diff --git "a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil" "b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil" --- "a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil" +++ "b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil" @@ -1,233 +1,239 @@ 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]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; - 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_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")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor stacked_axes_0 = const()[name = tensor("stacked_axes_0"), val = tensor([1])]; + tensor stacked = expand_dims(axes = stacked_axes_0, x = features)[name = tensor("stacked")]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1, 1, 345])]; + tensor input_1 = reshape(shape = var_26, x = stacked)[name = tensor("input_1")]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1p+0)]; + tensor var_34 = const()[name = tensor("op_34"), val = tensor(true)]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor(0x1.4f8b58p-17)]; + tensor var_38 = const()[name = tensor("op_38"), val = tensor(0)]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor(2)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor(-1)]; + tensor var_43 = const()[name = tensor("op_43"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0x1.5798eep-27)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = 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 input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_35, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_173 = const()[name = tensor("op_173"), val = tensor(0x1p-1)]; + tensor var_174 = mul(x = input_13, y = var_173)[name = tensor("op_174")]; + tensor input_15 = add(x = var_174, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -238,139 +244,139 @@ program(1.0) 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, 1, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_188 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 1, 4, 64])]; + tensor var_190 = reshape(shape = var_189, x = var_188)[name = tensor("op_190")]; 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, 1, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_194 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor(0x1p-3)]; + tensor var_196 = mul(x = var_194, y = var_195)[name = tensor("op_196")]; + tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 1, 4, 64])]; + tensor var_198 = reshape(shape = var_197, x = var_196)[name = tensor("op_198")]; 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, 1, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_202 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 4, 64])]; + tensor var_204 = reshape(shape = var_203, x = var_202)[name = tensor("op_204")]; 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 input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_198)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_190)[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([1, 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 var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_215 = reshape(shape = var_214, x = sqrt_s_t_1)[name = tensor("op_215")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_215)[name = tensor("M_1")]; + tensor var_217 = mul(x = qk_1, y = M_1)[name = tensor("op_217")]; 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, 1, 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 v_1 = transpose(perm = v_1_perm_0, x = var_204)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_217, y = v_1)[name = tensor("inner_1")]; + tensor var_219_transpose_x_0 = const()[name = tensor("op_219_transpose_x_0"), val = tensor(false)]; + tensor var_219_transpose_y_0 = const()[name = tensor("op_219_transpose_y_0"), val = tensor(false)]; + tensor var_219 = matmul(transpose_x = var_219_transpose_x_0, transpose_y = var_219_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_219")]; + tensor var_220 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_220")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1, 1, 1])]; + tensor var_222 = reshape(shape = var_221, x = var_220)[name = tensor("op_222")]; + tensor cross_1 = mul(x = var_219, y = var_222)[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(0x1p+0)]; - 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 var_225 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_225")]; + tensor var_227_transpose_x_1 = const()[name = tensor("op_227_transpose_x_1"), val = tensor(true)]; + tensor var_227_transpose_y_1 = const()[name = tensor("op_227_transpose_y_1"), val = tensor(false)]; + tensor var_227 = matmul(transpose_x = var_227_transpose_x_1, transpose_y = var_227_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_227")]; + tensor new_kv_unnorm_1 = add(x = var_225, y = var_227)[name = tensor("new_kv_unnorm_1")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_229)[name = tensor("new_scale_1")]; + tensor var_231 = sqrt(x = new_scale_1)[name = tensor("op_231")]; + tensor var_232 = real_div(x = new_kv_unnorm_1, y = var_231)[name = tensor("op_232")]; + tensor var_233_perm_0 = const()[name = tensor("op_233_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, 1, 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 var_233 = transpose(perm = var_233_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_43, x = var_233)[name = tensor("out_3")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_237, x = out_3)[name = tensor("out_5")]; + tensor var_239 = silu(x = input_19)[name = tensor("op_239")]; + tensor input_21 = mul(x = var_239, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_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 var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 16, 256])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, true, true])]; + tensor var_250 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = window_1)[name = tensor("op_250")]; 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, x_3))[name = tensor("window_3")]; - 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)[name = tensor("input_21")]; + tensor window_3 = concat(axis = var_52, interleave = window_3_interleave_0, values = (var_250, x_3))[name = tensor("window_3")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_38, interleave = input_23_interleave_0, values = window_3)[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = 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 x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_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_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; - tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; - tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; - tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; - tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_275_split_sizes_0 = const()[name = tensor("op_275_split_sizes_0"), val = tensor([256, 256])]; + tensor var_275_axis_0 = const()[name = tensor("op_275_axis_0"), val = tensor(1)]; + tensor var_275_0, tensor var_275_1 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = inputs_3)[name = tensor("op_275")]; + tensor var_277 = sigmoid(x = var_275_1)[name = tensor("op_277")]; + tensor inputs_5 = mul(x = var_275_0, y = var_277)[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([1, 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 outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([1, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; - tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; - tensor var_282_end_mask_0 = const()[name = tensor("op_282_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_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; - tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; - tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_284)[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_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; - tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; - tensor input_41 = add(x = var_308, 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 var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, -1, 0])]; + tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 16, 256])]; + tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_308 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = conv_out_1)[name = tensor("op_308")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([1, 0, 2])]; + tensor var_310 = transpose(perm = var_310_perm_0, x = var_308)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_310)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor(0x1p-1)]; + tensor var_334 = mul(x = input_41, y = var_333)[name = tensor("op_334")]; + tensor input_43 = add(x = var_334, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = 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_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; - tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; - tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor(0x1p-1)]; + tensor var_364 = mul(x = input_53, y = var_363)[name = tensor("op_364")]; + tensor input_55 = add(x = var_364, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -381,139 +387,139 @@ program(1.0) 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_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; - tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor var_378 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1, 4, 64])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = tensor("op_380")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; - tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_384 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor(0x1p-3)]; + tensor var_386 = mul(x = var_384, y = var_385)[name = tensor("op_386")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 1, 4, 64])]; + tensor var_388 = reshape(shape = var_387, x = var_386)[name = tensor("op_388")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; - tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor var_392 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 4, 64])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; 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 input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_388)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_380)[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_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; - tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; - tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_405 = reshape(shape = var_404, x = sqrt_s_t_3)[name = tensor("op_405")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_405)[name = tensor("M_3")]; + tensor var_407 = mul(x = qk_3, y = M_3)[name = tensor("op_407")]; 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_368)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; - tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; - tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; - tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; - tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; - tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_394)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_407, y = v_3)[name = tensor("inner_3")]; + tensor var_409_transpose_x_0 = const()[name = tensor("op_409_transpose_x_0"), val = tensor(false)]; + tensor var_409_transpose_y_0 = const()[name = tensor("op_409_transpose_y_0"), val = tensor(false)]; + tensor var_409 = matmul(transpose_x = var_409_transpose_x_0, transpose_y = var_409_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_409")]; + tensor var_410 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_410")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1, 1, 1])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + tensor cross_3 = mul(x = var_409, y = var_412)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; - tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; - tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; - tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; - tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; - tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; - tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_415 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_415")]; + tensor var_417_transpose_x_1 = const()[name = tensor("op_417_transpose_x_1"), val = tensor(true)]; + tensor var_417_transpose_y_1 = const()[name = tensor("op_417_transpose_y_1"), val = tensor(false)]; + tensor var_417 = matmul(transpose_x = var_417_transpose_x_1, transpose_y = var_417_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_417")]; + tensor new_kv_unnorm_3 = add(x = var_415, y = var_417)[name = tensor("new_kv_unnorm_3")]; + tensor var_419 = const()[name = tensor("op_419"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_419)[name = tensor("new_scale_3")]; + tensor var_421 = sqrt(x = new_scale_3)[name = tensor("op_421")]; + tensor var_422 = real_div(x = new_kv_unnorm_3, y = var_421)[name = tensor("op_422")]; + tensor var_423_perm_0 = const()[name = tensor("op_423_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_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; - tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; - tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; - tensor input_59 = mul(x = var_403, 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 var_423 = transpose(perm = var_423_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_43, x = var_423)[name = tensor("out_9")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_427, x = out_9)[name = tensor("out_11")]; + tensor var_429 = silu(x = input_59)[name = tensor("op_429")]; + tensor input_61 = mul(x = var_429, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; - tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; - tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; - tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; - tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 1, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 16, 256])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_440 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = window_5)[name = tensor("op_440")]; 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_414, x_9))[name = tensor("window_7")]; - 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_7)[name = tensor("input_61")]; + tensor window_7 = concat(axis = var_52, interleave = window_7_interleave_0, values = (var_440, x_9))[name = tensor("window_7")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_38, interleave = input_63_interleave_0, values = window_7)[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = 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 x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_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_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; - tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; - tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; - tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; - tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_465_split_sizes_0 = const()[name = tensor("op_465_split_sizes_0"), val = tensor([256, 256])]; + tensor var_465_axis_0 = const()[name = tensor("op_465_axis_0"), val = tensor(1)]; + tensor var_465_0, tensor var_465_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = inputs_13)[name = tensor("op_465")]; + tensor var_467 = sigmoid(x = var_465_1)[name = tensor("op_467")]; + tensor inputs_15 = mul(x = var_465_0, y = var_467)[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([1, 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 outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([1, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; - tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; - tensor var_472_end_mask_0 = const()[name = tensor("op_472_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_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; - tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; - tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_474)[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_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; - tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; - tensor input_81 = add(x = var_498, 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 var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, -1, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 16, 256])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_498 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = conv_out_3)[name = tensor("op_498")]; + tensor var_500_perm_0 = const()[name = tensor("op_500_perm_0"), val = tensor([1, 0, 2])]; + tensor var_500 = transpose(perm = var_500_perm_0, x = var_498)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_500)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_523 = const()[name = tensor("op_523"), val = tensor(0x1p-1)]; + tensor var_524 = mul(x = input_81, y = var_523)[name = tensor("op_524")]; + tensor input_83 = add(x = var_524, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = 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_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; - tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; - tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor(0x1p-1)]; + tensor var_554 = mul(x = input_93, y = var_553)[name = tensor("op_554")]; + tensor input_95 = add(x = var_554, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -524,139 +530,139 @@ program(1.0) 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_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; - tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor var_568 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1, 4, 64])]; + tensor var_570 = reshape(shape = var_569, x = var_568)[name = tensor("op_570")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; - tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; - tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; - tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor var_574 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-3)]; + tensor var_576 = mul(x = var_574, y = var_575)[name = tensor("op_576")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1, 4, 64])]; + tensor var_578 = reshape(shape = var_577, x = var_576)[name = tensor("op_578")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; - tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor var_582 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; 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 input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_578)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_570)[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_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; - tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; - tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1, 1])]; + tensor var_595 = reshape(shape = var_594, x = sqrt_s_t_5)[name = tensor("op_595")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_595)[name = tensor("M_5")]; + tensor var_597 = mul(x = qk_5, y = M_5)[name = tensor("op_597")]; 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_558)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; - tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; - tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; - tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; - tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; - tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_584)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_597, y = v_5)[name = tensor("inner_5")]; + tensor var_599_transpose_x_0 = const()[name = tensor("op_599_transpose_x_0"), val = tensor(false)]; + tensor var_599_transpose_y_0 = const()[name = tensor("op_599_transpose_y_0"), val = tensor(false)]; + tensor var_599 = matmul(transpose_x = var_599_transpose_x_0, transpose_y = var_599_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_599")]; + tensor var_600 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_600")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 1, 1, 1])]; + tensor var_602 = reshape(shape = var_601, x = var_600)[name = tensor("op_602")]; + tensor cross_5 = mul(x = var_599, y = var_602)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; - tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; - tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; - tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; - tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; - tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; - tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; - tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_605 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_605")]; + tensor var_607_transpose_x_1 = const()[name = tensor("op_607_transpose_x_1"), val = tensor(true)]; + tensor var_607_transpose_y_1 = const()[name = tensor("op_607_transpose_y_1"), val = tensor(false)]; + tensor var_607 = matmul(transpose_x = var_607_transpose_x_1, transpose_y = var_607_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_607")]; + tensor new_kv_unnorm_5 = add(x = var_605, y = var_607)[name = tensor("new_kv_unnorm_5")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_609)[name = tensor("new_scale_5")]; + tensor var_611 = sqrt(x = new_scale_5)[name = tensor("op_611")]; + tensor var_612 = real_div(x = new_kv_unnorm_5, y = var_611)[name = tensor("op_612")]; + tensor var_613_perm_0 = const()[name = tensor("op_613_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_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; - tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; - tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; - tensor input_99 = mul(x = var_593, 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 var_613 = transpose(perm = var_613_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_43, x = var_613)[name = tensor("out_15")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_617, x = out_15)[name = tensor("out_17")]; + tensor var_619 = silu(x = input_99)[name = tensor("op_619")]; + tensor input_101 = mul(x = var_619, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 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_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; - tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; - tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; - tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor var_630_begin_0 = const()[name = tensor("op_630_begin_0"), val = tensor([0, 1, 0])]; + tensor var_630_end_0 = const()[name = tensor("op_630_end_0"), val = tensor([1, 16, 256])]; + tensor var_630_end_mask_0 = const()[name = tensor("op_630_end_mask_0"), val = tensor([true, true, true])]; + tensor var_630 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, x = window_9)[name = tensor("op_630")]; 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_604, x_15))[name = tensor("window_11")]; - 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_11)[name = tensor("input_101")]; + tensor window_11 = concat(axis = var_52, interleave = window_11_interleave_0, values = (var_630, x_15))[name = tensor("window_11")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_38, interleave = input_103_interleave_0, values = window_11)[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = 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 x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_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_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; - tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; - tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; - tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; - tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_655_split_sizes_0 = const()[name = tensor("op_655_split_sizes_0"), val = tensor([256, 256])]; + tensor var_655_axis_0 = const()[name = tensor("op_655_axis_0"), val = tensor(1)]; + tensor var_655_0, tensor var_655_1 = split(axis = var_655_axis_0, split_sizes = var_655_split_sizes_0, x = inputs_23)[name = tensor("op_655")]; + tensor var_657 = sigmoid(x = var_655_1)[name = tensor("op_657")]; + tensor inputs_25 = mul(x = var_655_0, y = var_657)[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([1, 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 outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([1, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; - tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; - tensor var_662_end_mask_0 = const()[name = tensor("op_662_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_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; - tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; - tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_664)[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_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; - tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; - tensor input_121 = add(x = var_688, 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 var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, -1, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 16, 256])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_688 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, x = conv_out_5)[name = tensor("op_688")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([1, 0, 2])]; + tensor var_690 = transpose(perm = var_690_perm_0, x = var_688)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_690)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1p-1)]; + tensor var_714 = mul(x = input_121, y = var_713)[name = tensor("op_714")]; + tensor input_123 = add(x = var_714, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = 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_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; - tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; - tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor(0x1p-1)]; + tensor var_744 = mul(x = input_133, y = var_743)[name = tensor("op_744")]; + tensor input_135 = add(x = var_744, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -667,175 +673,168 @@ program(1.0) 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_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; - tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor var_758 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 1, 4, 64])]; + tensor var_760 = reshape(shape = var_759, x = var_758)[name = tensor("op_760")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; - tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; - tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor var_764 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-3)]; + tensor var_766 = mul(x = var_764, y = var_765)[name = tensor("op_766")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 4, 64])]; + tensor var_768 = reshape(shape = var_767, x = var_766)[name = tensor("op_768")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; - tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor var_772 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor([1, 1, 4, 64])]; + tensor var_774 = reshape(shape = var_773, x = var_772)[name = tensor("op_774")]; 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 input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_768)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_760)[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_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; - tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; - tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor([1, 1])]; + tensor var_785 = reshape(shape = var_784, x = sqrt_s_t_7)[name = tensor("op_785")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_785)[name = tensor("M_7")]; + tensor var_787 = mul(x = qk_7, y = M_7)[name = tensor("op_787")]; 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_748)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; - tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; - tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; - tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; - tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; - tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_774)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_787, y = v_7)[name = tensor("inner_7")]; + tensor var_789_transpose_x_0 = const()[name = tensor("op_789_transpose_x_0"), val = tensor(false)]; + tensor var_789_transpose_y_0 = const()[name = tensor("op_789_transpose_y_0"), val = tensor(false)]; + tensor var_789 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_789")]; + tensor var_790 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_790")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 1, 1, 1])]; + tensor var_792 = reshape(shape = var_791, x = var_790)[name = tensor("op_792")]; + tensor cross_7 = mul(x = var_789, y = var_792)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; - tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; - tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; - tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; - tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; - tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; - tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; - tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_795 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_795")]; + tensor var_797_transpose_x_1 = const()[name = tensor("op_797_transpose_x_1"), val = tensor(true)]; + tensor var_797_transpose_y_1 = const()[name = tensor("op_797_transpose_y_1"), val = tensor(false)]; + tensor var_797 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_797")]; + tensor new_kv_unnorm_7 = add(x = var_795, y = var_797)[name = tensor("new_kv_unnorm_7")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_799)[name = tensor("new_scale_7")]; + tensor var_801 = sqrt(x = new_scale_7)[name = tensor("op_801")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_801)[name = tensor("nkv_1")]; + tensor var_803_perm_0 = const()[name = tensor("op_803_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_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; - tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; - tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; - tensor input_139 = mul(x = var_783, 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 var_803 = transpose(perm = var_803_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_43, x = var_803)[name = tensor("out_21")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_807, x = out_21)[name = tensor("out_23")]; + tensor var_809 = silu(x = input_139)[name = tensor("op_809")]; + tensor input_141 = mul(x = var_809, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; - tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; - tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; - tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor var_820_begin_0 = const()[name = tensor("op_820_begin_0"), val = tensor([0, 1, 0])]; + tensor var_820_end_0 = const()[name = tensor("op_820_end_0"), val = tensor([1, 16, 256])]; + tensor var_820_end_mask_0 = const()[name = tensor("op_820_end_mask_0"), val = tensor([true, true, true])]; + tensor var_820 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = window_13)[name = tensor("op_820")]; 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_794, x_21))[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)[name = tensor("input_141")]; + tensor window = concat(axis = var_52, interleave = window_interleave_0, values = (var_820, x_21))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_38, interleave = input_143_interleave_0, values = window)[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = 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 x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_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_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; - tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; - tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; - tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; - tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_845_split_sizes_0 = const()[name = tensor("op_845_split_sizes_0"), val = tensor([256, 256])]; + tensor var_845_axis_0 = const()[name = tensor("op_845_axis_0"), val = tensor(1)]; + tensor var_845_0, tensor var_845_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = inputs_33)[name = tensor("op_845")]; + tensor var_847 = sigmoid(x = var_845_1)[name = tensor("op_847")]; + tensor inputs_35 = mul(x = var_845_0, y = var_847)[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([1, 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 outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([1, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; - tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; - tensor var_852_end_mask_0 = const()[name = tensor("op_852_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_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; - tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; - tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_854)[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_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; - tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; - tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor var_878_begin_0 = const()[name = tensor("op_878_begin_0"), val = tensor([0, -1, 0])]; + tensor var_878_end_0 = const()[name = tensor("op_878_end_0"), val = tensor([1, 16, 256])]; + tensor var_878_end_mask_0 = const()[name = tensor("op_878_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_878 = slice_by_index(begin = var_878_begin_0, end = var_878_end_0, end_mask = var_878_end_mask_0, x = conv_out_7)[name = tensor("op_878")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([1, 0, 2])]; + tensor var_880 = transpose(perm = var_880_perm_0, x = var_878)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_880)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor(0x1p-1)]; + tensor var_904 = mul(x = input_161, y = var_903)[name = tensor("op_904")]; + tensor input_163 = add(x = var_904, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_40, 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_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 0, 1])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 256, 19])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = cat)[name = tensor("op_922")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_925 = reduce_l2_norm(axes = var_924, keep_dims = var_34, x = input_165)[name = tensor("op_925")]; 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_899)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; - tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; - tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; - tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; - tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; - tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; - tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; - tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor clip_0 = clip(alpha = var_48, beta = const_12, x = var_925)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_929_axis_0, values = (var_232, var_422, var_612, nkv_1))[name = tensor("op_929")]; + tensor var_931_axis_0 = const()[name = tensor("op_931_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_931_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_931")]; + tensor var_933_axis_0 = const()[name = tensor("op_933_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_933_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_933")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1001_axes_0 = const()[name = tensor("op_1001_axes_0"), val = tensor([2])]; + tensor var_1001 = expand_dims(axes = var_1001_axes_0, x = emb)[name = tensor("op_1001")]; 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_998)[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_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[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_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([6, 1, 256])]; - tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1001)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_41, interleave = input_167_interleave_0, values = (emb_exp, var_996))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1009_perm_0 = const()[name = tensor("op_1009_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([6, 1, 256])]; + tensor var_1009 = transpose(perm = var_1009_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1013, x = var_1009)[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])]; @@ -846,131 +845,131 @@ program(1.0) 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_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([6, 1, 4, 64])]; - tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor var_1021 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([6, 1, 4, 64])]; + tensor var_1023 = reshape(shape = var_1022, x = var_1021)[name = tensor("op_1023")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; - tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([6, 1, 4, 64])]; - tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor var_1027 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor(0x1p-3)]; + tensor var_1029 = mul(x = var_1027, y = var_1028)[name = tensor("op_1029")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([6, 1, 4, 64])]; + tensor var_1031 = reshape(shape = var_1030, x = var_1029)[name = tensor("op_1031")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([6, 1, 4, 64])]; - tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor var_1035 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1036 = const()[name = tensor("op_1036"), val = tensor([6, 1, 4, 64])]; + tensor var_1037 = reshape(shape = var_1036, x = var_1035)[name = tensor("op_1037")]; 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 input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_38, 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_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_28, 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_1028)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1031)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1023)[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_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; - tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; - tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; - tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; - tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor var_1050 = reshape(shape = var_1049, x = valid_mask)[name = tensor("op_1050")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1])]; + tensor var_1053 = reshape(shape = var_1052, x = sqrt_s_t_9)[name = tensor("op_1053")]; + tensor M_9 = real_div(x = var_1050, y = var_1053)[name = tensor("M_9")]; + tensor var_1055 = mul(x = qk_9, y = M_9)[name = tensor("op_1055")]; 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_1034)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; - tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; - tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; - tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; - tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; - tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; - tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; - tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1037)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1055, y = v_9)[name = tensor("inner_9")]; + tensor var_1057_transpose_x_0 = const()[name = tensor("op_1057_transpose_x_0"), val = tensor(false)]; + tensor var_1057_transpose_y_0 = const()[name = tensor("op_1057_transpose_y_0"), val = tensor(false)]; + tensor var_1057 = matmul(transpose_x = var_1057_transpose_x_0, transpose_y = var_1057_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1057")]; + tensor var_1058 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1, 1, 1, 1])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor cross_9 = mul(x = var_1057, y = var_1060)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; - tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; - tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; - tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; - tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; - tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; - tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; - tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; - tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; - tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; - tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; - tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 1, 1])]; + tensor var_1064 = reshape(shape = var_1063, x = valid_mask)[name = tensor("op_1064")]; + tensor v_masked_1 = mul(x = v_9, y = var_1064)[name = tensor("v_masked_1")]; + tensor var_1066 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1066")]; + tensor var_1068_transpose_x_1 = const()[name = tensor("op_1068_transpose_x_1"), val = tensor(true)]; + tensor var_1068_transpose_y_1 = const()[name = tensor("op_1068_transpose_y_1"), val = tensor(false)]; + tensor var_1068 = matmul(transpose_x = var_1068_transpose_x_1, transpose_y = var_1068_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1068")]; + tensor new_kv_unnorm_9 = add(x = var_1066, y = var_1068)[name = tensor("new_kv_unnorm_9")]; + tensor var_1070_keep_dims_0 = const()[name = tensor("op_1070_keep_dims_0"), val = tensor(false)]; + tensor var_1070 = reduce_sum(keep_dims = var_1070_keep_dims_0, x = valid_mask)[name = tensor("op_1070")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1])]; + tensor var_1072 = reshape(shape = var_1071, x = var_1070)[name = tensor("op_1072")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1072)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_28, 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_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; - tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1076 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1076")]; + tensor var_1077_perm_0 = const()[name = tensor("op_1077_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_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([6, 1, 256])]; - tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; - tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; - tensor input_171 = mul(x = var_1080, 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 var_1077 = transpose(perm = var_1077_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_43, x = var_1077)[name = tensor("out_27")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([6, 1, 256])]; + tensor out_29 = reshape(shape = var_1081, x = out_27)[name = tensor("out_29")]; + tensor var_1083 = silu(x = input_171)[name = tensor("op_1083")]; + tensor input_173 = mul(x = var_1083, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 6, 1, 256])]; - tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; - tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 6, 256])]; - tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_35, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 6, 1, 256])]; + tensor var_1094 = reshape(shape = var_1093, x = xt_1)[name = tensor("op_1094")]; + tensor var_1095_perm_0 = const()[name = tensor("op_1095_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 6, 256])]; + tensor var_1095 = transpose(perm = var_1095_perm_0, x = var_1094)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1098, x = var_1095)[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_1118 = 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 var_1121 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 1, 3, 256])]; - tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; - tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; - tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; - tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; - tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; - tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor var_1123 = reshape(shape = concat_1, x = var_1121)[name = tensor("op_1123")]; + tensor var_1124_axes_0 = const()[name = tensor("op_1124_axes_0"), val = tensor([0])]; + tensor var_1124 = expand_dims(axes = var_1124_axes_0, x = var_1123)[name = tensor("op_1124")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-2])]; + tensor var_1125 = transpose(perm = var_1125_perm_0, x = var_1124)[name = tensor("transpose_21")]; + tensor var_1126 = squeeze(axes = var_1126_axes_0, x = var_1125)[name = tensor("op_1126")]; 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, 1, 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_1123)[name = tensor("q_11")]; + 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_1126)[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, 1, 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_1123)[name = tensor("k_11")]; + 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_1126)[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, 1, 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_1123)[name = tensor("v_11")]; - tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([6, 4, 64])]; - tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + 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_1126)[name = tensor("v_11")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([6, 4, 64])]; + tensor var_1135 = reshape(shape = var_1134, x = q_11)[name = tensor("op_1135")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([6, 4, 64])]; - tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([6, 4, 64])]; + tensor var_1142 = reshape(shape = var_1141, x = k_11)[name = tensor("op_1142")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([6, 4, 64])]; - tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = v_11)[name = tensor("op_1149")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; - tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; - tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1135)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1152, x = q_13)[name = tensor("q_15")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1142)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1154, x = k_13)[name = tensor("k_15")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1149)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1156, 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)]; @@ -981,30 +980,30 @@ program(1.0) 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_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; - tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([6, 256])]; - tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[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_1165 = const()[name = tensor("op_1165"), val = tensor([6, 1, 256])]; - tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([2, 0, 1, 3])]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([6, 256])]; + tensor var_1160 = transpose(perm = var_1159, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1164, x = var_1160)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([6, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1168, 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_921, 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 input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_35, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 6, 256])]; - tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; - tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([6, 1, 256])]; - tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_35, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([1, 1, 6, 256])]; + tensor x_31 = reshape(shape = var_1188, x = xt_3)[name = tensor("x_31")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([6, 1, 256])]; + tensor var_1190 = transpose(perm = var_1190_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1194, x = var_1190)[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])]; @@ -1015,120 +1014,120 @@ program(1.0) 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_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([6, 1, 4, 64])]; - tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor var_1202 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([6, 1, 4, 64])]; + tensor var_1204 = reshape(shape = var_1203, x = var_1202)[name = tensor("op_1204")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; - tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; - tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([6, 1, 4, 64])]; - tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor var_1208 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor(0x1p-3)]; + tensor var_1210 = mul(x = var_1208, y = var_1209)[name = tensor("op_1210")]; + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([6, 1, 4, 64])]; + tensor var_1212 = reshape(shape = var_1211, x = var_1210)[name = tensor("op_1212")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([6, 1, 4, 64])]; - tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor var_1216 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([6, 1, 4, 64])]; + tensor var_1218 = reshape(shape = var_1217, x = var_1216)[name = tensor("op_1218")]; 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 input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_28, 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_1209)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1212)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1204)[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_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; - tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; - tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; - tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; - 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_1215)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; - tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; - tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; - tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; - tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; - tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; - tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; - tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; - tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; - tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; - tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; - tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; - tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor var_1234 = reshape(shape = var_1233, x = sqrt_s_t)[name = tensor("op_1234")]; + tensor M = real_div(x = var_1050, y = var_1234)[name = tensor("M")]; + tensor var_1236 = mul(x = qk, y = M)[name = tensor("op_1236")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1218)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1236, y = v_17)[name = tensor("inner_11")]; + tensor var_1238_transpose_x_0 = const()[name = tensor("op_1238_transpose_x_0"), val = tensor(false)]; + tensor var_1238_transpose_y_0 = const()[name = tensor("op_1238_transpose_y_0"), val = tensor(false)]; + tensor var_1238 = matmul(transpose_x = var_1238_transpose_x_0, transpose_y = var_1238_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1238")]; + tensor var_1239 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1, 1, 1, 1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor cross = mul(x = var_1238, y = var_1241)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1064)[name = tensor("v_masked")]; + tensor var_1247 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1247")]; + tensor var_1249_transpose_x_1 = const()[name = tensor("op_1249_transpose_x_1"), val = tensor(true)]; + tensor var_1249_transpose_y_1 = const()[name = tensor("op_1249_transpose_y_1"), val = tensor(false)]; + tensor var_1249 = matmul(transpose_x = var_1249_transpose_x_1, transpose_y = var_1249_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1249")]; + tensor new_kv_unnorm = add(x = var_1247, y = var_1249)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1072)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_28, 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_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258_perm_0 = const()[name = tensor("op_1258_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_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; - tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([6, 1, 256])]; - tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; - tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; - tensor input_189 = mul(x = var_1261, 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 var_1258 = transpose(perm = var_1258_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_43, x = var_1258)[name = tensor("out_33")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([6, 1, 256])]; + tensor out = reshape(shape = var_1262, x = out_33)[name = tensor("out")]; + tensor var_1264 = silu(x = input_189)[name = tensor("op_1264")]; + tensor input_191 = mul(x = var_1264, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 6, 1, 256])]; - tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; - tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 6, 256])]; - tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_35, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 6, 1, 256])]; + tensor var_1275 = reshape(shape = var_1274, x = xt_5)[name = tensor("op_1275")]; + tensor var_1276_perm_0 = const()[name = tensor("op_1276_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 6, 256])]; + tensor var_1276 = transpose(perm = var_1276_perm_0, x = var_1275)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1279, x = var_1276)[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_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1302 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 1, 3, 256])]; - tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; - tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; - tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; - tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; - tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; - tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor var_1304 = reshape(shape = concat_2, x = var_1302)[name = tensor("op_1304")]; + tensor var_1305_axes_0 = const()[name = tensor("op_1305_axes_0"), val = tensor([0])]; + tensor var_1305 = expand_dims(axes = var_1305_axes_0, x = var_1304)[name = tensor("op_1305")]; + tensor var_1306_perm_0 = const()[name = tensor("op_1306_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1307_axes_0 = const()[name = tensor("op_1307_axes_0"), val = tensor([-2])]; + tensor var_1306 = transpose(perm = var_1306_perm_0, x = var_1305)[name = tensor("transpose_8")]; + tensor var_1307 = squeeze(axes = var_1307_axes_0, x = var_1306)[name = tensor("op_1307")]; 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, 1, 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_1304)[name = tensor("q_19")]; + 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_1307)[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, 1, 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_1304)[name = tensor("k_19")]; + 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_1307)[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, 1, 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_1304)[name = tensor("v_19")]; - tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([6, 4, 64])]; - tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + 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_1307)[name = tensor("v_19")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([6, 4, 64])]; + tensor var_1316 = reshape(shape = var_1315, x = q_19)[name = tensor("op_1316")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([6, 4, 64])]; - tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([6, 4, 64])]; + tensor var_1323 = reshape(shape = var_1322, x = k_19)[name = tensor("op_1323")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([6, 4, 64])]; - tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = v_19)[name = tensor("op_1330")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; - tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1316)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1333, x = q_21)[name = tensor("q")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1323)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1335, x = k_21)[name = tensor("k")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1330)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1337, 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)]; @@ -1139,34 +1138,34 @@ program(1.0) 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_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([6, 256])]; - tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[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_1346 = const()[name = tensor("op_1346"), val = tensor([6, 1, 256])]; - tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([2, 0, 1, 3])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([6, 256])]; + tensor var_1341 = transpose(perm = var_1340, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1345, x = var_1341)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1349 = const()[name = tensor("op_1349"), val = tensor([6, 1, 256])]; + tensor attn_output = reshape(shape = var_1349, 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_921, 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 input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_35, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 6, 256])]; - tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; - tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; - tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_35, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1, 1, 6, 256])]; + tensor input = reshape(shape = var_1369, x = xt)[name = tensor("input")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([-1])]; + tensor var_1372 = reduce_l2_norm(axes = var_1371, keep_dims = var_34, x = input)[name = tensor("op_1372")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; - tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor clip_5 = clip(alpha = var_48, beta = const_42, x = var_1372)[name = tensor("clip_5")]; + tensor var_1374 = real_div(x = input, y = clip_5)[name = tensor("op_1374")]; 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([1, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1374)[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)]; @@ -1175,10 +1174,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 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 = matmul_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; - tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; - tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1378")]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1380_axis_0, values = (var_1076, nkv))[name = tensor("op_1380")]; + tensor var_1382_axis_0 = const()[name = tensor("op_1382_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1382_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1382")]; } -> (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