> enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor<string, []>("op_1059")];
             tensor<fp32, [1, 4, 6, 256]> pos = const()[name = tensor<string, []>("pos"), val = tensor<fp32, [1, 4, 6, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44395584)))];
-            tensor<int32, [1]> var_1094_axes_0 = const()[name = tensor<string, []>("op_1094_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<fp32, [1, 4, 1, 256]> var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor<string, []>("op_1094")];
+            tensor<int32, [1]> var_1127_axes_0 = const()[name = tensor<string, []>("op_1127_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 4, 1, 256]> var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor<string, []>("op_1127")];
             tensor<int32, [4]> emb_exp_reps_0 = const()[name = tensor<string, []>("emb_exp_reps_0"), val = tensor<int32, [4]>([1, 1, 6, 1])];
-            tensor<fp32, [1, 4, 6, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[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, 4, 6, 512]> input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_165")];
-            tensor<fp32, [1, 4, 6, 256]> x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor<string, []>("linear_37")];
-            tensor<int32, [4]> var_1102_perm_0 = const()[name = tensor<string, []>("op_1102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1106 = const()[name = tensor<string, []>("op_1106"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [1, 6, 4, 256]> var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
-            tensor<fp32, [6, 4, 256]> x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor<string, []>("x_29")];
+            tensor<fp32, [1, 4, 6, 256]> emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[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, 4, 6, 512]> input_167 = concat(axis = var_71, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor<string, []>("input_167")];
+            tensor<fp32, [1, 4, 6, 256]> x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor<string, []>("linear_37")];
+            tensor<int32, [4]> var_1135_perm_0 = const()[name = tensor<string, []>("op_1135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1139 = const()[name = tensor<string, []>("op_1139"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [1, 6, 4, 256]> var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor<string, []>("transpose_28")];
+            tensor<fp32, [6, 4, 256]> x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor<string, []>("x_29")];
             tensor<int32, [5]> prev_kv_9_begin_0 = const()[name = tensor<string, []>("prev_kv_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_9_end_0 = const()[name = tensor<string, []>("prev_kv_9_end_0"), val = tensor<int32, [5]>([1, 6, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_9_end_mask_0 = const()[name = tensor<string, []>("prev_kv_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -983,132 +998,132 @@ program(1.0)
             tensor<bool, [2]> prev_scale_9_end_mask_0 = const()[name = tensor<string, []>("prev_scale_9_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_9_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_9_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor<string, []>("prev_scale_9")];
-            tensor<fp32, [6, 4, 256]> var_1114 = 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_1115 = const()[name = tensor<string, []>("op_1115"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor<string, []>("op_1116")];
+            tensor<fp32, [6, 4, 256]> var_1147 = 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_1148 = const()[name = tensor<string, []>("op_1148"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor<string, []>("op_1149")];
             tensor<int32, [4]> q_9_perm_0 = const()[name = tensor<string, []>("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_39")];
-            tensor<fp32, []> var_1121 = const()[name = tensor<string, []>("op_1121"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [6, 4, 256]> var_1122 = mul(x = var_1120, y = var_1121)[name = tensor<string, []>("op_1122")];
-            tensor<int32, [4]> var_1123 = const()[name = tensor<string, []>("op_1123"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor<string, []>("op_1124")];
+            tensor<fp32, [6, 4, 256]> var_1153 = 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_1154 = const()[name = tensor<string, []>("op_1154"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [6, 4, 256]> var_1155 = mul(x = var_1153, y = var_1154)[name = tensor<string, []>("op_1155")];
+            tensor<int32, [4]> var_1156 = const()[name = tensor<string, []>("op_1156"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor<string, []>("op_1157")];
             tensor<int32, [4]> k_9_perm_0 = const()[name = tensor<string, []>("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> var_1128 = 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_1129 = const()[name = tensor<string, []>("op_1129"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor<string, []>("op_1130")];
+            tensor<fp32, [6, 4, 256]> var_1161 = 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_1162 = const()[name = tensor<string, []>("op_1162"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor<string, []>("op_1163")];
             tensor<int32, [4]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor<string, []>("linear_41")];
+            tensor<fp32, [6, 4, 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, [4]> cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor<string, []>("cumsum_mask_1")];
+            tensor<fp32, [4]> cumsum_mask_1 = cumsum(axis = var_68, 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, [4]> 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, [4]> clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
+            tensor<fp32, [4]> clip_1 = clip(alpha = var_58, beta = const_20, x = s_t_9)[name = tensor<string, []>("clip_1")];
             tensor<fp32, [4]> sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor<string, []>("sqrt_s_t_9")];
             tensor<bool, []> qk_9_transpose_x_1 = const()[name = tensor<string, []>("qk_9_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_9_transpose_y_1 = const()[name = tensor<string, []>("qk_9_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [6, 4, 4, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor<string, []>("transpose_26")];
-            tensor<fp32, [6, 4, 4, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor<string, []>("transpose_27")];
+            tensor<fp32, [6, 4, 4, 64]> k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor<string, []>("transpose_26")];
+            tensor<fp32, [6, 4, 4, 64]> q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor<string, []>("transpose_27")];
             tensor<fp32, [6, 4, 4, 4]> 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_1142 = const()[name = tensor<string, []>("op_1142"), val = tensor<int32, [2]>([1, 4])];
-            tensor<fp32, [1, 4]> var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor<string, []>("op_1143")];
-            tensor<fp32, [4, 4]> causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor<string, []>("causal_with_valid_1")];
-            tensor<int32, [2]> var_1145 = const()[name = tensor<string, []>("op_1145"), val = tensor<int32, [2]>([4, 1])];
-            tensor<fp32, [4, 1]> var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor<string, []>("op_1146")];
-            tensor<fp32, [4, 4]> M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor<string, []>("M_9")];
-            tensor<fp32, [6, 4, 4, 4]> var_1148 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1148")];
+            tensor<int32, [2]> var_1175 = const()[name = tensor<string, []>("op_1175"), val = tensor<int32, [2]>([1, 4])];
+            tensor<fp32, [1, 4]> var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor<string, []>("op_1176")];
+            tensor<fp32, [4, 4]> causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor<string, []>("causal_with_valid_1")];
+            tensor<int32, [2]> var_1178 = const()[name = tensor<string, []>("op_1178"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor<string, []>("op_1179")];
+            tensor<fp32, [4, 4]> M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor<string, []>("M_9")];
+            tensor<fp32, [6, 4, 4, 4]> var_1181 = mul(x = qk_9, y = M_9)[name = tensor<string, []>("op_1181")];
             tensor<bool, []> inner_9_transpose_x_0 = const()[name = tensor<string, []>("inner_9_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> inner_9_transpose_y_0 = const()[name = tensor<string, []>("inner_9_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 4, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor<string, []>("transpose_25")];
-            tensor<fp32, [6, 4, 4, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor<string, []>("inner_9")];
-            tensor<bool, []> var_1150_transpose_x_0 = const()[name = tensor<string, []>("op_1150_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1150_transpose_y_0 = const()[name = tensor<string, []>("op_1150_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 4, 64]> var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1150")];
-            tensor<fp32, [4]> var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1151")];
-            tensor<int32, [4]> var_1152 = const()[name = tensor<string, []>("op_1152"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor<string, []>("op_1153")];
-            tensor<fp32, [6, 4, 4, 64]> cross_9 = mul(x = var_1150, y = var_1153)[name = tensor<string, []>("cross_9")];
+            tensor<fp32, [6, 4, 4, 64]> v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor<string, []>("transpose_25")];
+            tensor<fp32, [6, 4, 4, 64]> inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor<string, []>("inner_9")];
+            tensor<bool, []> var_1183_transpose_x_0 = const()[name = tensor<string, []>("op_1183_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1183_transpose_y_0 = const()[name = tensor<string, []>("op_1183_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 4, 64]> var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor<string, []>("op_1183")];
+            tensor<fp32, [4]> var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor<string, []>("op_1184")];
+            tensor<int32, [4]> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor<string, []>("op_1186")];
+            tensor<fp32, [6, 4, 4, 64]> cross_9 = mul(x = var_1183, y = var_1186)[name = tensor<string, []>("cross_9")];
             tensor<fp32, [6, 4, 4, 64]> out_25 = add(x = inner_9, y = cross_9)[name = tensor<string, []>("out_25")];
-            tensor<int32, [4]> var_1156 = const()[name = tensor<string, []>("op_1156"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor<string, []>("op_1157")];
-            tensor<fp32, [6, 4, 4, 64]> v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor<string, []>("v_masked_1")];
-            tensor<fp32, [6, 4, 64, 64]> var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1159")];
-            tensor<bool, []> var_1161_transpose_x_1 = const()[name = tensor<string, []>("op_1161_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1161_transpose_y_1 = const()[name = tensor<string, []>("op_1161_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 64, 64]> var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1161")];
-            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor<string, []>("new_kv_unnorm_9")];
-            tensor<bool, []> var_1163_keep_dims_0 = const()[name = tensor<string, []>("op_1163_keep_dims_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, []> var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1163")];
-            tensor<int32, [1]> var_1164 = const()[name = tensor<string, []>("op_1164"), val = tensor<int32, [1]>([1])];
-            tensor<fp32, [1]> var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor<string, []>("op_1165")];
-            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor<string, []>("new_scale_9")];
+            tensor<int32, [4]> var_1189 = const()[name = tensor<string, []>("op_1189"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor<string, []>("op_1190")];
+            tensor<fp32, [6, 4, 4, 64]> v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor<string, []>("v_masked_1")];
+            tensor<fp32, [6, 4, 64, 64]> var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor<string, []>("op_1192")];
+            tensor<bool, []> var_1194_transpose_x_1 = const()[name = tensor<string, []>("op_1194_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1194_transpose_y_1 = const()[name = tensor<string, []>("op_1194_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 64, 64]> var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor<string, []>("op_1194")];
+            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor<string, []>("new_kv_unnorm_9")];
+            tensor<bool, []> var_1196_keep_dims_0 = const()[name = tensor<string, []>("op_1196_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, []> var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor<string, []>("op_1196")];
+            tensor<int32, [1]> var_1197 = const()[name = tensor<string, []>("op_1197"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1]> var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor<string, []>("op_1198")];
+            tensor<fp32, [1]> new_scale_9 = add(x = prev_scale_9, y = var_1198)[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_1022, beta = const_21, x = new_scale_9)[name = tensor<string, []>("clip_2")];
+            tensor<fp32, [1]> clip_2 = clip(alpha = var_58, beta = const_21, x = new_scale_9)[name = tensor<string, []>("clip_2")];
             tensor<fp32, [1]> sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor<string, []>("sqrt_new_scale_1")];
-            tensor<fp32, [6, 4, 64, 64]> var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1169")];
-            tensor<int32, [4]> var_1170_perm_0 = const()[name = tensor<string, []>("op_1170_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<fp32, [6, 4, 64, 64]> var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor<string, []>("op_1202")];
+            tensor<int32, [4]> var_1203_perm_0 = const()[name = tensor<string, []>("op_1203_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_27_axes_0 = const()[name = tensor<string, []>("out_27_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 4, 4, 64]> var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
-            tensor<fp32, [6, 4, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor<string, []>("out_27")];
-            tensor<int32, [3]> var_1174 = const()[name = tensor<string, []>("op_1174"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [6, 4, 256]> out_29 = reshape(shape = var_1174, x = out_27)[name = tensor<string, []>("out_29")];
-            tensor<fp32, [6, 4, 256]> var_1176 = silu(x = input_169)[name = tensor<string, []>("op_1176")];
-            tensor<fp32, [6, 4, 256]> input_171 = mul(x = var_1176, y = out_29)[name = tensor<string, []>("input_171")];
-            tensor<fp32, [6, 4, 256]> ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor<string, []>("linear_42")];
-            tensor<fp32, [6, 4, 256]> input_173 = add(x = x_29, y = ret_out_9)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [6, 4, 4, 64]> var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor<string, []>("transpose_24")];
+            tensor<fp32, [6, 4, 4, 64]> out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_73, x = var_1203)[name = tensor<string, []>("out_27")];
+            tensor<int32, [3]> var_1207 = const()[name = tensor<string, []>("op_1207"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [6, 4, 256]> out_29 = reshape(shape = var_1207, x = out_27)[name = tensor<string, []>("out_29")];
+            tensor<fp32, [6, 4, 256]> var_1209 = silu(x = input_171)[name = tensor<string, []>("op_1209")];
+            tensor<fp32, [6, 4, 256]> input_173 = mul(x = var_1209, y = out_29)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [6, 4, 256]> ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor<string, []>("linear_42")];
+            tensor<fp32, [6, 4, 256]> input_175 = add(x = x_29, y = ret_out_9)[name = tensor<string, []>("input_175")];
             tensor<int32, [1]> xt_1_axes_0 = const()[name = tensor<string, []>("xt_1_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 4, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor<string, []>("xt_1")];
-            tensor<int32, [4]> var_1186 = const()[name = tensor<string, []>("op_1186"), val = tensor<int32, [4]>([1, 6, 4, 256])];
-            tensor<fp32, [1, 6, 4, 256]> var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor<string, []>("op_1187")];
-            tensor<int32, [4]> var_1188_perm_0 = const()[name = tensor<string, []>("op_1188_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1191 = const()[name = tensor<string, []>("op_1191"), val = tensor<int32, [3]>([4, 6, 256])];
-            tensor<fp32, [1, 4, 6, 256]> var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor<string, []>("transpose_23")];
-            tensor<fp32, [4, 6, 256]> query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor<string, []>("query_1")];
+            tensor<fp32, [6, 4, 256]> xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_65, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor<string, []>("xt_1")];
+            tensor<int32, [4]> var_1219 = const()[name = tensor<string, []>("op_1219"), val = tensor<int32, [4]>([1, 6, 4, 256])];
+            tensor<fp32, [1, 6, 4, 256]> var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor<string, []>("op_1220")];
+            tensor<int32, [4]> var_1221_perm_0 = const()[name = tensor<string, []>("op_1221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1224 = const()[name = tensor<string, []>("op_1224"), val = tensor<int32, [3]>([4, 6, 256])];
+            tensor<fp32, [1, 4, 6, 256]> var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor<string, []>("transpose_23")];
+            tensor<fp32, [4, 6, 256]> query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor<string, []>("query_1")];
             tensor<int32, [3]> query_3_perm_0 = const()[name = tensor<string, []>("query_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [6, 4, 256]> query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor<string, []>("transpose_22")];
-            tensor<fp32, [6, 4, 768]> var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor<string, []>("linear_43")];
+            tensor<fp32, [6, 4, 768]> var_1247 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor<string, []>("linear_43")];
             tensor<int32, [4]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<int32, [4]>([6, 4, 3, 256])];
-            tensor<fp32, [6, 4, 3, 256]> var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor<string, []>("op_1216")];
-            tensor<int32, [1]> var_1217_axes_0 = const()[name = tensor<string, []>("op_1217_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 6, 4, 3, 256]> var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor<string, []>("op_1217")];
-            tensor<int32, [5]> var_1218_perm_0 = const()[name = tensor<string, []>("op_1218_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1219_axes_0 = const()[name = tensor<string, []>("op_1219_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 6, 4, 1, 256]> var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor<string, []>("transpose_21")];
-            tensor<fp32, [3, 6, 4, 256]> var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor<string, []>("op_1219")];
+            tensor<fp32, [6, 4, 3, 256]> var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor<string, []>("op_1249")];
+            tensor<int32, [1]> var_1250_axes_0 = const()[name = tensor<string, []>("op_1250_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 6, 4, 3, 256]> var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor<string, []>("op_1250")];
+            tensor<int32, [5]> var_1251_perm_0 = const()[name = tensor<string, []>("op_1251_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1252_axes_0 = const()[name = tensor<string, []>("op_1252_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 6, 4, 1, 256]> var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor<string, []>("transpose_21")];
+            tensor<fp32, [3, 6, 4, 256]> var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor<string, []>("op_1252")];
             tensor<int32, [4]> q_11_begin_0 = const()[name = tensor<string, []>("q_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [4]> q_11_end_0 = const()[name = tensor<string, []>("q_11_end_0"), val = tensor<int32, [4]>([1, 6, 4, 256])];
             tensor<bool, [4]> q_11_end_mask_0 = const()[name = tensor<string, []>("q_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> q_11_squeeze_mask_0 = const()[name = tensor<string, []>("q_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1219)[name = tensor<string, []>("q_11")];
+            tensor<fp32, [6, 4, 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_1252)[name = tensor<string, []>("q_11")];
             tensor<int32, [4]> k_11_begin_0 = const()[name = tensor<string, []>("k_11_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> k_11_end_0 = const()[name = tensor<string, []>("k_11_end_0"), val = tensor<int32, [4]>([2, 6, 4, 256])];
             tensor<bool, [4]> k_11_end_mask_0 = const()[name = tensor<string, []>("k_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> k_11_squeeze_mask_0 = const()[name = tensor<string, []>("k_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1219)[name = tensor<string, []>("k_11")];
+            tensor<fp32, [6, 4, 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_1252)[name = tensor<string, []>("k_11")];
             tensor<int32, [4]> v_11_begin_0 = const()[name = tensor<string, []>("v_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> v_11_end_0 = const()[name = tensor<string, []>("v_11_end_0"), val = tensor<int32, [4]>([3, 6, 4, 256])];
             tensor<bool, [4]> v_11_end_mask_0 = const()[name = tensor<string, []>("v_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> v_11_squeeze_mask_0 = const()[name = tensor<string, []>("v_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1219)[name = tensor<string, []>("v_11")];
-            tensor<int32, [3]> var_1227 = const()[name = tensor<string, []>("op_1227"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor<string, []>("op_1228")];
+            tensor<fp32, [6, 4, 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_1252)[name = tensor<string, []>("v_11")];
+            tensor<int32, [3]> var_1260 = const()[name = tensor<string, []>("op_1260"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor<string, []>("op_1261")];
             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_1234 = const()[name = tensor<string, []>("op_1234"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor<string, []>("op_1235")];
+            tensor<int32, [3]> var_1267 = const()[name = tensor<string, []>("op_1267"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor<string, []>("op_1268")];
             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_1241 = const()[name = tensor<string, []>("op_1241"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor<string, []>("op_1242")];
+            tensor<int32, [3]> var_1274 = const()[name = tensor<string, []>("op_1274"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor<string, []>("op_1275")];
             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_1245 = const()[name = tensor<string, []>("op_1245"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor<string, []>("transpose_20")];
-            tensor<fp32, [4, 4, 6, 64]> q_15 = reshape(shape = var_1245, x = q_13)[name = tensor<string, []>("q_15")];
-            tensor<int32, [4]> var_1247 = const()[name = tensor<string, []>("op_1247"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor<string, []>("transpose_19")];
-            tensor<fp32, [4, 4, 6, 64]> k_15 = reshape(shape = var_1247, x = k_13)[name = tensor<string, []>("k_15")];
-            tensor<int32, [4]> var_1249 = const()[name = tensor<string, []>("op_1249"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor<string, []>("transpose_18")];
-            tensor<fp32, [4, 4, 6, 64]> v_15 = reshape(shape = var_1249, x = v_13)[name = tensor<string, []>("v_15")];
+            tensor<int32, [4]> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor<string, []>("transpose_20")];
+            tensor<fp32, [4, 4, 6, 64]> q_15 = reshape(shape = var_1278, x = q_13)[name = tensor<string, []>("q_15")];
+            tensor<int32, [4]> var_1280 = const()[name = tensor<string, []>("op_1280"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor<string, []>("transpose_19")];
+            tensor<fp32, [4, 4, 6, 64]> k_15 = reshape(shape = var_1280, x = k_13)[name = tensor<string, []>("k_15")];
+            tensor<int32, [4]> var_1282 = const()[name = tensor<string, []>("op_1282"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor<string, []>("transpose_18")];
+            tensor<fp32, [4, 4, 6, 64]> v_15 = reshape(shape = var_1282, 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, [4, 4, 6, 64]> mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor<string, []>("mul_1")];
             tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
@@ -1119,30 +1134,30 @@ 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, [4, 4, 6, 64]> attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor<string, []>("attn_output_1")];
-            tensor<int32, [4]> var_1252 = const()[name = tensor<string, []>("op_1252"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1257 = const()[name = tensor<string, []>("op_1257"), val = tensor<int32, [2]>([24, 256])];
-            tensor<fp32, [6, 4, 4, 64]> var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
-            tensor<fp32, [24, 256]> attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor<string, []>("attn_output_3")];
-            tensor<fp32, [24, 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_1261 = const()[name = tensor<string, []>("op_1261"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [6, 4, 256]> attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor<string, []>("attn_output_7")];
+            tensor<int32, [4]> var_1285 = const()[name = tensor<string, []>("op_1285"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1290 = const()[name = tensor<string, []>("op_1290"), val = tensor<int32, [2]>([24, 256])];
+            tensor<fp32, [6, 4, 4, 64]> var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor<string, []>("transpose_17")];
+            tensor<fp32, [24, 256]> attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor<string, []>("attn_output_3")];
+            tensor<fp32, [24, 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_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [6, 4, 256]> attn_output_7 = reshape(shape = var_1294, 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, [4, 6, 256]> sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor<string, []>("transpose_16")];
-            tensor<fp32, [4, 6, 256]> input_175 = add(x = query_1, y = sa2_out_1)[name = tensor<string, []>("input_175")];
-            tensor<int32, [1]> input_177_axes_0 = const()[name = tensor<string, []>("input_177_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [4, 6, 256]> input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor<string, []>("input_177")];
-            tensor<fp32, [4, 6, 2048]> input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor<string, []>("linear_45")];
-            tensor<fp32, [4, 6, 2048]> input_181 = relu(x = input_179)[name = tensor<string, []>("input_181")];
-            tensor<fp32, [4, 6, 256]> ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor<string, []>("linear_46")];
-            tensor<fp32, [4, 6, 256]> input_183 = add(x = input_177, y = ffn_out_1)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [4, 6, 256]> input_177 = add(x = query_1, y = sa2_out_1)[name = tensor<string, []>("input_177")];
+            tensor<int32, [1]> input_179_axes_0 = const()[name = tensor<string, []>("input_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [4, 6, 256]> input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_65, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor<string, []>("input_179")];
+            tensor<fp32, [4, 6, 2048]> input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor<string, []>("linear_45")];
+            tensor<fp32, [4, 6, 2048]> input_183 = relu(x = input_181)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [4, 6, 256]> ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor<string, []>("linear_46")];
+            tensor<fp32, [4, 6, 256]> input_185 = add(x = input_179, y = ffn_out_1)[name = tensor<string, []>("input_185")];
             tensor<int32, [1]> xt_3_axes_0 = const()[name = tensor<string, []>("xt_3_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [4, 6, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor<string, []>("xt_3")];
-            tensor<int32, [4]> var_1281 = const()[name = tensor<string, []>("op_1281"), val = tensor<int32, [4]>([1, 4, 6, 256])];
-            tensor<fp32, [1, 4, 6, 256]> x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor<string, []>("x_31")];
-            tensor<int32, [4]> var_1283_perm_0 = const()[name = tensor<string, []>("op_1283_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1287 = const()[name = tensor<string, []>("op_1287"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [1, 6, 4, 256]> var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
-            tensor<fp32, [6, 4, 256]> x = reshape(shape = var_1287, x = var_1283)[name = tensor<string, []>("x")];
+            tensor<fp32, [4, 6, 256]> xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_65, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor<string, []>("xt_3")];
+            tensor<int32, [4]> var_1314 = const()[name = tensor<string, []>("op_1314"), val = tensor<int32, [4]>([1, 4, 6, 256])];
+            tensor<fp32, [1, 4, 6, 256]> x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor<string, []>("x_31")];
+            tensor<int32, [4]> var_1316_perm_0 = const()[name = tensor<string, []>("op_1316_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1320 = const()[name = tensor<string, []>("op_1320"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [1, 6, 4, 256]> var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor<string, []>("transpose_15")];
+            tensor<fp32, [6, 4, 256]> x = reshape(shape = var_1320, x = var_1316)[name = tensor<string, []>("x")];
             tensor<int32, [5]> prev_kv_begin_0 = const()[name = tensor<string, []>("prev_kv_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
             tensor<int32, [5]> prev_kv_end_0 = const()[name = tensor<string, []>("prev_kv_end_0"), val = tensor<int32, [5]>([2, 6, 4, 64, 64])];
             tensor<bool, [5]> prev_kv_end_mask_0 = const()[name = tensor<string, []>("prev_kv_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
@@ -1153,120 +1168,120 @@ program(1.0)
             tensor<bool, [2]> prev_scale_end_mask_0 = const()[name = tensor<string, []>("prev_scale_end_mask_0"), val = tensor<bool, [2]>([false, true])];
             tensor<bool, [2]> prev_scale_squeeze_mask_0 = const()[name = tensor<string, []>("prev_scale_squeeze_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<fp32, [1]> prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor<string, []>("prev_scale")];
-            tensor<fp32, [6, 4, 256]> var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor<string, []>("linear_47")];
-            tensor<int32, [4]> var_1296 = const()[name = tensor<string, []>("op_1296"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor<string, []>("op_1297")];
+            tensor<fp32, [6, 4, 256]> var_1328 = 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_1329 = const()[name = tensor<string, []>("op_1329"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor<string, []>("op_1330")];
             tensor<int32, [4]> q_17_perm_0 = const()[name = tensor<string, []>("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor<string, []>("linear_48")];
-            tensor<fp32, []> var_1302 = const()[name = tensor<string, []>("op_1302"), val = tensor<fp32, []>(0x1p-3)];
-            tensor<fp32, [6, 4, 256]> var_1303 = mul(x = var_1301, y = var_1302)[name = tensor<string, []>("op_1303")];
-            tensor<int32, [4]> var_1304 = const()[name = tensor<string, []>("op_1304"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor<string, []>("op_1305")];
+            tensor<fp32, [6, 4, 256]> var_1334 = 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_1335 = const()[name = tensor<string, []>("op_1335"), val = tensor<fp32, []>(0x1p-3)];
+            tensor<fp32, [6, 4, 256]> var_1336 = mul(x = var_1334, y = var_1335)[name = tensor<string, []>("op_1336")];
+            tensor<int32, [4]> var_1337 = const()[name = tensor<string, []>("op_1337"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor<string, []>("op_1338")];
             tensor<int32, [4]> k_17_perm_0 = const()[name = tensor<string, []>("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor<string, []>("linear_49")];
-            tensor<int32, [4]> var_1310 = const()[name = tensor<string, []>("op_1310"), val = tensor<int32, [4]>([6, 4, 4, 64])];
-            tensor<fp32, [6, 4, 4, 64]> var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor<string, []>("op_1311")];
+            tensor<fp32, [6, 4, 256]> var_1342 = 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_1343 = const()[name = tensor<string, []>("op_1343"), val = tensor<int32, [4]>([6, 4, 4, 64])];
+            tensor<fp32, [6, 4, 4, 64]> var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor<string, []>("op_1344")];
             tensor<int32, [4]> v_17_perm_0 = const()[name = tensor<string, []>("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<fp32, [6, 4, 256]> input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor<string, []>("linear_50")];
+            tensor<fp32, [6, 4, 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, [4]> 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, [4]> clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
+            tensor<fp32, [4]> clip_3 = clip(alpha = var_58, beta = const_32, x = s_t)[name = tensor<string, []>("clip_3")];
             tensor<fp32, [4]> sqrt_s_t = sqrt(x = clip_3)[name = tensor<string, []>("sqrt_s_t")];
             tensor<bool, []> qk_transpose_x_1 = const()[name = tensor<string, []>("qk_transpose_x_1"), val = tensor<bool, []>(false)];
             tensor<bool, []> qk_transpose_y_1 = const()[name = tensor<string, []>("qk_transpose_y_1"), val = tensor<bool, []>(true)];
-            tensor<fp32, [6, 4, 4, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor<string, []>("transpose_13")];
-            tensor<fp32, [6, 4, 4, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor<string, []>("transpose_14")];
+            tensor<fp32, [6, 4, 4, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor<string, []>("transpose_13")];
+            tensor<fp32, [6, 4, 4, 64]> q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor<string, []>("transpose_14")];
             tensor<fp32, [6, 4, 4, 4]> 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_1326 = const()[name = tensor<string, []>("op_1326"), val = tensor<int32, [2]>([4, 1])];
-            tensor<fp32, [4, 1]> var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor<string, []>("op_1327")];
-            tensor<fp32, [4, 4]> M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor<string, []>("M")];
-            tensor<fp32, [6, 4, 4, 4]> var_1329 = mul(x = qk, y = M)[name = tensor<string, []>("op_1329")];
-            tensor<bool, []> inner_transpose_x_0 = const()[name = tensor<string, []>("inner_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> inner_transpose_y_0 = const()[name = tensor<string, []>("inner_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 4, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor<string, []>("transpose_12")];
-            tensor<fp32, [6, 4, 4, 64]> inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor<string, []>("inner")];
-            tensor<bool, []> var_1331_transpose_x_0 = const()[name = tensor<string, []>("op_1331_transpose_x_0"), val = tensor<bool, []>(false)];
-            tensor<bool, []> var_1331_transpose_y_0 = const()[name = tensor<string, []>("op_1331_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 4, 64]> var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1331")];
-            tensor<fp32, [4]> var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1332")];
-            tensor<int32, [4]> var_1333 = const()[name = tensor<string, []>("op_1333"), val = tensor<int32, [4]>([1, 1, 4, 1])];
-            tensor<fp32, [1, 1, 4, 1]> var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor<string, []>("op_1334")];
-            tensor<fp32, [6, 4, 4, 64]> cross = mul(x = var_1331, y = var_1334)[name = tensor<string, []>("cross")];
-            tensor<fp32, [6, 4, 4, 64]> out_31 = add(x = inner, y = cross)[name = tensor<string, []>("out_31")];
-            tensor<fp32, [6, 4, 4, 64]> v_masked = mul(x = v_17, y = var_1157)[name = tensor<string, []>("v_masked")];
-            tensor<fp32, [6, 4, 64, 64]> var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1340")];
-            tensor<bool, []> var_1342_transpose_x_1 = const()[name = tensor<string, []>("op_1342_transpose_x_1"), val = tensor<bool, []>(true)];
-            tensor<bool, []> var_1342_transpose_y_1 = const()[name = tensor<string, []>("op_1342_transpose_y_1"), val = tensor<bool, []>(false)];
-            tensor<fp32, [6, 4, 64, 64]> var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1342")];
-            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor<string, []>("new_kv_unnorm")];
-            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1165)[name = tensor<string, []>("new_scale")];
+            tensor<int32, [2]> var_1359 = const()[name = tensor<string, []>("op_1359"), val = tensor<int32, [2]>([4, 1])];
+            tensor<fp32, [4, 1]> var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor<string, []>("op_1360")];
+            tensor<fp32, [4, 4]> M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor<string, []>("M")];
+            tensor<fp32, [6, 4, 4, 4]> var_1362 = mul(x = qk, y = M)[name = tensor<string, []>("op_1362")];
+            tensor<bool, []> inner_11_transpose_x_0 = const()[name = tensor<string, []>("inner_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> inner_11_transpose_y_0 = const()[name = tensor<string, []>("inner_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 4, 64]> v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor<string, []>("transpose_12")];
+            tensor<fp32, [6, 4, 4, 64]> inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1362, y = v_17)[name = tensor<string, []>("inner_11")];
+            tensor<bool, []> var_1364_transpose_x_0 = const()[name = tensor<string, []>("op_1364_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_1364_transpose_y_0 = const()[name = tensor<string, []>("op_1364_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 4, 64]> var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor<string, []>("op_1364")];
+            tensor<fp32, [4]> var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor<string, []>("op_1365")];
+            tensor<int32, [4]> var_1366 = const()[name = tensor<string, []>("op_1366"), val = tensor<int32, [4]>([1, 1, 4, 1])];
+            tensor<fp32, [1, 1, 4, 1]> var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor<string, []>("op_1367")];
+            tensor<fp32, [6, 4, 4, 64]> cross = mul(x = var_1364, y = var_1367)[name = tensor<string, []>("cross")];
+            tensor<fp32, [6, 4, 4, 64]> out_31 = add(x = inner_11, y = cross)[name = tensor<string, []>("out_31")];
+            tensor<fp32, [6, 4, 4, 64]> v_masked = mul(x = v_17, y = var_1190)[name = tensor<string, []>("v_masked")];
+            tensor<fp32, [6, 4, 64, 64]> var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor<string, []>("op_1373")];
+            tensor<bool, []> var_1375_transpose_x_1 = const()[name = tensor<string, []>("op_1375_transpose_x_1"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_1375_transpose_y_1 = const()[name = tensor<string, []>("op_1375_transpose_y_1"), val = tensor<bool, []>(false)];
+            tensor<fp32, [6, 4, 64, 64]> var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor<string, []>("op_1375")];
+            tensor<fp32, [6, 4, 64, 64]> new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor<string, []>("new_kv_unnorm")];
+            tensor<fp32, [1]> new_scale = add(x = prev_scale, y = var_1198)[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_1022, beta = const_33, x = new_scale)[name = tensor<string, []>("clip_4")];
+            tensor<fp32, [1]> clip_4 = clip(alpha = var_58, beta = const_33, x = new_scale)[name = tensor<string, []>("clip_4")];
             tensor<fp32, [1]> sqrt_new_scale = sqrt(x = clip_4)[name = tensor<string, []>("sqrt_new_scale")];
             tensor<fp32, [6, 4, 64, 64]> nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor<string, []>("nkv")];
-            tensor<int32, [4]> var_1351_perm_0 = const()[name = tensor<string, []>("op_1351_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_1384_perm_0 = const()[name = tensor<string, []>("op_1384_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
             tensor<int32, [1]> out_33_axes_0 = const()[name = tensor<string, []>("out_33_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 4, 4, 64]> var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
-            tensor<fp32, [6, 4, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor<string, []>("out_33")];
-            tensor<int32, [3]> var_1355 = const()[name = tensor<string, []>("op_1355"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [6, 4, 256]> out = reshape(shape = var_1355, x = out_33)[name = tensor<string, []>("out")];
-            tensor<fp32, [6, 4, 256]> var_1357 = silu(x = input_187)[name = tensor<string, []>("op_1357")];
-            tensor<fp32, [6, 4, 256]> input_189 = mul(x = var_1357, y = out)[name = tensor<string, []>("input_189")];
-            tensor<fp32, [6, 4, 256]> ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor<string, []>("linear_51")];
-            tensor<fp32, [6, 4, 256]> input_191 = add(x = x, y = ret_out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [6, 4, 4, 64]> var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor<string, []>("transpose_11")];
+            tensor<fp32, [6, 4, 4, 64]> out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_73, x = var_1384)[name = tensor<string, []>("out_33")];
+            tensor<int32, [3]> var_1388 = const()[name = tensor<string, []>("op_1388"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [6, 4, 256]> out = reshape(shape = var_1388, x = out_33)[name = tensor<string, []>("out")];
+            tensor<fp32, [6, 4, 256]> var_1390 = silu(x = input_189)[name = tensor<string, []>("op_1390")];
+            tensor<fp32, [6, 4, 256]> input_191 = mul(x = var_1390, y = out)[name = tensor<string, []>("input_191")];
+            tensor<fp32, [6, 4, 256]> ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor<string, []>("linear_51")];
+            tensor<fp32, [6, 4, 256]> input_193 = add(x = x, y = ret_out)[name = tensor<string, []>("input_193")];
             tensor<int32, [1]> xt_5_axes_0 = const()[name = tensor<string, []>("xt_5_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [6, 4, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor<string, []>("xt_5")];
-            tensor<int32, [4]> var_1367 = const()[name = tensor<string, []>("op_1367"), val = tensor<int32, [4]>([1, 6, 4, 256])];
-            tensor<fp32, [1, 6, 4, 256]> var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor<string, []>("op_1368")];
-            tensor<int32, [4]> var_1369_perm_0 = const()[name = tensor<string, []>("op_1369_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1372 = const()[name = tensor<string, []>("op_1372"), val = tensor<int32, [3]>([4, 6, 256])];
-            tensor<fp32, [1, 4, 6, 256]> var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor<string, []>("transpose_10")];
-            tensor<fp32, [4, 6, 256]> query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor<string, []>("query_5")];
+            tensor<fp32, [6, 4, 256]> xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_65, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor<string, []>("xt_5")];
+            tensor<int32, [4]> var_1400 = const()[name = tensor<string, []>("op_1400"), val = tensor<int32, [4]>([1, 6, 4, 256])];
+            tensor<fp32, [1, 6, 4, 256]> var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor<string, []>("op_1401")];
+            tensor<int32, [4]> var_1402_perm_0 = const()[name = tensor<string, []>("op_1402_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1405 = const()[name = tensor<string, []>("op_1405"), val = tensor<int32, [3]>([4, 6, 256])];
+            tensor<fp32, [1, 4, 6, 256]> var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor<string, []>("transpose_10")];
+            tensor<fp32, [4, 6, 256]> query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor<string, []>("query_5")];
             tensor<int32, [3]> query_perm_0 = const()[name = tensor<string, []>("query_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<fp32, [6, 4, 256]> query = transpose(perm = query_perm_0, x = query_5)[name = tensor<string, []>("transpose_9")];
-            tensor<fp32, [6, 4, 768]> var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor<string, []>("linear_52")];
+            tensor<fp32, [6, 4, 768]> var_1428 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor<string, []>("linear_52")];
             tensor<int32, [4]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [4]>([6, 4, 3, 256])];
-            tensor<fp32, [6, 4, 3, 256]> var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor<string, []>("op_1397")];
-            tensor<int32, [1]> var_1398_axes_0 = const()[name = tensor<string, []>("op_1398_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [1, 6, 4, 3, 256]> var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor<string, []>("op_1398")];
-            tensor<int32, [5]> var_1399_perm_0 = const()[name = tensor<string, []>("op_1399_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
-            tensor<int32, [1]> var_1400_axes_0 = const()[name = tensor<string, []>("op_1400_axes_0"), val = tensor<int32, [1]>([-2])];
-            tensor<fp32, [3, 6, 4, 1, 256]> var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor<string, []>("transpose_8")];
-            tensor<fp32, [3, 6, 4, 256]> var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor<string, []>("op_1400")];
+            tensor<fp32, [6, 4, 3, 256]> var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor<string, []>("op_1430")];
+            tensor<int32, [1]> var_1431_axes_0 = const()[name = tensor<string, []>("op_1431_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 6, 4, 3, 256]> var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor<string, []>("op_1431")];
+            tensor<int32, [5]> var_1432_perm_0 = const()[name = tensor<string, []>("op_1432_perm_0"), val = tensor<int32, [5]>([-2, 1, 2, 0, 4])];
+            tensor<int32, [1]> var_1433_axes_0 = const()[name = tensor<string, []>("op_1433_axes_0"), val = tensor<int32, [1]>([-2])];
+            tensor<fp32, [3, 6, 4, 1, 256]> var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor<string, []>("transpose_8")];
+            tensor<fp32, [3, 6, 4, 256]> var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor<string, []>("op_1433")];
             tensor<int32, [4]> q_19_begin_0 = const()[name = tensor<string, []>("q_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
             tensor<int32, [4]> q_19_end_0 = const()[name = tensor<string, []>("q_19_end_0"), val = tensor<int32, [4]>([1, 6, 4, 256])];
             tensor<bool, [4]> q_19_end_mask_0 = const()[name = tensor<string, []>("q_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> q_19_squeeze_mask_0 = const()[name = tensor<string, []>("q_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1400)[name = tensor<string, []>("q_19")];
+            tensor<fp32, [6, 4, 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_1433)[name = tensor<string, []>("q_19")];
             tensor<int32, [4]> k_19_begin_0 = const()[name = tensor<string, []>("k_19_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
             tensor<int32, [4]> k_19_end_0 = const()[name = tensor<string, []>("k_19_end_0"), val = tensor<int32, [4]>([2, 6, 4, 256])];
             tensor<bool, [4]> k_19_end_mask_0 = const()[name = tensor<string, []>("k_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> k_19_squeeze_mask_0 = const()[name = tensor<string, []>("k_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1400)[name = tensor<string, []>("k_19")];
+            tensor<fp32, [6, 4, 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_1433)[name = tensor<string, []>("k_19")];
             tensor<int32, [4]> v_19_begin_0 = const()[name = tensor<string, []>("v_19_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
             tensor<int32, [4]> v_19_end_0 = const()[name = tensor<string, []>("v_19_end_0"), val = tensor<int32, [4]>([3, 6, 4, 256])];
             tensor<bool, [4]> v_19_end_mask_0 = const()[name = tensor<string, []>("v_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
             tensor<bool, [4]> v_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
-            tensor<fp32, [6, 4, 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_1400)[name = tensor<string, []>("v_19")];
-            tensor<int32, [3]> var_1408 = const()[name = tensor<string, []>("op_1408"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor<string, []>("op_1409")];
+            tensor<fp32, [6, 4, 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_1433)[name = tensor<string, []>("v_19")];
+            tensor<int32, [3]> var_1441 = const()[name = tensor<string, []>("op_1441"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor<string, []>("op_1442")];
             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_1415 = const()[name = tensor<string, []>("op_1415"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor<string, []>("op_1416")];
+            tensor<int32, [3]> var_1448 = const()[name = tensor<string, []>("op_1448"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor<string, []>("op_1449")];
             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_1422 = const()[name = tensor<string, []>("op_1422"), val = tensor<int32, [3]>([6, 16, 64])];
-            tensor<fp32, [6, 16, 64]> var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor<string, []>("op_1423")];
+            tensor<int32, [3]> var_1455 = const()[name = tensor<string, []>("op_1455"), val = tensor<int32, [3]>([6, 16, 64])];
+            tensor<fp32, [6, 16, 64]> var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor<string, []>("op_1456")];
             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_1426 = const()[name = tensor<string, []>("op_1426"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor<string, []>("transpose_7")];
-            tensor<fp32, [4, 4, 6, 64]> q = reshape(shape = var_1426, x = q_21)[name = tensor<string, []>("q")];
-            tensor<int32, [4]> var_1428 = const()[name = tensor<string, []>("op_1428"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor<string, []>("transpose_6")];
-            tensor<fp32, [4, 4, 6, 64]> k = reshape(shape = var_1428, x = k_21)[name = tensor<string, []>("k")];
-            tensor<int32, [4]> var_1430 = const()[name = tensor<string, []>("op_1430"), val = tensor<int32, [4]>([4, 4, 6, 64])];
-            tensor<fp32, [16, 6, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor<string, []>("transpose_5")];
-            tensor<fp32, [4, 4, 6, 64]> v = reshape(shape = var_1430, x = v_21)[name = tensor<string, []>("v")];
+            tensor<int32, [4]> var_1459 = const()[name = tensor<string, []>("op_1459"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor<string, []>("transpose_7")];
+            tensor<fp32, [4, 4, 6, 64]> q = reshape(shape = var_1459, x = q_21)[name = tensor<string, []>("q")];
+            tensor<int32, [4]> var_1461 = const()[name = tensor<string, []>("op_1461"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor<string, []>("transpose_6")];
+            tensor<fp32, [4, 4, 6, 64]> k = reshape(shape = var_1461, x = k_21)[name = tensor<string, []>("k")];
+            tensor<int32, [4]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [4]>([4, 4, 6, 64])];
+            tensor<fp32, [16, 6, 64]> v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor<string, []>("transpose_5")];
+            tensor<fp32, [4, 4, 6, 64]> v = reshape(shape = var_1463, 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, [4, 4, 6, 64]> mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor<string, []>("mul_3")];
             tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
@@ -1277,36 +1292,36 @@ 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, [4, 4, 6, 64]> attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor<string, []>("attn_output_9")];
-            tensor<int32, [4]> var_1433 = const()[name = tensor<string, []>("op_1433"), val = tensor<int32, [4]>([2, 0, 1, 3])];
-            tensor<int32, [2]> var_1438 = const()[name = tensor<string, []>("op_1438"), val = tensor<int32, [2]>([24, 256])];
-            tensor<fp32, [6, 4, 4, 64]> var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
-            tensor<fp32, [24, 256]> attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor<string, []>("attn_output_11")];
-            tensor<fp32, [24, 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_1442 = const()[name = tensor<string, []>("op_1442"), val = tensor<int32, [3]>([6, 4, 256])];
-            tensor<fp32, [6, 4, 256]> attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor<string, []>("attn_output")];
+            tensor<int32, [4]> var_1466 = const()[name = tensor<string, []>("op_1466"), val = tensor<int32, [4]>([2, 0, 1, 3])];
+            tensor<int32, [2]> var_1471 = const()[name = tensor<string, []>("op_1471"), val = tensor<int32, [2]>([24, 256])];
+            tensor<fp32, [6, 4, 4, 64]> var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor<string, []>("transpose_4")];
+            tensor<fp32, [24, 256]> attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor<string, []>("attn_output_11")];
+            tensor<fp32, [24, 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_1475 = const()[name = tensor<string, []>("op_1475"), val = tensor<int32, [3]>([6, 4, 256])];
+            tensor<fp32, [6, 4, 256]> attn_output = reshape(shape = var_1475, 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, [4, 6, 256]> sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor<string, []>("transpose_3")];
-            tensor<fp32, [4, 6, 256]> input_193 = add(x = query_5, y = sa2_out)[name = tensor<string, []>("input_193")];
-            tensor<int32, [1]> input_195_axes_0 = const()[name = tensor<string, []>("input_195_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [4, 6, 256]> input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor<string, []>("input_195")];
-            tensor<fp32, [4, 6, 2048]> input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor<string, []>("linear_54")];
-            tensor<fp32, [4, 6, 2048]> input_199 = relu(x = input_197)[name = tensor<string, []>("input_199")];
-            tensor<fp32, [4, 6, 256]> ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor<string, []>("linear_55")];
-            tensor<fp32, [4, 6, 256]> input_201 = add(x = input_195, y = ffn_out)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [4, 6, 256]> input_195 = add(x = query_5, y = sa2_out)[name = tensor<string, []>("input_195")];
+            tensor<int32, [1]> input_197_axes_0 = const()[name = tensor<string, []>("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [4, 6, 256]> input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_65, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor<string, []>("input_197")];
+            tensor<fp32, [4, 6, 2048]> input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor<string, []>("linear_54")];
+            tensor<fp32, [4, 6, 2048]> input_201 = relu(x = input_199)[name = tensor<string, []>("input_201")];
+            tensor<fp32, [4, 6, 256]> ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor<string, []>("linear_55")];
+            tensor<fp32, [4, 6, 256]> input_203 = add(x = input_197, y = ffn_out)[name = tensor<string, []>("input_203")];
             tensor<int32, [1]> xt_axes_0 = const()[name = tensor<string, []>("xt_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [4, 6, 256]> xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor<string, []>("xt")];
-            tensor<int32, [4]> var_1462 = const()[name = tensor<string, []>("op_1462"), val = tensor<int32, [4]>([1, 4, 6, 256])];
-            tensor<fp32, [1, 4, 6, 256]> input = reshape(shape = var_1462, x = xt)[name = tensor<string, []>("input")];
-            tensor<int32, [1]> var_1464 = const()[name = tensor<string, []>("op_1464"), val = tensor<int32, [1]>([-1])];
-            tensor<fp32, [1, 4, 6, 1]> var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor<string, []>("op_1465")];
+            tensor<fp32, [4, 6, 256]> xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_65, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor<string, []>("xt")];
+            tensor<int32, [4]> var_1495 = const()[name = tensor<string, []>("op_1495"), val = tensor<int32, [4]>([1, 4, 6, 256])];
+            tensor<fp32, [1, 4, 6, 256]> input = reshape(shape = var_1495, x = xt)[name = tensor<string, []>("input")];
+            tensor<int32, [1]> var_1497 = const()[name = tensor<string, []>("op_1497"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 4, 6, 1]> var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_64, x = input)[name = tensor<string, []>("op_1498")];
             tensor<fp32, []> const_42 = const()[name = tensor<string, []>("const_42"), val = tensor<fp32, []>(0x1.fffffep+127)];
-            tensor<fp32, [1, 4, 6, 1]> clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor<string, []>("clip_5")];
-            tensor<fp32, [1, 4, 6, 256]> var_1467 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1467")];
+            tensor<fp32, [1, 4, 6, 1]> clip_5 = clip(alpha = var_78, beta = const_42, x = var_1498)[name = tensor<string, []>("clip_5")];
+            tensor<fp32, [1, 4, 6, 256]> var_1500 = real_div(x = input, y = clip_5)[name = tensor<string, []>("op_1500")];
             tensor<int32, [3]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<int32, [3]>([4, 1, 256])];
             tensor<fp32, [4, 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]>([4, 256, 6])];
-            tensor<fp32, [1, 4, 256, 6]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 4, 256, 6]> transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor<string, []>("transpose_2")];
             tensor<fp32, [4, 256, 6]> reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor<string, []>("reshape_1")];
             tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(false)];
@@ -1317,10 +1332,10 @@ program(1.0)
             tensor<int32, [3]> output_end_0 = const()[name = tensor<string, []>("output_end_0"), val = tensor<int32, [3]>([1, 4, 5])];
             tensor<bool, [3]> output_end_mask_0 = const()[name = tensor<string, []>("output_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
             tensor<fp32, [1, 4, 4]> output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor<string, []>("output")];
-            tensor<fp32, [1, 4, 4]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1471")];
-            tensor<int32, []> var_1473_axis_0 = const()[name = tensor<string, []>("op_1473_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 6, 4, 64, 64]> dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor<string, []>("op_1473")];
-            tensor<int32, []> var_1475_axis_0 = const()[name = tensor<string, []>("op_1475_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1475")];
+            tensor<fp32, [1, 4, 4]> probs = sigmoid(x = output)[name = tensor<string, []>("op_1504")];
+            tensor<int32, []> var_1506_axis_0 = const()[name = tensor<string, []>("op_1506_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 6, 4, 64, 64]> dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor<string, []>("op_1506")];
+            tensor<int32, []> var_1508_axis_0 = const()[name = tensor<string, []>("op_1508_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1]> dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor<string, []>("op_1508")];
         } -> (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