diff --git "a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" "b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" @@ -0,0 +1,1286 @@ +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([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + 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(4981952)))]; + 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(4983040)))]; + 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(5336384)))]; + 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(5337472)))]; + 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(5338560)))]; + 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(5339648)))]; + 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(5340736)))]; + 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(5344896)))]; + 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(6393536)))]; + 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(6394624)))]; + 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(7443264)))]; + 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(7444352)))]; + 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(7445440)))]; + 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(7446528)))]; + 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(7708736)))]; + 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(7709824)))]; + 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(7972032)))]; + 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(7973120)))]; + 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(8235328)))]; + 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(8236416)))]; + 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(8498624)))]; + 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(8499712)))]; + 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(8761920)))]; + 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(8763008)))]; + 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(8764096)))]; + 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(8766208)))]; + 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(9290560)))]; + 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(9307008)))]; + 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(9308096)))]; + 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(9309184)))]; + 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(9310272)))]; + 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(9311360)))]; + 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(9312448)))]; + 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(9574656)))]; + 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(9575744)))]; + 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(9576832)))]; + 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(9580992)))]; + 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(10629632)))]; + 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(10630720)))]; + 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(11679360)))]; + 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(11680448)))]; + 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(11681536)))]; + 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(11682624)))]; + 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(11683712)))]; + 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(11687872)))]; + 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(12736512)))]; + 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(12737600)))]; + 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(13786240)))]; + 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(13787328)))]; + 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(13788416)))]; + 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(13789504)))]; + 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(14051712)))]; + 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(14052800)))]; + 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(14315008)))]; + 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(14316096)))]; + 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(14578304)))]; + 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(14579392)))]; + 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(14841600)))]; + 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(14842688)))]; + 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(15104896)))]; + 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(15105984)))]; + 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(15107072)))]; + 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(15109184)))]; + 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(15633536)))]; + 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(15649984)))]; + 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(15651072)))]; + 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(15652160)))]; + 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(15653248)))]; + 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(15654336)))]; + 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(15655424)))]; + 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(15917632)))]; + 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(15918720)))]; + 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(15919808)))]; + 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(15923968)))]; + 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(16972608)))]; + 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(16973696)))]; + 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(18022336)))]; + 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(18023424)))]; + 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(18024512)))]; + 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(18025600)))]; + 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(18026688)))]; + 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(18030848)))]; + 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(19079488)))]; + 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(19080576)))]; + 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(20129216)))]; + 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(20130304)))]; + 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(20131392)))]; + 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(20132480)))]; + 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(20394688)))]; + 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(20395776)))]; + 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(20657984)))]; + 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(20659072)))]; + 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(20921280)))]; + 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(20922368)))]; + 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(21184576)))]; + 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(21185664)))]; + 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(21447872)))]; + 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(21448960)))]; + 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(21450048)))]; + 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(21452160)))]; + 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(21976512)))]; + 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(21992960)))]; + 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(21994048)))]; + 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(21995136)))]; + 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(21996224)))]; + 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(21997312)))]; + 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(21998400)))]; + 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(22260608)))]; + 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(22261696)))]; + 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(22262784)))]; + 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(22266944)))]; + 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(23315584)))]; + 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(23316672)))]; + 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(24365312)))]; + 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(24366400)))]; + 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(24367488)))]; + 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(24368576)))]; + 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(24369664)))]; + 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(24373824)))]; + 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(25422464)))]; + 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(25423552)))]; + 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(26472192)))]; + 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(26473280)))]; + 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(26474368)))]; + 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(26475456)))]; + 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(26737664)))]; + 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(26738752)))]; + 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(27000960)))]; + 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(27002048)))]; + 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(27264256)))]; + 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(27265344)))]; + 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(27527552)))]; + 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(27528640)))]; + 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(27790848)))]; + 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(27791936)))]; + 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(27793024)))]; + 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(27795136)))]; + 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(28319488)))]; + 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(28335936)))]; + 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(28337024)))]; + 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(28338112)))]; + 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(28339200)))]; + 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(28340288)))]; + 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(28341376)))]; + 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(28603584)))]; + 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(28604672)))]; + 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(28605760)))]; + 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(28609920)))]; + 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(29658560)))]; + 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(29659648)))]; + 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(30708288)))]; + 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(30709376)))]; + tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + 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(31235904)))]; + 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(31236992)))]; + 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(31499200)))]; + 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(31500288)))]; + 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(31762496)))]; + 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(31763584)))]; + 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(32025792)))]; + 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(32026880)))]; + 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(32289088)))]; + 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(32290176)))]; + tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + 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(32554560)))]; + 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(32555648)))]; + 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(32817856)))]; + 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(32820992)))]; + tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + 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(37815616)))]; + 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(37816704)))]; + 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(38078912)))]; + 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(38080000)))]; + 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(38342208)))]; + 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(38343296)))]; + 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(38605504)))]; + 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(38606592)))]; + 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(38868800)))]; + 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(38869888)))]; + tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + 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(39134272)))]; + 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(39135360)))]; + 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(39397568)))]; + 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(39400704)))]; + tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; + tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; + tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; + tensor var_29 = const()[name = tensor("op_29"), 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_28, 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_28, 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_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; + tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; + tensor input_13 = add(x = var_148, 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_28, 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_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; + tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; + tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; + tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; + tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; + tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + 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_172)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[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_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; + tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; + tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; + tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + 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_178)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; + tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; + tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; + tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; + tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; + tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; + tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; + tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; + tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; + tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; + tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; + tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; + tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; + tensor var_207_perm_0 = const()[name = tensor("op_207_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_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; + tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; + tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; + tensor input_19 = mul(x = var_213, 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_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; + tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; + tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; + tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; + tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; + tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; + tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; + tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; + tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; + tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; + tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; + tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; + tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; + tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; + tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; + tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; + tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; + tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; + tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; + tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; + tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; + tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; + tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; + tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; + tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; + tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[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_28, 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_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; + tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; + tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; + tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; + tensor inputs_5 = mul(x = var_265_0, y = var_267)[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([3, 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_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("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_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_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_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; + tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; + tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; + tensor input_31 = add(x = x_3, y = var_300)[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_28, 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_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; + tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; + tensor input_41 = add(x = var_324, 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_28, 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_28, 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_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; + tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; + tensor input_53 = add(x = var_354, 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_28, 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_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; + tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; + tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; + tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; + tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; + tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + 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_378)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[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_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; + tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; + tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; + tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + 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_384)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; + tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; + tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; + tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; + tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; + tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; + tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; + tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; + tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; + tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; + tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; + tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; + tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; + tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; + tensor var_413_perm_0 = const()[name = tensor("op_413_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_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; + tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; + tensor input_59 = mul(x = var_419, 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_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; + tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; + tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; + tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; + tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; + tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; + tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; + tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; + tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; + tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; + tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; + tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; + tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; + tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; + tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; + tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; + tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; + tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; + tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; + tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; + tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; + tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; + tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[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_28, 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_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; + tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; + tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; + tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; + tensor inputs_15 = mul(x = var_471_0, y = var_473)[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([3, 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_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("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_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; + tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; + tensor var_504_end_mask_0 = const()[name = tensor("op_504_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_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; + tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; + tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; + tensor input_71 = add(x = x_9, y = var_506)[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_28, 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_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; + tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; + tensor input_81 = add(x = var_530, 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_28, 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_28, 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_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; + tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; + tensor input_93 = add(x = var_560, 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_28, 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_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; + tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; + tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; + tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + 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_584)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[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_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; + tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; + tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; + tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + 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_590)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; + tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; + tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; + tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; + tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; + tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; + tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; + tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; + tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; + tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; + tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; + tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; + tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; + tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; + tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; + tensor var_619_perm_0 = const()[name = tensor("op_619_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_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; + tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; + tensor input_99 = mul(x = var_625, 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_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; + tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; + tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; + tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; + tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; + tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; + tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; + tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; + tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; + tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; + tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; + tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; + tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; + tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; + tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; + tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; + tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; + tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; + tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; + tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; + tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; + tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[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_28, 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_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; + tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; + tensor inputs_25 = mul(x = var_677_0, y = var_679)[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([3, 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_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("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_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; + tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; + tensor var_710_end_mask_0 = const()[name = tensor("op_710_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_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; + tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; + tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; + tensor input_111 = add(x = x_15, y = var_712)[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_28, 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_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; + tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; + tensor input_121 = add(x = var_736, 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_28, 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_28, 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_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; + tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; + tensor input_133 = add(x = var_766, 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_28, 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_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; + tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; + tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; + tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; + tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; + tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + 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_790)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[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_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; + tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; + tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; + tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + 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_796)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; + tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; + tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; + tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; + tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; + tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; + tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; + tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; + tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; + tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; + tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; + tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; + tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; + tensor var_825_perm_0 = const()[name = tensor("op_825_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_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; + tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; + tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; + tensor input_139 = mul(x = var_831, 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_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 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_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; + tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; + tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; + tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; + tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; + tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; + tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; + tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; + tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; + tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; + tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; + tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; + tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; + tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; + tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; + tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; + tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; + tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; + tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; + tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; + tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; + tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; + tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; + tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; + tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; + tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, 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_28, 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_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; + tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; + tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; + tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; + tensor inputs_35 = mul(x = var_883_0, y = var_885)[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([3, 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_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("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_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; + tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; + tensor var_916_end_mask_0 = const()[name = tensor("op_916_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_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; + tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; + tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; + tensor input_151 = add(x = x_21, y = var_918)[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_28, 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_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; + tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; + tensor input_161 = add(x = var_942, 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_28, 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_20, 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_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; + tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; + tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; + tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; + tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + 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_963)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; + tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; + tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; + tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; + tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + 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_1062)[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_994, 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_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([6, 3, 256])]; + tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1074, x = var_1070)[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_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([6, 3, 4, 64])]; + tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; + tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([6, 3, 4, 64])]; + tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([6, 3, 4, 64])]; + tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + 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_1000, 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_990, 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_1092)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[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_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; + tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; + tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; + tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; + tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + 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_1098)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; + tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; + tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; + tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; + tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; + tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; + tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; + tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; + tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; + tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; + tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; + tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; + tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; + tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; + tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; + tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_990, 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_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; + tensor var_1138_perm_0 = const()[name = tensor("op_1138_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_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; + tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([6, 3, 256])]; + tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; + tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; + tensor input_171 = mul(x = var_1144, 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_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 6, 3, 256])]; + tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; + tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 6, 256])]; + tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1159, x = var_1156)[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_1182 = 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, 3, 3, 256])]; + tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; + tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; + tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; + tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; + tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; + tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + 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, 3, 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_1187)[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, 3, 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_1187)[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, 3, 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_1187)[name = tensor("v_11")]; + tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([6, 12, 64])]; + tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([6, 12, 64])]; + tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([6, 12, 64])]; + tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; + tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1217, 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_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([18, 256])]; + tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[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_1229 = const()[name = tensor("op_1229"), val = tensor([6, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1229, 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_985, 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_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 6, 256])]; + tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([6, 3, 256])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1255, x = var_1251)[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_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([6, 3, 4, 64])]; + tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; + tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([6, 3, 4, 64])]; + tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([6, 3, 4, 64])]; + tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + 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_990, 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_1273)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[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_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; + tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; + tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; + 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_1279)[name = tensor("transpose_12")]; + tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; + tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; + tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; + tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; + tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; + tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; + tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; + tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; + tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; + tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; + tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_990, 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_1319_perm_0 = const()[name = tensor("op_1319_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_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([6, 3, 256])]; + tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; + tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; + tensor input_189 = mul(x = var_1325, 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_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 6, 3, 256])]; + tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; + tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 6, 256])]; + tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1340, x = var_1337)[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_1363 = 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, 3, 3, 256])]; + tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; + tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; + tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; + tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; + tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; + tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + 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, 3, 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_1368)[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, 3, 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_1368)[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, 3, 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_1368)[name = tensor("v_19")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([6, 12, 64])]; + tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([6, 12, 64])]; + tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([6, 12, 64])]; + tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; + tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1398, 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_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([18, 256])]; + tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[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_1410 = const()[name = tensor("op_1410"), val = tensor([6, 3, 256])]; + tensor attn_output = reshape(shape = var_1410, 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_985, 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_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 6, 256])]; + tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; + tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; + tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; + tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 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([3, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[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, 3, 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, 3, 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_1439")]; + tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; + tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + } -> (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