diff --git "a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil" "b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil" @@ -0,0 +1,1341 @@ +program(1.0) +[buildInfo = dict, tensor>({{"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(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 1, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_49")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor("stacked")]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 4, 345])]; + tensor input_1 = reshape(shape = var_56, x = stacked)[name = tensor("input_1")]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0x1p+0)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(true)]; + tensor var_65 = const()[name = tensor("op_65"), val = tensor(0x1.4f8b58p-17)]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(0)]; + tensor var_70 = const()[name = tensor("op_70"), val = tensor(2)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(-1)]; + tensor var_73 = const()[name = tensor("op_73"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor(0x1.5798eep-27)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor 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("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor 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("input_7")]; + tensor 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("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor 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("linear_2")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p-1)]; + tensor var_204 = mul(x = input_13, y = var_203)[name = tensor("op_204")]; + tensor input_15 = add(x = var_204, y = input_5)[name = tensor("input_15")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor 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("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale_1")]; + tensor var_218 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 4, 4, 64])]; + tensor var_220 = reshape(shape = var_219, x = var_218)[name = tensor("op_220")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_224 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(0x1p-3)]; + tensor var_226 = mul(x = var_224, y = var_225)[name = tensor("op_226")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 4, 4, 64])]; + tensor var_228 = reshape(shape = var_227, x = var_226)[name = tensor("op_228")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_232 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 4, 4, 64])]; + tensor var_234 = reshape(shape = var_233, x = var_232)[name = tensor("op_234")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor("transpose_58")]; + tensor 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("qk_1")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([4, 1])]; + tensor var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor("op_245")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor("M_1")]; + tensor var_247 = mul(x = qk_1, y = M_1)[name = tensor("op_247")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor("transpose_56")]; + tensor 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("inner_1")]; + tensor var_249_transpose_x_0 = const()[name = tensor("op_249_transpose_x_0"), val = tensor(false)]; + tensor var_249_transpose_y_0 = const()[name = tensor("op_249_transpose_y_0"), val = tensor(false)]; + tensor 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("op_249")]; + tensor var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_250")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1, 4, 1])]; + tensor var_252 = reshape(shape = var_251, x = var_250)[name = tensor("op_252")]; + tensor cross_1 = mul(x = var_249, y = var_252)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_255")]; + tensor var_257_transpose_x_1 = const()[name = tensor("op_257_transpose_x_1"), val = tensor(true)]; + tensor var_257_transpose_y_1 = const()[name = tensor("op_257_transpose_y_1"), val = tensor(false)]; + tensor 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("op_257")]; + tensor new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor("new_kv_unnorm_1")]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor("new_scale_1")]; + tensor var_261 = sqrt(x = new_scale_1)[name = tensor("op_261")]; + tensor var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor("op_262")]; + tensor var_263_perm_0 = const()[name = tensor("op_263_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_73, x = var_263)[name = tensor("out_3")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_267, x = out_3)[name = tensor("out_5")]; + tensor var_269 = silu(x = input_19)[name = tensor("op_269")]; + tensor input_21 = mul(x = var_269, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("window_1")]; + tensor var_277_begin_0 = const()[name = tensor("op_277_begin_0"), val = tensor([0, 0, 0])]; + tensor var_277_end_0 = const()[name = tensor("op_277_end_0"), val = tensor([1, 1, 256])]; + tensor var_277_end_mask_0 = const()[name = tensor("op_277_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_277")]; + tensor var_280_begin_0 = const()[name = tensor("op_280_begin_0"), val = tensor([0, 1, 0])]; + tensor var_280_end_0 = const()[name = tensor("op_280_end_0"), val = tensor([1, 16, 256])]; + tensor var_280_end_mask_0 = const()[name = tensor("op_280_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_280")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor("window_3")]; + tensor var_285_begin_0 = const()[name = tensor("op_285_begin_0"), val = tensor([0, 1, 0])]; + tensor var_285_end_0 = const()[name = tensor("op_285_end_0"), val = tensor([1, 2, 256])]; + tensor var_285_end_mask_0 = const()[name = tensor("op_285_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_285")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 1, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 16, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_288")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor("window_5")]; + tensor var_293_begin_0 = const()[name = tensor("op_293_begin_0"), val = tensor([0, 2, 0])]; + tensor var_293_end_0 = const()[name = tensor("op_293_end_0"), val = tensor([1, 3, 256])]; + tensor var_293_end_mask_0 = const()[name = tensor("op_293_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_293")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 16, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_296")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor("window_7")]; + tensor var_301_begin_0 = const()[name = tensor("op_301_begin_0"), val = tensor([0, 3, 0])]; + tensor var_301_end_0 = const()[name = tensor("op_301_end_0"), val = tensor([1, 1, 256])]; + tensor var_301_end_mask_0 = const()[name = tensor("op_301_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_301")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 1, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 16, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_304")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor("window_9")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_68, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_23")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor 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("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor 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("inputs_3")]; + tensor var_329_split_sizes_0 = const()[name = tensor("op_329_split_sizes_0"), val = tensor([256, 256])]; + tensor var_329_axis_0 = const()[name = tensor("op_329_axis_0"), val = tensor(1)]; + tensor var_329_0, tensor var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor("op_329")]; + tensor var_331 = sigmoid(x = var_329_1)[name = tensor("op_331")]; + tensor inputs_5 = mul(x = var_329_0, y = var_331)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor 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("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([4, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor 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("input_27")]; + tensor 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("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor 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("input_31")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_362_begin_0 = const()[name = tensor("op_362_begin_0"), val = tensor([0, -1, 0])]; + tensor var_362_end_0 = const()[name = tensor("op_362_end_0"), val = tensor([4, 16, 256])]; + tensor var_362_end_mask_0 = const()[name = tensor("op_362_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor 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("op_362")]; + tensor var_364_perm_0 = const()[name = tensor("op_364_perm_0"), val = tensor([1, 0, 2])]; + tensor var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_364)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor 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("input_35")]; + tensor 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("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor 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("linear_9")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor(0x1p-1)]; + tensor var_388 = mul(x = input_41, y = var_387)[name = tensor("op_388")]; + tensor input_43 = add(x = var_388, y = input_33)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor 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("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor 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("input_47")]; + tensor 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("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor 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("linear_11")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p-1)]; + tensor var_418 = mul(x = input_53, y = var_417)[name = tensor("op_418")]; + tensor input_55 = add(x = var_418, y = input_45)[name = tensor("input_55")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor 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("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale_3")]; + tensor var_432 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_433 = const()[name = tensor("op_433"), val = tensor([1, 4, 4, 64])]; + tensor var_434 = reshape(shape = var_433, x = var_432)[name = tensor("op_434")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_438 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor(0x1p-3)]; + tensor var_440 = mul(x = var_438, y = var_439)[name = tensor("op_440")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 4, 4, 64])]; + tensor var_442 = reshape(shape = var_441, x = var_440)[name = tensor("op_442")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_446 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 4, 4, 64])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor("transpose_51")]; + tensor 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("qk_3")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([4, 1])]; + tensor var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor("op_459")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor("M_3")]; + tensor var_461 = mul(x = qk_3, y = M_3)[name = tensor("op_461")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor("transpose_49")]; + tensor 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("inner_3")]; + tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; + tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; + tensor 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("op_463")]; + tensor var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_464")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 1, 4, 1])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; + tensor cross_3 = mul(x = var_463, y = var_466)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_469")]; + tensor var_471_transpose_x_1 = const()[name = tensor("op_471_transpose_x_1"), val = tensor(true)]; + tensor var_471_transpose_y_1 = const()[name = tensor("op_471_transpose_y_1"), val = tensor(false)]; + tensor 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("op_471")]; + tensor new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor("new_kv_unnorm_3")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor("new_scale_3")]; + tensor var_475 = sqrt(x = new_scale_3)[name = tensor("op_475")]; + tensor var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor("op_476")]; + tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_73, x = var_477)[name = tensor("out_9")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_481, x = out_9)[name = tensor("out_11")]; + tensor var_483 = silu(x = input_59)[name = tensor("op_483")]; + tensor input_61 = mul(x = var_483, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; + tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("window_11")]; + tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 0, 0])]; + tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 1, 256])]; + tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_491")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 16, 256])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_494")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor("window_13")]; + tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 1, 0])]; + tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 2, 256])]; + tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_499")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 16, 256])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_502")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor("window_15")]; + tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 2, 0])]; + tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 3, 256])]; + tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_507")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 16, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_510")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor("window_17")]; + tensor var_515_begin_0 = const()[name = tensor("op_515_begin_0"), val = tensor([0, 3, 0])]; + tensor var_515_end_0 = const()[name = tensor("op_515_end_0"), val = tensor([1, 1, 256])]; + tensor var_515_end_mask_0 = const()[name = tensor("op_515_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_515")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 16, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_518")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor("window_19")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_68, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_63")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor 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("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor 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("inputs_13")]; + tensor var_543_split_sizes_0 = const()[name = tensor("op_543_split_sizes_0"), val = tensor([256, 256])]; + tensor var_543_axis_0 = const()[name = tensor("op_543_axis_0"), val = tensor(1)]; + tensor var_543_0, tensor var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor("op_543")]; + tensor var_545 = sigmoid(x = var_543_1)[name = tensor("op_545")]; + tensor inputs_15 = mul(x = var_543_0, y = var_545)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor 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("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([4, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor 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("input_67")]; + tensor 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("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor 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("input_71")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_576_begin_0 = const()[name = tensor("op_576_begin_0"), val = tensor([0, -1, 0])]; + tensor var_576_end_0 = const()[name = tensor("op_576_end_0"), val = tensor([4, 16, 256])]; + tensor var_576_end_mask_0 = const()[name = tensor("op_576_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor 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("op_576")]; + tensor var_578_perm_0 = const()[name = tensor("op_578_perm_0"), val = tensor([1, 0, 2])]; + tensor var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_578)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor 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("input_75")]; + tensor 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("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor 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("linear_18")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-1)]; + tensor var_602 = mul(x = input_81, y = var_601)[name = tensor("op_602")]; + tensor input_83 = add(x = var_602, y = input_73)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor 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("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor 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("input_87")]; + tensor 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("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor 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("linear_20")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p-1)]; + tensor var_632 = mul(x = input_93, y = var_631)[name = tensor("op_632")]; + tensor input_95 = add(x = var_632, y = input_85)[name = tensor("input_95")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor 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("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale_5")]; + tensor var_646 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 4, 4, 64])]; + tensor var_648 = reshape(shape = var_647, x = var_646)[name = tensor("op_648")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_652 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor(0x1p-3)]; + tensor var_654 = mul(x = var_652, y = var_653)[name = tensor("op_654")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 4, 4, 64])]; + tensor var_656 = reshape(shape = var_655, x = var_654)[name = tensor("op_656")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_660 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 4, 4, 64])]; + tensor var_662 = reshape(shape = var_661, x = var_660)[name = tensor("op_662")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor("transpose_44")]; + tensor 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("qk_5")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([4, 1])]; + tensor var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor("op_673")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor("M_5")]; + tensor var_675 = mul(x = qk_5, y = M_5)[name = tensor("op_675")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor("transpose_42")]; + tensor 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("inner_5")]; + tensor var_677_transpose_x_0 = const()[name = tensor("op_677_transpose_x_0"), val = tensor(false)]; + tensor var_677_transpose_y_0 = const()[name = tensor("op_677_transpose_y_0"), val = tensor(false)]; + tensor 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("op_677")]; + tensor var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_678")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1, 4, 1])]; + tensor var_680 = reshape(shape = var_679, x = var_678)[name = tensor("op_680")]; + tensor cross_5 = mul(x = var_677, y = var_680)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_683")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(true)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(false)]; + tensor 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("op_685")]; + tensor new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor("new_kv_unnorm_5")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor("new_scale_5")]; + tensor var_689 = sqrt(x = new_scale_5)[name = tensor("op_689")]; + tensor var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor("op_690")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_73, x = var_691)[name = tensor("out_15")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_695, x = out_15)[name = tensor("out_17")]; + tensor var_697 = silu(x = input_99)[name = tensor("op_697")]; + tensor input_101 = mul(x = var_697, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; + tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("window_21")]; + tensor var_705_begin_0 = const()[name = tensor("op_705_begin_0"), val = tensor([0, 0, 0])]; + tensor var_705_end_0 = const()[name = tensor("op_705_end_0"), val = tensor([1, 1, 256])]; + tensor var_705_end_mask_0 = const()[name = tensor("op_705_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_705")]; + tensor var_708_begin_0 = const()[name = tensor("op_708_begin_0"), val = tensor([0, 1, 0])]; + tensor var_708_end_0 = const()[name = tensor("op_708_end_0"), val = tensor([1, 16, 256])]; + tensor var_708_end_mask_0 = const()[name = tensor("op_708_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_708")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor("window_23")]; + tensor var_713_begin_0 = const()[name = tensor("op_713_begin_0"), val = tensor([0, 1, 0])]; + tensor var_713_end_0 = const()[name = tensor("op_713_end_0"), val = tensor([1, 2, 256])]; + tensor var_713_end_mask_0 = const()[name = tensor("op_713_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_713")]; + tensor var_716_begin_0 = const()[name = tensor("op_716_begin_0"), val = tensor([0, 1, 0])]; + tensor var_716_end_0 = const()[name = tensor("op_716_end_0"), val = tensor([1, 16, 256])]; + tensor var_716_end_mask_0 = const()[name = tensor("op_716_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_716")]; + tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; + tensor window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor("window_25")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 2, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 3, 256])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_721")]; + tensor var_724_begin_0 = const()[name = tensor("op_724_begin_0"), val = tensor([0, 1, 0])]; + tensor var_724_end_0 = const()[name = tensor("op_724_end_0"), val = tensor([1, 16, 256])]; + tensor var_724_end_mask_0 = const()[name = tensor("op_724_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_724")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor("window_27")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 3, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 1, 256])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_729")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 1, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 16, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_732")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor("window_29")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_68, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_103")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor 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("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor 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("inputs_23")]; + tensor var_757_split_sizes_0 = const()[name = tensor("op_757_split_sizes_0"), val = tensor([256, 256])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_0, tensor var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor("op_757")]; + tensor var_759 = sigmoid(x = var_757_1)[name = tensor("op_759")]; + tensor inputs_25 = mul(x = var_757_0, y = var_759)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor 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("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([4, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor 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("input_107")]; + tensor 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("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor 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("input_111")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_790_begin_0 = const()[name = tensor("op_790_begin_0"), val = tensor([0, -1, 0])]; + tensor var_790_end_0 = const()[name = tensor("op_790_end_0"), val = tensor([4, 16, 256])]; + tensor var_790_end_mask_0 = const()[name = tensor("op_790_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor 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("op_790")]; + tensor var_792_perm_0 = const()[name = tensor("op_792_perm_0"), val = tensor([1, 0, 2])]; + tensor var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_792)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor 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("input_115")]; + tensor 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("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor 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("linear_27")]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor(0x1p-1)]; + tensor var_816 = mul(x = input_121, y = var_815)[name = tensor("op_816")]; + tensor input_123 = add(x = var_816, y = input_113)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor 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("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor 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("input_127")]; + tensor 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("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor 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("linear_29")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p-1)]; + tensor var_846 = mul(x = input_133, y = var_845)[name = tensor("op_846")]; + tensor input_135 = add(x = var_846, y = input_125)[name = tensor("input_135")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor 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("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale_7")]; + tensor var_860 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 4, 4, 64])]; + tensor var_862 = reshape(shape = var_861, x = var_860)[name = tensor("op_862")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_866 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1p-3)]; + tensor var_868 = mul(x = var_866, y = var_867)[name = tensor("op_868")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 4, 4, 64])]; + tensor var_870 = reshape(shape = var_869, x = var_868)[name = tensor("op_870")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_874 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 4, 4, 64])]; + tensor var_876 = reshape(shape = var_875, x = var_874)[name = tensor("op_876")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor("transpose_37")]; + tensor 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("qk_7")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([4, 1])]; + tensor var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor("op_887")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor("M_7")]; + tensor var_889 = mul(x = qk_7, y = M_7)[name = tensor("op_889")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor("transpose_35")]; + tensor 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("inner_7")]; + tensor var_891_transpose_x_0 = const()[name = tensor("op_891_transpose_x_0"), val = tensor(false)]; + tensor var_891_transpose_y_0 = const()[name = tensor("op_891_transpose_y_0"), val = tensor(false)]; + tensor 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("op_891")]; + tensor var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_892")]; + tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 1, 4, 1])]; + tensor var_894 = reshape(shape = var_893, x = var_892)[name = tensor("op_894")]; + tensor cross_7 = mul(x = var_891, y = var_894)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_897")]; + tensor var_899_transpose_x_1 = const()[name = tensor("op_899_transpose_x_1"), val = tensor(true)]; + tensor var_899_transpose_y_1 = const()[name = tensor("op_899_transpose_y_1"), val = tensor(false)]; + tensor 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("op_899")]; + tensor new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor("new_kv_unnorm_7")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor("new_scale_7")]; + tensor var_903 = sqrt(x = new_scale_7)[name = tensor("op_903")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor("nkv_1")]; + tensor var_905_perm_0 = const()[name = tensor("op_905_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_73, x = var_905)[name = tensor("out_21")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_909, x = out_21)[name = tensor("out_23")]; + tensor var_911 = silu(x = input_139)[name = tensor("op_911")]; + tensor input_141 = mul(x = var_911, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; + tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("window_31")]; + tensor var_919_begin_0 = const()[name = tensor("op_919_begin_0"), val = tensor([0, 0, 0])]; + tensor var_919_end_0 = const()[name = tensor("op_919_end_0"), val = tensor([1, 1, 256])]; + tensor var_919_end_mask_0 = const()[name = tensor("op_919_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_919")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 1, 0])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 16, 256])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_922")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor("window_33")]; + tensor var_927_begin_0 = const()[name = tensor("op_927_begin_0"), val = tensor([0, 1, 0])]; + tensor var_927_end_0 = const()[name = tensor("op_927_end_0"), val = tensor([1, 2, 256])]; + tensor var_927_end_mask_0 = const()[name = tensor("op_927_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_927")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 1, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 16, 256])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_930")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor("window_35")]; + tensor var_935_begin_0 = const()[name = tensor("op_935_begin_0"), val = tensor([0, 2, 0])]; + tensor var_935_end_0 = const()[name = tensor("op_935_end_0"), val = tensor([1, 3, 256])]; + tensor var_935_end_mask_0 = const()[name = tensor("op_935_end_mask_0"), val = tensor([true, false, true])]; + tensor 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("op_935")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 1, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 16, 256])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_938")]; + tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; + tensor window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor("window_37")]; + tensor var_943_begin_0 = const()[name = tensor("op_943_begin_0"), val = tensor([0, 3, 0])]; + tensor var_943_end_0 = const()[name = tensor("op_943_end_0"), val = tensor([1, 1, 256])]; + tensor var_943_end_mask_0 = const()[name = tensor("op_943_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_943")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 1, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 16, 256])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_946")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_68, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_143")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor 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("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor 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("inputs_33")]; + tensor var_971_split_sizes_0 = const()[name = tensor("op_971_split_sizes_0"), val = tensor([256, 256])]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(1)]; + tensor var_971_0, tensor var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor("op_971")]; + tensor var_973 = sigmoid(x = var_971_1)[name = tensor("op_973")]; + tensor inputs_35 = mul(x = var_971_0, y = var_973)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor 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("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([4, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor 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("input_147")]; + tensor 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("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor 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("input_151")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1004_begin_0 = const()[name = tensor("op_1004_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1004_end_0 = const()[name = tensor("op_1004_end_0"), val = tensor([4, 16, 256])]; + tensor var_1004_end_mask_0 = const()[name = tensor("op_1004_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor 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("op_1004")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1006)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor 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("input_155")]; + tensor 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("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor 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("linear_36")]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor(0x1p-1)]; + tensor var_1030 = mul(x = input_161, y = var_1029)[name = tensor("op_1030")]; + tensor input_163 = add(x = var_1030, y = input_153)[name = tensor("input_163")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor 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("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_70, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor 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("conv_out")]; + tensor var_1048_begin_0 = const()[name = tensor("op_1048_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1048_end_0 = const()[name = tensor("op_1048_end_0"), val = tensor([1, 256, 22])]; + tensor var_1048_end_mask_0 = const()[name = tensor("op_1048_end_mask_0"), val = tensor([true, true, true])]; + tensor 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("op_1048")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_64, x = input_165)[name = tensor("op_1051")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_78, beta = const_12, x = var_1051)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1055_axis_0 = const()[name = tensor("op_1055_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor("op_1055")]; + tensor var_1057_axis_0 = const()[name = tensor("op_1057_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1057")]; + tensor var_1059_axis_0 = const()[name = tensor("op_1059_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1059")]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_71, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([6, 4, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale_9")]; + tensor var_1147 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 4, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1153 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([6, 4, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1161 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([6, 4, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_68, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_58, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor 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("qk_9")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 4])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([4, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor 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("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor 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("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 4, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 4, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor 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("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_58, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_73, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([6, 4, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_171)[name = tensor("op_1209")]; + tensor input_173 = mul(x = var_1209, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor 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("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 6, 4, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([4, 6, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor 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("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 4, 3, 256])]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 4, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 4, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 4, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([6, 16, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([6, 16, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([6, 16, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([4, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([4, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([4, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor 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("attn_output_1")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([24, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor 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("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([6, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor 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("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor 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("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 4, 6, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([6, 4, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor 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("prev_scale")]; + tensor var_1328 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 4, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([6, 4, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([6, 4, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_58, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([4, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor 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("inner_11")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor 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("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 4, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor 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("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_58, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_73, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([6, 4, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_189)[name = tensor("op_1390")]; + tensor input_191 = mul(x = var_1390, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor 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("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 6, 4, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([4, 6, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor 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("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 4, 3, 256])]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 4, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 4, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 4, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor 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("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([6, 16, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([6, 16, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([6, 16, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([4, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([4, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([4, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor 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("attn_output_9")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([24, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor 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("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([6, 4, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor 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("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor 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("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 4, 6, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_64, x = input)[name = tensor("op_1498")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_78, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 4, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("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