diff --git "a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil" "b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil" @@ -0,0 +1,1366 @@ +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 encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; + tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; + tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; + tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; + tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + 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 = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; + tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; + tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; + tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; + tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; + tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + 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 = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[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_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; + tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; + tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; + tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; + tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; + tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = 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 = 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_173)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[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_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; + tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; + tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + 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_179)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; + tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; + tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; + tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; + tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; + tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; + tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; + tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; + tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; + tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; + tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; + tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; + tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; + tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; + tensor var_208_perm_0 = const()[name = tensor("op_208_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_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; + tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; + tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; + tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_13, 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_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; + tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; + tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; + tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; + tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; + tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; + tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; + tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; + tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; + tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; + tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; + tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; + tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; + tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; + tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; + tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; + tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; + tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; + tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; + tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; + tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; + tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; + tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; + tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; + tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; + tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; + tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; + tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; + tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; + tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; + tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; + tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; + tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; + tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; + tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; + tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + 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 = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; + tensor input_23_perm_0 = const()[name = tensor("input_23_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_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; + tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; + tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; + tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; + tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; + tensor inputs_5 = mul(x = var_282_0, y = var_284)[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 = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; + tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; + tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; + tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; + tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; + tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; + tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; + tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; + tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; + tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; + tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; + tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; + tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; + tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; + tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; + tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[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 = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; + tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; + tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; + tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; + tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; + tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + 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 = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[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_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; + tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; + tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; + tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; + tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; + tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = 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 = 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_395)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[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_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; + tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; + tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + 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_401)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; + tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; + tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; + tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; + tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; + tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; + tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; + tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; + tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; + tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; + tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; + tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; + tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; + tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; + tensor var_430_perm_0 = const()[name = tensor("op_430_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_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; + tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; + tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; + tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; + tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; + tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; + tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; + tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; + tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; + tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; + tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; + tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; + tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; + tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; + tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; + tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; + tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; + tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; + tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; + tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; + tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; + tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; + tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; + tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; + tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; + tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; + tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; + tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; + tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; + tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; + tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + 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 = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; + tensor input_63_perm_0 = const()[name = tensor("input_63_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_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; + tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; + tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; + tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; + tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; + tensor inputs_15 = mul(x = var_504_0, y = var_506)[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 = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; + tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; + tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; + tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; + tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; + tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; + tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; + tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; + tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; + tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; + tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; + tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; + tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; + tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; + tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[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 = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; + tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; + tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; + tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; + tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; + tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + 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 = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[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_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; + tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; + tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; + tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; + tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; + tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = 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 = 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_617)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[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_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; + tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; + tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + 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_623)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; + tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; + tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; + tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; + tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; + tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; + tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; + tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; + tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; + tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; + tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; + tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; + tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; + tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; + tensor var_652_perm_0 = const()[name = tensor("op_652_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_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; + tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; + tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; + tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; + tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; + tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; + tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; + tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; + tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; + tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; + tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; + tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; + tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; + tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; + tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; + tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; + tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; + tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; + tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; + tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; + tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; + tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; + tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; + tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; + tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + 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 = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; + tensor input_103_perm_0 = const()[name = tensor("input_103_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_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; + tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; + tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; + tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; + tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; + tensor inputs_25 = mul(x = var_726_0, y = var_728)[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 = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; + tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; + tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; + tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; + tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; + tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; + tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; + tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; + tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; + tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; + tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; + tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; + tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; + tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[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 = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; + tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; + tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; + tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; + tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; + tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; + tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + 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 = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[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_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; + tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; + tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; + tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; + tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; + tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = 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 = 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_839)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[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_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; + tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; + tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + 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_845)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; + tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; + tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; + tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; + tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; + tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; + tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; + tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; + tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; + tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; + tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; + tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; + tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; + tensor var_874_perm_0 = const()[name = tensor("op_874_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_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; + tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; + tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; + tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; + tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; + tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; + tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; + tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; + tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; + tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; + tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; + tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; + tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; + tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; + tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; + tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; + tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; + tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; + tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; + tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; + tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; + tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; + tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; + tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; + tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; + tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; + tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; + tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; + tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; + tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; + tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; + tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; + tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + 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 = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; + tensor input_143_perm_0 = const()[name = tensor("input_143_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_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; + tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; + tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; + tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; + tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; + tensor inputs_35 = mul(x = var_948_0, y = var_950)[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 = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; + tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; + tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; + tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; + tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; + tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; + tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; + tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; + tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; + tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; + tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; + tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; + tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; + tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + 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 = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[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_21, 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 = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; + tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; + tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; + tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; + tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; + tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; + 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_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; + tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; + tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[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, 5, 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 = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 5, 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 = decoder_k_proj_0_bias, weight = 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, 5, 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 = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([6, 5, 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_169 = linear(bias = decoder_g_proj_0_bias, weight = 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_1065, 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_1055, 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, 5])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 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, 5, 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, 5, 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_1055, 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_1052, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([6, 5, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; + tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; + tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; + tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + 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 = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 6, 5, 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([5, 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 = decoder_self_attn2_0_in_proj_bias, weight = 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, 5, 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, 5, 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, 5, 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, 5, 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, 20, 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, 20, 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, 20, 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([5, 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([5, 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([5, 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([30, 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 = decoder_self_attn2_0_out_proj_bias, weight = 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, 5, 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_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; + tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; + tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; + tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; + tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + 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 = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 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, 5, 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 = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 5, 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 = decoder_k_proj_1_bias, weight = 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, 5, 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 = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([6, 5, 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_187 = linear(bias = decoder_g_proj_1_bias, weight = 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_1055, 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([5, 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_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; + tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; + 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, 5, 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, 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_1055, 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_1052, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([6, 5, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; + tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; + tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; + tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + 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 = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 6, 5, 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([5, 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 = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 5, 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, 5, 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, 5, 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, 5, 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, 20, 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, 20, 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, 20, 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([5, 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([5, 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([5, 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([30, 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 = decoder_self_attn2_1_out_proj_bias, weight = 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, 5, 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_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; + tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; + tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; + tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; + tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 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_1053, 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_1045, 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([5, 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([5, 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, 5, 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, 5, 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