| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "1.13.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] |
| { |
| func main<ios18>(tensor<fp32, [2, 12, 4, 64, 64]> dec_ret_kv, tensor<fp32, [2, 12, 4]> dec_ret_scale, tensor<fp32, [1]> decode, tensor<fp32, [4, 1, 15, 256]> enc_conv_cache, tensor<fp32, [4, 1, 4, 64, 64]> enc_ret_kv, tensor<fp32, [4, 1, 4]> enc_ret_scale, tensor<fp32, [1, 1, 345]> frame, tensor<fp32, [1]> ingest, tensor<fp32, [1, 19, 256]> top_buffer) { |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn2_out_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn2_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn2_out_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn2_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1152)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn2_out_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn2_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263360)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn2_out_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn2_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264448)))]; |
| tensor<fp32, [256]> model_cnn_bias = const()[name = string("model_cnn_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526656)))]; |
| tensor<fp32, [256, 256, 19]> model_cnn_weight = const()[name = string("model_cnn_weight"), val = tensor<fp32, [256, 256, 19]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527744)))]; |
| tensor<fp32, [256]> model_enc_encoder_input_projection_linear_bias = const()[name = string("model_enc_encoder_input_projection_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5508544)))]; |
| tensor<fp32, [256, 345]> model_enc_encoder_input_projection_linear_weight = const()[name = string("model_enc_encoder_input_projection_linear_weight"), val = tensor<fp32, [256, 345]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5509632)))]; |
| tensor<fp32, [256]> model_enc_encoder_layer_norm_bias = const()[name = string("model_enc_encoder_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5862976)))]; |
| tensor<fp32, [256]> model_enc_encoder_layer_norm_weight = const()[name = string("model_enc_encoder_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5864064)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_0_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5865152)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_0_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5866240)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5867328)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5871488)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6920128)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6921216)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_layer_norm_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7969856)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_layer_norm_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7970944)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7972032)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7973120)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8235328)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8236416)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8498624)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8499712)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8761920)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8763008)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_bias = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9025216)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_weight = const()[name = string("model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9026304)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_2_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9288512)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_2_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9289600)))]; |
| tensor<fp32, [512]> model_enc_encoder_layers_0_sequential_2_module_sequential_2_conv_bias = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9290688)))]; |
| tensor<fp32, [512, 256, 1]> model_enc_encoder_layers_0_sequential_2_module_sequential_2_conv_weight = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9292800)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_2_module_sequential_7_conv_bias = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9817152)))]; |
| tensor<fp32, [256, 256, 1]> model_enc_encoder_layers_0_sequential_2_module_sequential_7_conv_weight = const()[name = string("model_enc_encoder_layers_0_sequential_2_module_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9818240)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_3_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10080448)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_3_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10081536)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10082624)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10086784)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11135424)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11136512)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_4_bias = const()[name = string("model_enc_encoder_layers_0_sequential_4_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12185152)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_0_sequential_4_weight = const()[name = string("model_enc_encoder_layers_0_sequential_4_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12186240)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_0_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12187328)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_0_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12188416)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12189504)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12193664)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13242304)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13243392)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_layer_norm_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14292032)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_layer_norm_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14293120)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14294208)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14295296)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14557504)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14558592)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14820800)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14821888)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15084096)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15085184)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_bias = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15347392)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_weight = const()[name = string("model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15348480)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_2_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15610688)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_2_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15611776)))]; |
| tensor<fp32, [512]> model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_bias = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15612864)))]; |
| tensor<fp32, [512, 256, 1]> model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_weight = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15614976)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_bias = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16139328)))]; |
| tensor<fp32, [256, 256, 1]> model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_weight = const()[name = string("model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16140416)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_3_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16402624)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_3_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16403712)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16404800)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16408960)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17457600)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17458688)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_4_bias = const()[name = string("model_enc_encoder_layers_1_sequential_4_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18507328)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_1_sequential_4_weight = const()[name = string("model_enc_encoder_layers_1_sequential_4_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18508416)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_0_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18509504)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_0_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18510592)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18511680)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18515840)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19564480)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19565568)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_layer_norm_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20614208)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_layer_norm_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20615296)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20616384)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20617472)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20879680)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20880768)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21142976)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21144064)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21406272)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21407360)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_bias = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21669568)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_weight = const()[name = string("model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21670656)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_2_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21932864)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_2_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21933952)))]; |
| tensor<fp32, [512]> model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_bias = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21935040)))]; |
| tensor<fp32, [512, 256, 1]> model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_weight = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21937152)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_bias = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22461504)))]; |
| tensor<fp32, [256, 256, 1]> model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_weight = const()[name = string("model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22462592)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_3_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22724800)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_3_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22725888)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22726976)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22731136)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23779776)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23780864)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_4_bias = const()[name = string("model_enc_encoder_layers_2_sequential_4_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24829504)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_2_sequential_4_weight = const()[name = string("model_enc_encoder_layers_2_sequential_4_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24830592)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_0_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24831680)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_0_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24832768)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24833856)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24838016)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25886656)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25887744)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_layer_norm_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_layer_norm_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26936384)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_layer_norm_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26937472)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26938560)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26939648)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27201856)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27202944)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27465152)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27466240)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27728448)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27729536)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_bias = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27991744)))]; |
| tensor<fp32, [256, 256]> model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_weight = const()[name = string("model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27992832)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_2_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28255040)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_2_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28256128)))]; |
| tensor<fp32, [512]> model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_bias = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28257216)))]; |
| tensor<fp32, [512, 256, 1]> model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_weight = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_weight"), val = tensor<fp32, [512, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28259328)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_bias = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28783680)))]; |
| tensor<fp32, [256, 256, 1]> model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_weight = const()[name = string("model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_weight"), val = tensor<fp32, [256, 256, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28784768)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_3_module_sequential_0_bias = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29046976)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_3_module_sequential_0_weight = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_0_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29048064)))]; |
| tensor<fp32, [1024]> model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_bias = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29049152)))]; |
| tensor<fp32, [1024, 256]> model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_weight = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29053312)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_bias = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30101952)))]; |
| tensor<fp32, [256, 1024]> model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_weight = const()[name = string("model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30103040)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_4_bias = const()[name = string("model_enc_encoder_layers_3_sequential_4_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31151680)))]; |
| tensor<fp32, [256]> model_enc_encoder_layers_3_sequential_4_weight = const()[name = string("model_enc_encoder_layers_3_sequential_4_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31152768)))]; |
| tensor<fp32, [256]> model_dec_convert_bias = const()[name = string("model_dec_convert_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31153856)))]; |
| tensor<fp32, [256, 512]> model_dec_convert_weight = const()[name = string("model_dec_convert_weight"), val = tensor<fp32, [256, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31154944)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn1_q_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31679296)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn1_q_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31680384)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn1_k_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31942592)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn1_k_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31943680)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn1_v_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32205888)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn1_v_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32206976)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn1_g_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32469184)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn1_g_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32470272)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_self_attn1_out_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32732480)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_0_self_attn1_out_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_0_self_attn1_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32733568)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm11_bias = const()[name = string("model_dec_attractor_decoder_layers_0_norm11_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32995776)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm11_weight = const()[name = string("model_dec_attractor_decoder_layers_0_norm11_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32996864)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm21_bias = const()[name = string("model_dec_attractor_decoder_layers_0_norm21_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32997952)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm21_weight = const()[name = string("model_dec_attractor_decoder_layers_0_norm21_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32999040)))]; |
| tensor<fp32, [2048]> model_dec_attractor_decoder_layers_0_linear1_bias = const()[name = string("model_dec_attractor_decoder_layers_0_linear1_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33000128)))]; |
| tensor<fp32, [2048, 256]> model_dec_attractor_decoder_layers_0_linear1_weight = const()[name = string("model_dec_attractor_decoder_layers_0_linear1_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33008384)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_linear2_bias = const()[name = string("model_dec_attractor_decoder_layers_0_linear2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35105600)))]; |
| tensor<fp32, [256, 2048]> model_dec_attractor_decoder_layers_0_linear2_weight = const()[name = string("model_dec_attractor_decoder_layers_0_linear2_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(35106688)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm22_bias = const()[name = string("model_dec_attractor_decoder_layers_0_norm22_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37203904)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_0_norm22_weight = const()[name = string("model_dec_attractor_decoder_layers_0_norm22_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37204992)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn1_q_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_q_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37206080)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn1_q_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_q_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37207168)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn1_k_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_k_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37469376)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn1_k_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_k_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37470464)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn1_v_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_v_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37732672)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn1_v_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_v_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37733760)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn1_g_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_g_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37995968)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn1_g_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_g_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37997056)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_self_attn1_out_proj_bias = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_out_proj_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38259264)))]; |
| tensor<fp32, [256, 256]> model_dec_attractor_decoder_layers_1_self_attn1_out_proj_weight = const()[name = string("model_dec_attractor_decoder_layers_1_self_attn1_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38260352)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm11_bias = const()[name = string("model_dec_attractor_decoder_layers_1_norm11_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38522560)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm11_weight = const()[name = string("model_dec_attractor_decoder_layers_1_norm11_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38523648)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm21_bias = const()[name = string("model_dec_attractor_decoder_layers_1_norm21_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38524736)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm21_weight = const()[name = string("model_dec_attractor_decoder_layers_1_norm21_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38525824)))]; |
| tensor<fp32, [2048]> model_dec_attractor_decoder_layers_1_linear1_bias = const()[name = string("model_dec_attractor_decoder_layers_1_linear1_bias"), val = tensor<fp32, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38526912)))]; |
| tensor<fp32, [2048, 256]> model_dec_attractor_decoder_layers_1_linear1_weight = const()[name = string("model_dec_attractor_decoder_layers_1_linear1_weight"), val = tensor<fp32, [2048, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38535168)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_linear2_bias = const()[name = string("model_dec_attractor_decoder_layers_1_linear2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40632384)))]; |
| tensor<fp32, [256, 2048]> model_dec_attractor_decoder_layers_1_linear2_weight = const()[name = string("model_dec_attractor_decoder_layers_1_linear2_weight"), val = tensor<fp32, [256, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40633472)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm22_bias = const()[name = string("model_dec_attractor_decoder_layers_1_norm22_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42730688)))]; |
| tensor<fp32, [256]> model_dec_attractor_decoder_layers_1_norm22_weight = const()[name = string("model_dec_attractor_decoder_layers_1_norm22_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42731776)))]; |
| tensor<int32, [3]> var_765 = const()[name = string("op_765"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<fp32, [1, 1, 1]> ingest_scalar = reshape(shape = var_765, x = ingest)[name = string("ingest_scalar")]; |
| tensor<int32, [3]> var_776 = const()[name = string("op_776"), val = tensor<int32, [3]>([1, 1, 1])]; |
| tensor<fp32, [1, 1, 1]> decode_scalar = reshape(shape = var_776, x = decode)[name = string("decode_scalar")]; |
| tensor<int32, [2]> var_786 = const()[name = string("op_786"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<fp32, [1, 1]> ingest_vec = reshape(shape = var_786, x = ingest)[name = string("ingest_vec")]; |
| tensor<int32, [2]> var_796 = const()[name = string("op_796"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<fp32, [1, 1]> decode_vec = reshape(shape = var_796, x = decode)[name = string("decode_vec")]; |
| tensor<fp32, [1, 1, 256]> input_1 = linear(bias = model_enc_encoder_input_projection_linear_bias, weight = model_enc_encoder_input_projection_linear_weight, x = frame)[name = string("linear_0")]; |
| fp32 var_803 = const()[name = string("op_803"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_3_axes_0 = const()[name = string("input_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_3 = layer_norm(axes = input_3_axes_0, beta = model_enc_encoder_layer_norm_bias, epsilon = var_803, gamma = model_enc_encoder_layer_norm_weight, x = input_1)[name = string("input_3")]; |
| tensor<int32, [5]> old_kv_1_begin_0 = const()[name = string("old_kv_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_1_end_0 = const()[name = string("old_kv_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_1_end_mask_0 = const()[name = string("old_kv_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_1_squeeze_mask_0 = const()[name = string("old_kv_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, 64, 64]> old_kv_1 = slice_by_index(begin = old_kv_1_begin_0, end = old_kv_1_end_0, end_mask = old_kv_1_end_mask_0, squeeze_mask = old_kv_1_squeeze_mask_0, x = enc_ret_kv)[name = string("old_kv_1")]; |
| tensor<int32, [3]> old_scale_1_begin_0 = const()[name = string("old_scale_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> old_scale_1_end_0 = const()[name = string("old_scale_1_end_0"), val = tensor<int32, [3]>([1, 1, 4])]; |
| tensor<bool, [3]> old_scale_1_end_mask_0 = const()[name = string("old_scale_1_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_1_squeeze_mask_0 = const()[name = string("old_scale_1_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [1, 4]> old_scale_1 = slice_by_index(begin = old_scale_1_begin_0, end = old_scale_1_end_0, end_mask = old_scale_1_end_mask_0, squeeze_mask = old_scale_1_squeeze_mask_0, x = enc_ret_scale)[name = string("old_scale_1")]; |
| tensor<int32, [4]> old_cache_1_begin_0 = const()[name = string("old_cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])]; |
| tensor<int32, [4]> old_cache_1_end_0 = const()[name = string("old_cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 15, 256])]; |
| tensor<bool, [4]> old_cache_1_end_mask_0 = const()[name = string("old_cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; |
| tensor<bool, [4]> old_cache_1_squeeze_mask_0 = const()[name = string("old_cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; |
| tensor<fp32, [1, 15, 256]> old_cache_1 = slice_by_index(begin = old_cache_1_begin_0, end = old_cache_1_end_0, end_mask = old_cache_1_end_mask_0, squeeze_mask = old_cache_1_squeeze_mask_0, x = enc_conv_cache)[name = string("old_cache_1")]; |
| fp32 var_819 = const()[name = string("op_819"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_5_axes_0 = const()[name = string("input_5_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_5 = layer_norm(axes = input_5_axes_0, beta = model_enc_encoder_layers_0_sequential_0_module_sequential_0_bias, epsilon = var_819, gamma = model_enc_encoder_layers_0_sequential_0_module_sequential_0_weight, x = input_3)[name = string("input_5")]; |
| tensor<fp32, [1, 1, 1024]> inputs_1 = linear(bias = model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_0_sequential_0_module_sequential_1_linear_weight, x = input_5)[name = string("linear_1")]; |
| tensor<fp32, [1, 1, 1024]> input_7 = silu(x = inputs_1)[name = string("input_7")]; |
| tensor<fp32, [1, 1, 256]> input_11 = linear(bias = model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_0_sequential_0_module_sequential_4_linear_weight, x = input_7)[name = string("linear_2")]; |
| fp32 var_843 = const()[name = string("op_843"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_844 = mul(x = input_11, y = var_843)[name = string("op_844")]; |
| tensor<fp32, [1, 1, 256]> input_13 = add(x = var_844, y = input_3)[name = string("input_13")]; |
| fp32 var_850 = const()[name = string("op_850"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_1_axes_0 = const()[name = string("x_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_1 = layer_norm(axes = x_1_axes_0, beta = model_enc_encoder_layers_0_sequential_1_module_layer_norm_bias, epsilon = var_850, gamma = model_enc_encoder_layers_0_sequential_1_module_layer_norm_weight, x = input_13)[name = string("x_1")]; |
| tensor<fp32, [1, 1, 256]> q_1 = linear(bias = model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_bias, weight = model_enc_encoder_layers_0_sequential_1_module_self_attn_q_proj_weight, x = x_1)[name = string("linear_3")]; |
| tensor<fp32, [1, 1, 256]> k_1 = linear(bias = model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_bias, weight = model_enc_encoder_layers_0_sequential_1_module_self_attn_k_proj_weight, x = x_1)[name = string("linear_4")]; |
| tensor<fp32, [1, 1, 256]> v_1 = linear(bias = model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_bias, weight = model_enc_encoder_layers_0_sequential_1_module_self_attn_v_proj_weight, x = x_1)[name = string("linear_5")]; |
| tensor<fp32, [1, 1, 256]> input_17 = linear(bias = model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_bias, weight = model_enc_encoder_layers_0_sequential_1_module_self_attn_g_proj_weight, x = x_1)[name = string("linear_6")]; |
| fp32 var_881 = const()[name = string("op_881"), val = fp32(0x1p-3)]; |
| tensor<fp32, [1, 1, 256]> k_3 = mul(x = k_1, y = var_881)[name = string("k_3")]; |
| tensor<int32, [4]> var_885 = const()[name = string("op_885"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_886 = reshape(shape = var_885, x = q_1)[name = string("op_886")]; |
| tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_892 = const()[name = string("op_892"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_893 = reshape(shape = var_892, x = k_3)[name = string("op_893")]; |
| tensor<int32, [4]> k_5_perm_0 = const()[name = string("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_900 = const()[name = string("op_900"), val = tensor<int32, [4]>([1, 4, 64, 1])]; |
| tensor<fp32, [1, 4, 64, 1]> v_3 = reshape(shape = var_900, x = v_1)[name = string("v_3")]; |
| tensor<fp32, [1, 4, 1, 64]> k_5 = transpose(perm = k_5_perm_0, x = var_893)[name = string("transpose_50")]; |
| tensor<fp32, [1, 4, 64, 64]> kv_1 = mul(x = k_5, y = v_3)[name = string("kv_1")]; |
| fp32 var_916 = const()[name = string("op_916"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 4]> candidate_scale_1 = add(x = old_scale_1, y = var_916)[name = string("candidate_scale_1")]; |
| tensor<fp32, [1, 4]> var_918 = sqrt(x = old_scale_1)[name = string("op_918")]; |
| tensor<fp32, [1, 4]> var_920 = sqrt(x = candidate_scale_1)[name = string("op_920")]; |
| tensor<fp32, [1, 4]> var_921 = real_div(x = var_918, y = var_920)[name = string("op_921")]; |
| tensor<int32, [1]> var_923_axes_0 = const()[name = string("op_923_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_923 = expand_dims(axes = var_923_axes_0, x = var_921)[name = string("op_923")]; |
| tensor<int32, [1]> blend_1_axes_0 = const()[name = string("blend_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> blend_1 = expand_dims(axes = blend_1_axes_0, x = var_923)[name = string("blend_1")]; |
| tensor<fp32, [1, 4, 64, 64]> var_926 = mul(x = old_kv_1, y = blend_1)[name = string("op_926")]; |
| tensor<int32, [1]> var_929_axes_0 = const()[name = string("op_929_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_929 = expand_dims(axes = var_929_axes_0, x = var_920)[name = string("op_929")]; |
| tensor<int32, [1]> var_931_axes_0 = const()[name = string("op_931_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> var_931 = expand_dims(axes = var_931_axes_0, x = var_929)[name = string("op_931")]; |
| tensor<fp32, [1, 4, 64, 64]> var_932 = real_div(x = kv_1, y = var_931)[name = string("op_932")]; |
| tensor<fp32, [1, 4, 64, 64]> candidate_kv_1 = add(x = var_926, y = var_932)[name = string("candidate_kv_1")]; |
| tensor<fp32, [1, 4, 1, 64]> q_3 = transpose(perm = q_3_perm_0, x = var_886)[name = string("transpose_51")]; |
| tensor<fp32, [1, 4, 64, 64]> var_935 = mul(x = q_3, y = candidate_kv_1)[name = string("op_935")]; |
| tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_15_keep_dims_0 = const()[name = string("input_15_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 64]> input_15 = reduce_sum(axes = input_15_axes_0, keep_dims = input_15_keep_dims_0, x = var_935)[name = string("input_15")]; |
| fp32 var_942 = const()[name = string("op_942"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_946_axes_0 = const()[name = string("op_946_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 64]> var_946 = layer_norm(axes = var_946_axes_0, epsilon = var_942, x = input_15)[name = string("op_946")]; |
| tensor<int32, [3]> var_948 = const()[name = string("op_948"), val = tensor<int32, [3]>([1, 1, 256])]; |
| tensor<fp32, [1, 1, 256]> output_1 = reshape(shape = var_948, x = var_946)[name = string("output_1")]; |
| tensor<fp32, [1, 1, 256]> var_950 = silu(x = input_17)[name = string("op_950")]; |
| tensor<fp32, [1, 1, 256]> input_19 = mul(x = var_950, y = output_1)[name = string("input_19")]; |
| tensor<fp32, [1, 1, 256]> input_21 = linear(bias = model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_bias, weight = model_enc_encoder_layers_0_sequential_1_module_self_attn_out_proj_weight, x = input_19)[name = string("linear_7")]; |
| tensor<fp32, [1, 1, 256]> input_23 = add(x = input_13, y = input_21)[name = string("input_23")]; |
| fp32 var_961 = const()[name = string("op_961"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_3_axes_0 = const()[name = string("x_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_3 = layer_norm(axes = x_3_axes_0, beta = model_enc_encoder_layers_0_sequential_2_module_sequential_0_bias, epsilon = var_961, gamma = model_enc_encoder_layers_0_sequential_2_module_sequential_0_weight, x = input_23)[name = string("x_3")]; |
| tensor<int32, [3]> input_25_perm_0 = const()[name = string("input_25_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string inputs_3_pad_type_0 = const()[name = string("inputs_3_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> inputs_3_strides_0 = const()[name = string("inputs_3_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> inputs_3_pad_0 = const()[name = string("inputs_3_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> inputs_3_dilations_0 = const()[name = string("inputs_3_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 inputs_3_groups_0 = const()[name = string("inputs_3_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_25 = transpose(perm = input_25_perm_0, x = x_3)[name = string("transpose_49")]; |
| tensor<fp32, [1, 512, 1]> inputs_3 = conv(bias = model_enc_encoder_layers_0_sequential_2_module_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 = model_enc_encoder_layers_0_sequential_2_module_sequential_2_conv_weight, x = input_25)[name = string("inputs_3")]; |
| tensor<int32, [2]> var_982_split_sizes_0 = const()[name = string("op_982_split_sizes_0"), val = tensor<int32, [2]>([256, 256])]; |
| int32 var_982_axis_0 = const()[name = string("op_982_axis_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> var_982_0, tensor<fp32, [1, 256, 1]> var_982_1 = split(axis = var_982_axis_0, split_sizes = var_982_split_sizes_0, x = inputs_3)[name = string("op_982")]; |
| tensor<fp32, [1, 256, 1]> var_984 = sigmoid(x = var_982_1)[name = string("op_984")]; |
| tensor<fp32, [1, 256, 1]> current_1 = mul(x = var_982_0, y = var_984)[name = string("current_1")]; |
| tensor<int32, [3]> cache_1_perm_0 = const()[name = string("cache_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| int32 var_990 = const()[name = string("op_990"), val = int32(2)]; |
| bool depthwise_window_1_interleave_0 = const()[name = string("depthwise_window_1_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 256, 15]> cache_1 = transpose(perm = cache_1_perm_0, x = old_cache_1)[name = string("transpose_48")]; |
| tensor<fp32, [1, 256, 16]> depthwise_window_1 = concat(axis = var_990, interleave = depthwise_window_1_interleave_0, values = (cache_1, current_1))[name = string("depthwise_window_1")]; |
| string input_27_pad_type_0 = const()[name = string("input_27_pad_type_0"), val = string("valid")]; |
| int32 input_27_groups_0 = const()[name = string("input_27_groups_0"), val = int32(256)]; |
| tensor<int32, [1]> input_27_strides_0 = const()[name = string("input_27_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_27_pad_0 = const()[name = string("input_27_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_27_dilations_0 = const()[name = string("input_27_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [256, 1, 16]> const_34 = const()[name = string("const_34"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42732864)))]; |
| tensor<fp32, [256]> const_35 = const()[name = string("const_35"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42749312)))]; |
| tensor<fp32, [1, 256, 1]> inputs_5 = conv(bias = const_35, dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = const_34, x = depthwise_window_1)[name = string("inputs_5")]; |
| tensor<int32, [3]> var_1022_begin_0 = const()[name = string("op_1022_begin_0"), val = tensor<int32, [3]>([0, 0, 1])]; |
| tensor<int32, [3]> var_1022_end_0 = const()[name = string("op_1022_end_0"), val = tensor<int32, [3]>([1, 256, 16])]; |
| tensor<bool, [3]> var_1022_end_mask_0 = const()[name = string("op_1022_end_mask_0"), val = tensor<bool, [3]>([true, true, true])]; |
| tensor<fp32, [1, 256, 15]> var_1022 = slice_by_index(begin = var_1022_begin_0, end = var_1022_end_0, end_mask = var_1022_end_mask_0, x = depthwise_window_1)[name = string("op_1022")]; |
| tensor<int32, [3]> candidate_conv_1_perm_0 = const()[name = string("candidate_conv_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 256, 1]> input_29 = silu(x = inputs_5)[name = string("input_29")]; |
| string input_31_pad_type_0 = const()[name = string("input_31_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> input_31_strides_0 = const()[name = string("input_31_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_31_pad_0 = const()[name = string("input_31_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_31_dilations_0 = const()[name = string("input_31_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 input_31_groups_0 = const()[name = string("input_31_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_31 = conv(bias = model_enc_encoder_layers_0_sequential_2_module_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 = model_enc_encoder_layers_0_sequential_2_module_sequential_7_conv_weight, x = input_29)[name = string("input_31")]; |
| tensor<int32, [3]> conv_output_1_perm_0 = const()[name = string("conv_output_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 1, 256]> conv_output_1 = transpose(perm = conv_output_1_perm_0, x = input_31)[name = string("transpose_46")]; |
| tensor<fp32, [1, 1, 256]> input_33 = add(x = input_23, y = conv_output_1)[name = string("input_33")]; |
| fp32 var_1058 = const()[name = string("op_1058"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_35_axes_0 = const()[name = string("input_35_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_35 = layer_norm(axes = input_35_axes_0, beta = model_enc_encoder_layers_0_sequential_3_module_sequential_0_bias, epsilon = var_1058, gamma = model_enc_encoder_layers_0_sequential_3_module_sequential_0_weight, x = input_33)[name = string("input_35")]; |
| tensor<fp32, [1, 1, 1024]> inputs_7 = linear(bias = model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_0_sequential_3_module_sequential_1_linear_weight, x = input_35)[name = string("linear_8")]; |
| tensor<fp32, [1, 1, 1024]> input_37 = silu(x = inputs_7)[name = string("input_37")]; |
| tensor<fp32, [1, 1, 256]> input_41 = linear(bias = model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_0_sequential_3_module_sequential_4_linear_weight, x = input_37)[name = string("linear_9")]; |
| fp32 var_1082 = const()[name = string("op_1082"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1083 = mul(x = input_41, y = var_1082)[name = string("op_1083")]; |
| tensor<fp32, [1, 1, 256]> input_43 = add(x = var_1083, y = input_33)[name = string("input_43")]; |
| fp32 var_1089 = const()[name = string("op_1089"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_45 = layer_norm(axes = input_45_axes_0, beta = model_enc_encoder_layers_0_sequential_4_bias, epsilon = var_1089, gamma = model_enc_encoder_layers_0_sequential_4_weight, x = input_43)[name = string("input_45")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1096 = sub(x = candidate_kv_1, y = old_kv_1)[name = string("op_1096")]; |
| tensor<int32, [1]> var_1098_axes_0 = const()[name = string("op_1098_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 1, 1]> var_1098 = expand_dims(axes = var_1098_axes_0, x = ingest_scalar)[name = string("op_1098")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1099 = mul(x = var_1096, y = var_1098)[name = string("op_1099")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1101 = add(x = old_kv_1, y = var_1099)[name = string("op_1101")]; |
| tensor<fp32, [1, 4]> var_1103 = sub(x = candidate_scale_1, y = old_scale_1)[name = string("op_1103")]; |
| tensor<fp32, [1, 4]> var_1104 = mul(x = var_1103, y = ingest_vec)[name = string("op_1104")]; |
| tensor<fp32, [1, 4]> var_1106 = add(x = old_scale_1, y = var_1104)[name = string("op_1106")]; |
| tensor<fp32, [1, 15, 256]> candidate_conv_1 = transpose(perm = candidate_conv_1_perm_0, x = var_1022)[name = string("transpose_47")]; |
| tensor<fp32, [1, 15, 256]> var_1108 = sub(x = candidate_conv_1, y = old_cache_1)[name = string("op_1108")]; |
| tensor<fp32, [1, 15, 256]> var_1109 = mul(x = var_1108, y = ingest_scalar)[name = string("op_1109")]; |
| tensor<fp32, [1, 15, 256]> var_1111 = add(x = old_cache_1, y = var_1109)[name = string("op_1111")]; |
| tensor<int32, [5]> old_kv_3_begin_0 = const()[name = string("old_kv_3_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_3_end_0 = const()[name = string("old_kv_3_end_0"), val = tensor<int32, [5]>([2, 1, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_3_end_mask_0 = const()[name = string("old_kv_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_3_squeeze_mask_0 = const()[name = string("old_kv_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, 64, 64]> old_kv_3 = slice_by_index(begin = old_kv_3_begin_0, end = old_kv_3_end_0, end_mask = old_kv_3_end_mask_0, squeeze_mask = old_kv_3_squeeze_mask_0, x = enc_ret_kv)[name = string("old_kv_3")]; |
| tensor<int32, [3]> old_scale_3_begin_0 = const()[name = string("old_scale_3_begin_0"), val = tensor<int32, [3]>([1, 0, 0])]; |
| tensor<int32, [3]> old_scale_3_end_0 = const()[name = string("old_scale_3_end_0"), val = tensor<int32, [3]>([2, 1, 4])]; |
| tensor<bool, [3]> old_scale_3_end_mask_0 = const()[name = string("old_scale_3_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_3_squeeze_mask_0 = const()[name = string("old_scale_3_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [1, 4]> old_scale_3 = slice_by_index(begin = old_scale_3_begin_0, end = old_scale_3_end_0, end_mask = old_scale_3_end_mask_0, squeeze_mask = old_scale_3_squeeze_mask_0, x = enc_ret_scale)[name = string("old_scale_3")]; |
| tensor<int32, [4]> old_cache_3_begin_0 = const()[name = string("old_cache_3_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])]; |
| tensor<int32, [4]> old_cache_3_end_0 = const()[name = string("old_cache_3_end_0"), val = tensor<int32, [4]>([2, 1, 15, 256])]; |
| tensor<bool, [4]> old_cache_3_end_mask_0 = const()[name = string("old_cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; |
| tensor<bool, [4]> old_cache_3_squeeze_mask_0 = const()[name = string("old_cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; |
| tensor<fp32, [1, 15, 256]> old_cache_3 = slice_by_index(begin = old_cache_3_begin_0, end = old_cache_3_end_0, end_mask = old_cache_3_end_mask_0, squeeze_mask = old_cache_3_squeeze_mask_0, x = enc_conv_cache)[name = string("old_cache_3")]; |
| fp32 var_1122 = const()[name = string("op_1122"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_47 = layer_norm(axes = input_47_axes_0, beta = model_enc_encoder_layers_1_sequential_0_module_sequential_0_bias, epsilon = var_1122, gamma = model_enc_encoder_layers_1_sequential_0_module_sequential_0_weight, x = input_45)[name = string("input_47")]; |
| tensor<fp32, [1, 1, 1024]> inputs_9 = linear(bias = model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_1_sequential_0_module_sequential_1_linear_weight, x = input_47)[name = string("linear_10")]; |
| tensor<fp32, [1, 1, 1024]> input_49 = silu(x = inputs_9)[name = string("input_49")]; |
| tensor<fp32, [1, 1, 256]> input_53 = linear(bias = model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_1_sequential_0_module_sequential_4_linear_weight, x = input_49)[name = string("linear_11")]; |
| fp32 var_1146 = const()[name = string("op_1146"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1147 = mul(x = input_53, y = var_1146)[name = string("op_1147")]; |
| tensor<fp32, [1, 1, 256]> input_55 = add(x = var_1147, y = input_45)[name = string("input_55")]; |
| fp32 var_1153 = const()[name = string("op_1153"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_5_axes_0 = const()[name = string("x_5_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_5 = layer_norm(axes = x_5_axes_0, beta = model_enc_encoder_layers_1_sequential_1_module_layer_norm_bias, epsilon = var_1153, gamma = model_enc_encoder_layers_1_sequential_1_module_layer_norm_weight, x = input_55)[name = string("x_5")]; |
| tensor<fp32, [1, 1, 256]> q_5 = linear(bias = model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_bias, weight = model_enc_encoder_layers_1_sequential_1_module_self_attn_q_proj_weight, x = x_5)[name = string("linear_12")]; |
| tensor<fp32, [1, 1, 256]> k_7 = linear(bias = model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_bias, weight = model_enc_encoder_layers_1_sequential_1_module_self_attn_k_proj_weight, x = x_5)[name = string("linear_13")]; |
| tensor<fp32, [1, 1, 256]> v_5 = linear(bias = model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_bias, weight = model_enc_encoder_layers_1_sequential_1_module_self_attn_v_proj_weight, x = x_5)[name = string("linear_14")]; |
| tensor<fp32, [1, 1, 256]> input_59 = linear(bias = model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_bias, weight = model_enc_encoder_layers_1_sequential_1_module_self_attn_g_proj_weight, x = x_5)[name = string("linear_15")]; |
| fp32 var_1184 = const()[name = string("op_1184"), val = fp32(0x1p-3)]; |
| tensor<fp32, [1, 1, 256]> k_9 = mul(x = k_7, y = var_1184)[name = string("k_9")]; |
| tensor<int32, [4]> var_1188 = const()[name = string("op_1188"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1189 = reshape(shape = var_1188, x = q_5)[name = string("op_1189")]; |
| tensor<int32, [4]> q_7_perm_0 = const()[name = string("q_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1195 = const()[name = string("op_1195"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1196 = reshape(shape = var_1195, x = k_9)[name = string("op_1196")]; |
| tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1203 = const()[name = string("op_1203"), val = tensor<int32, [4]>([1, 4, 64, 1])]; |
| tensor<fp32, [1, 4, 64, 1]> v_7 = reshape(shape = var_1203, x = v_5)[name = string("v_7")]; |
| tensor<fp32, [1, 4, 1, 64]> k_11 = transpose(perm = k_11_perm_0, x = var_1196)[name = string("transpose_44")]; |
| tensor<fp32, [1, 4, 64, 64]> kv_3 = mul(x = k_11, y = v_7)[name = string("kv_3")]; |
| fp32 var_1218 = const()[name = string("op_1218"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 4]> candidate_scale_3 = add(x = old_scale_3, y = var_1218)[name = string("candidate_scale_3")]; |
| tensor<fp32, [1, 4]> var_1220 = sqrt(x = old_scale_3)[name = string("op_1220")]; |
| tensor<fp32, [1, 4]> var_1222 = sqrt(x = candidate_scale_3)[name = string("op_1222")]; |
| tensor<fp32, [1, 4]> var_1223 = real_div(x = var_1220, y = var_1222)[name = string("op_1223")]; |
| tensor<int32, [1]> var_1225_axes_0 = const()[name = string("op_1225_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1225 = expand_dims(axes = var_1225_axes_0, x = var_1223)[name = string("op_1225")]; |
| tensor<int32, [1]> blend_3_axes_0 = const()[name = string("blend_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> blend_3 = expand_dims(axes = blend_3_axes_0, x = var_1225)[name = string("blend_3")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1228 = mul(x = old_kv_3, y = blend_3)[name = string("op_1228")]; |
| tensor<int32, [1]> var_1231_axes_0 = const()[name = string("op_1231_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1231 = expand_dims(axes = var_1231_axes_0, x = var_1222)[name = string("op_1231")]; |
| tensor<int32, [1]> var_1233_axes_0 = const()[name = string("op_1233_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> var_1233 = expand_dims(axes = var_1233_axes_0, x = var_1231)[name = string("op_1233")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1234 = real_div(x = kv_3, y = var_1233)[name = string("op_1234")]; |
| tensor<fp32, [1, 4, 64, 64]> candidate_kv_3 = add(x = var_1228, y = var_1234)[name = string("candidate_kv_3")]; |
| tensor<fp32, [1, 4, 1, 64]> q_7 = transpose(perm = q_7_perm_0, x = var_1189)[name = string("transpose_45")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1237 = mul(x = q_7, y = candidate_kv_3)[name = string("op_1237")]; |
| tensor<int32, [1]> input_57_axes_0 = const()[name = string("input_57_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_57_keep_dims_0 = const()[name = string("input_57_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 64]> input_57 = reduce_sum(axes = input_57_axes_0, keep_dims = input_57_keep_dims_0, x = var_1237)[name = string("input_57")]; |
| fp32 var_1244 = const()[name = string("op_1244"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_1248_axes_0 = const()[name = string("op_1248_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 64]> var_1248 = layer_norm(axes = var_1248_axes_0, epsilon = var_1244, x = input_57)[name = string("op_1248")]; |
| tensor<int32, [3]> var_1250 = const()[name = string("op_1250"), val = tensor<int32, [3]>([1, 1, 256])]; |
| tensor<fp32, [1, 1, 256]> output_3 = reshape(shape = var_1250, x = var_1248)[name = string("output_3")]; |
| tensor<fp32, [1, 1, 256]> var_1252 = silu(x = input_59)[name = string("op_1252")]; |
| tensor<fp32, [1, 1, 256]> input_61 = mul(x = var_1252, y = output_3)[name = string("input_61")]; |
| tensor<fp32, [1, 1, 256]> input_63 = linear(bias = model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_bias, weight = model_enc_encoder_layers_1_sequential_1_module_self_attn_out_proj_weight, x = input_61)[name = string("linear_16")]; |
| tensor<fp32, [1, 1, 256]> input_65 = add(x = input_55, y = input_63)[name = string("input_65")]; |
| fp32 var_1263 = const()[name = string("op_1263"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_7_axes_0 = const()[name = string("x_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = model_enc_encoder_layers_1_sequential_2_module_sequential_0_bias, epsilon = var_1263, gamma = model_enc_encoder_layers_1_sequential_2_module_sequential_0_weight, x = input_65)[name = string("x_7")]; |
| tensor<int32, [3]> input_67_perm_0 = const()[name = string("input_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string inputs_11_pad_type_0 = const()[name = string("inputs_11_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> inputs_11_strides_0 = const()[name = string("inputs_11_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> inputs_11_pad_0 = const()[name = string("inputs_11_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> inputs_11_dilations_0 = const()[name = string("inputs_11_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 inputs_11_groups_0 = const()[name = string("inputs_11_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_67 = transpose(perm = input_67_perm_0, x = x_7)[name = string("transpose_43")]; |
| tensor<fp32, [1, 512, 1]> inputs_11 = conv(bias = model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_bias, dilations = inputs_11_dilations_0, groups = inputs_11_groups_0, pad = inputs_11_pad_0, pad_type = inputs_11_pad_type_0, strides = inputs_11_strides_0, weight = model_enc_encoder_layers_1_sequential_2_module_sequential_2_conv_weight, x = input_67)[name = string("inputs_11")]; |
| tensor<int32, [2]> var_1284_split_sizes_0 = const()[name = string("op_1284_split_sizes_0"), val = tensor<int32, [2]>([256, 256])]; |
| int32 var_1284_axis_0 = const()[name = string("op_1284_axis_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> var_1284_0, tensor<fp32, [1, 256, 1]> var_1284_1 = split(axis = var_1284_axis_0, split_sizes = var_1284_split_sizes_0, x = inputs_11)[name = string("op_1284")]; |
| tensor<fp32, [1, 256, 1]> var_1286 = sigmoid(x = var_1284_1)[name = string("op_1286")]; |
| tensor<fp32, [1, 256, 1]> current_3 = mul(x = var_1284_0, y = var_1286)[name = string("current_3")]; |
| tensor<int32, [3]> cache_3_perm_0 = const()[name = string("cache_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| int32 var_1292 = const()[name = string("op_1292"), val = int32(2)]; |
| bool depthwise_window_3_interleave_0 = const()[name = string("depthwise_window_3_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 256, 15]> cache_3 = transpose(perm = cache_3_perm_0, x = old_cache_3)[name = string("transpose_42")]; |
| tensor<fp32, [1, 256, 16]> depthwise_window_3 = concat(axis = var_1292, interleave = depthwise_window_3_interleave_0, values = (cache_3, current_3))[name = string("depthwise_window_3")]; |
| string input_69_pad_type_0 = const()[name = string("input_69_pad_type_0"), val = string("valid")]; |
| int32 input_69_groups_0 = const()[name = string("input_69_groups_0"), val = int32(256)]; |
| tensor<int32, [1]> input_69_strides_0 = const()[name = string("input_69_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_69_pad_0 = const()[name = string("input_69_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_69_dilations_0 = const()[name = string("input_69_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [256, 1, 16]> const_36 = const()[name = string("const_36"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42750400)))]; |
| tensor<fp32, [256]> const_37 = const()[name = string("const_37"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42766848)))]; |
| tensor<fp32, [1, 256, 1]> inputs_13 = conv(bias = const_37, 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 = const_36, x = depthwise_window_3)[name = string("inputs_13")]; |
| tensor<int32, [3]> var_1324_begin_0 = const()[name = string("op_1324_begin_0"), val = tensor<int32, [3]>([0, 0, 1])]; |
| tensor<int32, [3]> var_1324_end_0 = const()[name = string("op_1324_end_0"), val = tensor<int32, [3]>([1, 256, 16])]; |
| tensor<bool, [3]> var_1324_end_mask_0 = const()[name = string("op_1324_end_mask_0"), val = tensor<bool, [3]>([true, true, true])]; |
| tensor<fp32, [1, 256, 15]> var_1324 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, x = depthwise_window_3)[name = string("op_1324")]; |
| tensor<int32, [3]> candidate_conv_3_perm_0 = const()[name = string("candidate_conv_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 256, 1]> input_71 = silu(x = inputs_13)[name = string("input_71")]; |
| string input_73_pad_type_0 = const()[name = string("input_73_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> input_73_strides_0 = const()[name = string("input_73_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_73_pad_0 = const()[name = string("input_73_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_73_dilations_0 = const()[name = string("input_73_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 input_73_groups_0 = const()[name = string("input_73_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_73 = conv(bias = model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_bias, dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = model_enc_encoder_layers_1_sequential_2_module_sequential_7_conv_weight, x = input_71)[name = string("input_73")]; |
| tensor<int32, [3]> conv_output_3_perm_0 = const()[name = string("conv_output_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 1, 256]> conv_output_3 = transpose(perm = conv_output_3_perm_0, x = input_73)[name = string("transpose_40")]; |
| tensor<fp32, [1, 1, 256]> input_75 = add(x = input_65, y = conv_output_3)[name = string("input_75")]; |
| fp32 var_1360 = const()[name = string("op_1360"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_77 = layer_norm(axes = input_77_axes_0, beta = model_enc_encoder_layers_1_sequential_3_module_sequential_0_bias, epsilon = var_1360, gamma = model_enc_encoder_layers_1_sequential_3_module_sequential_0_weight, x = input_75)[name = string("input_77")]; |
| tensor<fp32, [1, 1, 1024]> inputs_15 = linear(bias = model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_1_sequential_3_module_sequential_1_linear_weight, x = input_77)[name = string("linear_17")]; |
| tensor<fp32, [1, 1, 1024]> input_79 = silu(x = inputs_15)[name = string("input_79")]; |
| tensor<fp32, [1, 1, 256]> input_83 = linear(bias = model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_1_sequential_3_module_sequential_4_linear_weight, x = input_79)[name = string("linear_18")]; |
| fp32 var_1384 = const()[name = string("op_1384"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1385 = mul(x = input_83, y = var_1384)[name = string("op_1385")]; |
| tensor<fp32, [1, 1, 256]> input_85 = add(x = var_1385, y = input_75)[name = string("input_85")]; |
| fp32 var_1391 = const()[name = string("op_1391"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_87 = layer_norm(axes = input_87_axes_0, beta = model_enc_encoder_layers_1_sequential_4_bias, epsilon = var_1391, gamma = model_enc_encoder_layers_1_sequential_4_weight, x = input_85)[name = string("input_87")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1398 = sub(x = candidate_kv_3, y = old_kv_3)[name = string("op_1398")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1401 = mul(x = var_1398, y = var_1098)[name = string("op_1401")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1403 = add(x = old_kv_3, y = var_1401)[name = string("op_1403")]; |
| tensor<fp32, [1, 4]> var_1405 = sub(x = candidate_scale_3, y = old_scale_3)[name = string("op_1405")]; |
| tensor<fp32, [1, 4]> var_1406 = mul(x = var_1405, y = ingest_vec)[name = string("op_1406")]; |
| tensor<fp32, [1, 4]> var_1408 = add(x = old_scale_3, y = var_1406)[name = string("op_1408")]; |
| tensor<fp32, [1, 15, 256]> candidate_conv_3 = transpose(perm = candidate_conv_3_perm_0, x = var_1324)[name = string("transpose_41")]; |
| tensor<fp32, [1, 15, 256]> var_1410 = sub(x = candidate_conv_3, y = old_cache_3)[name = string("op_1410")]; |
| tensor<fp32, [1, 15, 256]> var_1411 = mul(x = var_1410, y = ingest_scalar)[name = string("op_1411")]; |
| tensor<fp32, [1, 15, 256]> var_1413 = add(x = old_cache_3, y = var_1411)[name = string("op_1413")]; |
| tensor<int32, [5]> old_kv_5_begin_0 = const()[name = string("old_kv_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_5_end_0 = const()[name = string("old_kv_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_5_end_mask_0 = const()[name = string("old_kv_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_5_squeeze_mask_0 = const()[name = string("old_kv_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, 64, 64]> old_kv_5 = slice_by_index(begin = old_kv_5_begin_0, end = old_kv_5_end_0, end_mask = old_kv_5_end_mask_0, squeeze_mask = old_kv_5_squeeze_mask_0, x = enc_ret_kv)[name = string("old_kv_5")]; |
| tensor<int32, [3]> old_scale_5_begin_0 = const()[name = string("old_scale_5_begin_0"), val = tensor<int32, [3]>([2, 0, 0])]; |
| tensor<int32, [3]> old_scale_5_end_0 = const()[name = string("old_scale_5_end_0"), val = tensor<int32, [3]>([3, 1, 4])]; |
| tensor<bool, [3]> old_scale_5_end_mask_0 = const()[name = string("old_scale_5_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_5_squeeze_mask_0 = const()[name = string("old_scale_5_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [1, 4]> old_scale_5 = slice_by_index(begin = old_scale_5_begin_0, end = old_scale_5_end_0, end_mask = old_scale_5_end_mask_0, squeeze_mask = old_scale_5_squeeze_mask_0, x = enc_ret_scale)[name = string("old_scale_5")]; |
| tensor<int32, [4]> old_cache_5_begin_0 = const()[name = string("old_cache_5_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])]; |
| tensor<int32, [4]> old_cache_5_end_0 = const()[name = string("old_cache_5_end_0"), val = tensor<int32, [4]>([3, 1, 15, 256])]; |
| tensor<bool, [4]> old_cache_5_end_mask_0 = const()[name = string("old_cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; |
| tensor<bool, [4]> old_cache_5_squeeze_mask_0 = const()[name = string("old_cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; |
| tensor<fp32, [1, 15, 256]> old_cache_5 = slice_by_index(begin = old_cache_5_begin_0, end = old_cache_5_end_0, end_mask = old_cache_5_end_mask_0, squeeze_mask = old_cache_5_squeeze_mask_0, x = enc_conv_cache)[name = string("old_cache_5")]; |
| fp32 var_1424 = const()[name = string("op_1424"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_89_axes_0 = const()[name = string("input_89_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_89 = layer_norm(axes = input_89_axes_0, beta = model_enc_encoder_layers_2_sequential_0_module_sequential_0_bias, epsilon = var_1424, gamma = model_enc_encoder_layers_2_sequential_0_module_sequential_0_weight, x = input_87)[name = string("input_89")]; |
| tensor<fp32, [1, 1, 1024]> inputs_17 = linear(bias = model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_2_sequential_0_module_sequential_1_linear_weight, x = input_89)[name = string("linear_19")]; |
| tensor<fp32, [1, 1, 1024]> input_91 = silu(x = inputs_17)[name = string("input_91")]; |
| tensor<fp32, [1, 1, 256]> input_95 = linear(bias = model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_2_sequential_0_module_sequential_4_linear_weight, x = input_91)[name = string("linear_20")]; |
| fp32 var_1448 = const()[name = string("op_1448"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1449 = mul(x = input_95, y = var_1448)[name = string("op_1449")]; |
| tensor<fp32, [1, 1, 256]> input_97 = add(x = var_1449, y = input_87)[name = string("input_97")]; |
| fp32 var_1455 = const()[name = string("op_1455"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_9_axes_0 = const()[name = string("x_9_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_9 = layer_norm(axes = x_9_axes_0, beta = model_enc_encoder_layers_2_sequential_1_module_layer_norm_bias, epsilon = var_1455, gamma = model_enc_encoder_layers_2_sequential_1_module_layer_norm_weight, x = input_97)[name = string("x_9")]; |
| tensor<fp32, [1, 1, 256]> q_9 = linear(bias = model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_bias, weight = model_enc_encoder_layers_2_sequential_1_module_self_attn_q_proj_weight, x = x_9)[name = string("linear_21")]; |
| tensor<fp32, [1, 1, 256]> k_13 = linear(bias = model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_bias, weight = model_enc_encoder_layers_2_sequential_1_module_self_attn_k_proj_weight, x = x_9)[name = string("linear_22")]; |
| tensor<fp32, [1, 1, 256]> v_9 = linear(bias = model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_bias, weight = model_enc_encoder_layers_2_sequential_1_module_self_attn_v_proj_weight, x = x_9)[name = string("linear_23")]; |
| tensor<fp32, [1, 1, 256]> input_101 = linear(bias = model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_bias, weight = model_enc_encoder_layers_2_sequential_1_module_self_attn_g_proj_weight, x = x_9)[name = string("linear_24")]; |
| fp32 var_1486 = const()[name = string("op_1486"), val = fp32(0x1p-3)]; |
| tensor<fp32, [1, 1, 256]> k_15 = mul(x = k_13, y = var_1486)[name = string("k_15")]; |
| tensor<int32, [4]> var_1490 = const()[name = string("op_1490"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1491 = reshape(shape = var_1490, x = q_9)[name = string("op_1491")]; |
| tensor<int32, [4]> q_11_perm_0 = const()[name = string("q_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1497 = const()[name = string("op_1497"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1498 = reshape(shape = var_1497, x = k_15)[name = string("op_1498")]; |
| tensor<int32, [4]> k_17_perm_0 = const()[name = string("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1505 = const()[name = string("op_1505"), val = tensor<int32, [4]>([1, 4, 64, 1])]; |
| tensor<fp32, [1, 4, 64, 1]> v_11 = reshape(shape = var_1505, x = v_9)[name = string("v_11")]; |
| tensor<fp32, [1, 4, 1, 64]> k_17 = transpose(perm = k_17_perm_0, x = var_1498)[name = string("transpose_38")]; |
| tensor<fp32, [1, 4, 64, 64]> kv_5 = mul(x = k_17, y = v_11)[name = string("kv_5")]; |
| fp32 var_1520 = const()[name = string("op_1520"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 4]> candidate_scale_5 = add(x = old_scale_5, y = var_1520)[name = string("candidate_scale_5")]; |
| tensor<fp32, [1, 4]> var_1522 = sqrt(x = old_scale_5)[name = string("op_1522")]; |
| tensor<fp32, [1, 4]> var_1524 = sqrt(x = candidate_scale_5)[name = string("op_1524")]; |
| tensor<fp32, [1, 4]> var_1525 = real_div(x = var_1522, y = var_1524)[name = string("op_1525")]; |
| tensor<int32, [1]> var_1527_axes_0 = const()[name = string("op_1527_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1527 = expand_dims(axes = var_1527_axes_0, x = var_1525)[name = string("op_1527")]; |
| tensor<int32, [1]> blend_5_axes_0 = const()[name = string("blend_5_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> blend_5 = expand_dims(axes = blend_5_axes_0, x = var_1527)[name = string("blend_5")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1530 = mul(x = old_kv_5, y = blend_5)[name = string("op_1530")]; |
| tensor<int32, [1]> var_1533_axes_0 = const()[name = string("op_1533_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1533 = expand_dims(axes = var_1533_axes_0, x = var_1524)[name = string("op_1533")]; |
| tensor<int32, [1]> var_1535_axes_0 = const()[name = string("op_1535_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> var_1535 = expand_dims(axes = var_1535_axes_0, x = var_1533)[name = string("op_1535")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1536 = real_div(x = kv_5, y = var_1535)[name = string("op_1536")]; |
| tensor<fp32, [1, 4, 64, 64]> candidate_kv_5 = add(x = var_1530, y = var_1536)[name = string("candidate_kv_5")]; |
| tensor<fp32, [1, 4, 1, 64]> q_11 = transpose(perm = q_11_perm_0, x = var_1491)[name = string("transpose_39")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1539 = mul(x = q_11, y = candidate_kv_5)[name = string("op_1539")]; |
| tensor<int32, [1]> input_99_axes_0 = const()[name = string("input_99_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_99_keep_dims_0 = const()[name = string("input_99_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 64]> input_99 = reduce_sum(axes = input_99_axes_0, keep_dims = input_99_keep_dims_0, x = var_1539)[name = string("input_99")]; |
| fp32 var_1546 = const()[name = string("op_1546"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_1550_axes_0 = const()[name = string("op_1550_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 64]> var_1550 = layer_norm(axes = var_1550_axes_0, epsilon = var_1546, x = input_99)[name = string("op_1550")]; |
| tensor<int32, [3]> var_1552 = const()[name = string("op_1552"), val = tensor<int32, [3]>([1, 1, 256])]; |
| tensor<fp32, [1, 1, 256]> output_5 = reshape(shape = var_1552, x = var_1550)[name = string("output_5")]; |
| tensor<fp32, [1, 1, 256]> var_1554 = silu(x = input_101)[name = string("op_1554")]; |
| tensor<fp32, [1, 1, 256]> input_103 = mul(x = var_1554, y = output_5)[name = string("input_103")]; |
| tensor<fp32, [1, 1, 256]> input_105 = linear(bias = model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_bias, weight = model_enc_encoder_layers_2_sequential_1_module_self_attn_out_proj_weight, x = input_103)[name = string("linear_25")]; |
| tensor<fp32, [1, 1, 256]> input_107 = add(x = input_97, y = input_105)[name = string("input_107")]; |
| fp32 var_1565 = const()[name = string("op_1565"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_11_axes_0 = const()[name = string("x_11_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_11 = layer_norm(axes = x_11_axes_0, beta = model_enc_encoder_layers_2_sequential_2_module_sequential_0_bias, epsilon = var_1565, gamma = model_enc_encoder_layers_2_sequential_2_module_sequential_0_weight, x = input_107)[name = string("x_11")]; |
| tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string inputs_19_pad_type_0 = const()[name = string("inputs_19_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> inputs_19_strides_0 = const()[name = string("inputs_19_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> inputs_19_pad_0 = const()[name = string("inputs_19_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> inputs_19_dilations_0 = const()[name = string("inputs_19_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 inputs_19_groups_0 = const()[name = string("inputs_19_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_109 = transpose(perm = input_109_perm_0, x = x_11)[name = string("transpose_37")]; |
| tensor<fp32, [1, 512, 1]> inputs_19 = conv(bias = model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_bias, dilations = inputs_19_dilations_0, groups = inputs_19_groups_0, pad = inputs_19_pad_0, pad_type = inputs_19_pad_type_0, strides = inputs_19_strides_0, weight = model_enc_encoder_layers_2_sequential_2_module_sequential_2_conv_weight, x = input_109)[name = string("inputs_19")]; |
| tensor<int32, [2]> var_1586_split_sizes_0 = const()[name = string("op_1586_split_sizes_0"), val = tensor<int32, [2]>([256, 256])]; |
| int32 var_1586_axis_0 = const()[name = string("op_1586_axis_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> var_1586_0, tensor<fp32, [1, 256, 1]> var_1586_1 = split(axis = var_1586_axis_0, split_sizes = var_1586_split_sizes_0, x = inputs_19)[name = string("op_1586")]; |
| tensor<fp32, [1, 256, 1]> var_1588 = sigmoid(x = var_1586_1)[name = string("op_1588")]; |
| tensor<fp32, [1, 256, 1]> current_5 = mul(x = var_1586_0, y = var_1588)[name = string("current_5")]; |
| tensor<int32, [3]> cache_5_perm_0 = const()[name = string("cache_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| int32 var_1594 = const()[name = string("op_1594"), val = int32(2)]; |
| bool depthwise_window_5_interleave_0 = const()[name = string("depthwise_window_5_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 256, 15]> cache_5 = transpose(perm = cache_5_perm_0, x = old_cache_5)[name = string("transpose_36")]; |
| tensor<fp32, [1, 256, 16]> depthwise_window_5 = concat(axis = var_1594, interleave = depthwise_window_5_interleave_0, values = (cache_5, current_5))[name = string("depthwise_window_5")]; |
| string input_111_pad_type_0 = const()[name = string("input_111_pad_type_0"), val = string("valid")]; |
| int32 input_111_groups_0 = const()[name = string("input_111_groups_0"), val = int32(256)]; |
| tensor<int32, [1]> input_111_strides_0 = const()[name = string("input_111_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_111_pad_0 = const()[name = string("input_111_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_111_dilations_0 = const()[name = string("input_111_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [256, 1, 16]> const_38 = const()[name = string("const_38"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42767936)))]; |
| tensor<fp32, [256]> const_39 = const()[name = string("const_39"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42784384)))]; |
| tensor<fp32, [1, 256, 1]> inputs_21 = conv(bias = const_39, 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 = const_38, x = depthwise_window_5)[name = string("inputs_21")]; |
| tensor<int32, [3]> var_1626_begin_0 = const()[name = string("op_1626_begin_0"), val = tensor<int32, [3]>([0, 0, 1])]; |
| tensor<int32, [3]> var_1626_end_0 = const()[name = string("op_1626_end_0"), val = tensor<int32, [3]>([1, 256, 16])]; |
| tensor<bool, [3]> var_1626_end_mask_0 = const()[name = string("op_1626_end_mask_0"), val = tensor<bool, [3]>([true, true, true])]; |
| tensor<fp32, [1, 256, 15]> var_1626 = slice_by_index(begin = var_1626_begin_0, end = var_1626_end_0, end_mask = var_1626_end_mask_0, x = depthwise_window_5)[name = string("op_1626")]; |
| tensor<int32, [3]> candidate_conv_5_perm_0 = const()[name = string("candidate_conv_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 256, 1]> input_113 = silu(x = inputs_21)[name = string("input_113")]; |
| string input_115_pad_type_0 = const()[name = string("input_115_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> input_115_strides_0 = const()[name = string("input_115_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_115_pad_0 = const()[name = string("input_115_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_115_dilations_0 = const()[name = string("input_115_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 input_115_groups_0 = const()[name = string("input_115_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_115 = conv(bias = model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_bias, dilations = input_115_dilations_0, groups = input_115_groups_0, pad = input_115_pad_0, pad_type = input_115_pad_type_0, strides = input_115_strides_0, weight = model_enc_encoder_layers_2_sequential_2_module_sequential_7_conv_weight, x = input_113)[name = string("input_115")]; |
| tensor<int32, [3]> conv_output_5_perm_0 = const()[name = string("conv_output_5_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 1, 256]> conv_output_5 = transpose(perm = conv_output_5_perm_0, x = input_115)[name = string("transpose_34")]; |
| tensor<fp32, [1, 1, 256]> input_117 = add(x = input_107, y = conv_output_5)[name = string("input_117")]; |
| fp32 var_1662 = const()[name = string("op_1662"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_119_axes_0 = const()[name = string("input_119_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_119 = layer_norm(axes = input_119_axes_0, beta = model_enc_encoder_layers_2_sequential_3_module_sequential_0_bias, epsilon = var_1662, gamma = model_enc_encoder_layers_2_sequential_3_module_sequential_0_weight, x = input_117)[name = string("input_119")]; |
| tensor<fp32, [1, 1, 1024]> inputs_23 = linear(bias = model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_2_sequential_3_module_sequential_1_linear_weight, x = input_119)[name = string("linear_26")]; |
| tensor<fp32, [1, 1, 1024]> input_121 = silu(x = inputs_23)[name = string("input_121")]; |
| tensor<fp32, [1, 1, 256]> input_125 = linear(bias = model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_2_sequential_3_module_sequential_4_linear_weight, x = input_121)[name = string("linear_27")]; |
| fp32 var_1686 = const()[name = string("op_1686"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1687 = mul(x = input_125, y = var_1686)[name = string("op_1687")]; |
| tensor<fp32, [1, 1, 256]> input_127 = add(x = var_1687, y = input_117)[name = string("input_127")]; |
| fp32 var_1693 = const()[name = string("op_1693"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_129 = layer_norm(axes = input_129_axes_0, beta = model_enc_encoder_layers_2_sequential_4_bias, epsilon = var_1693, gamma = model_enc_encoder_layers_2_sequential_4_weight, x = input_127)[name = string("input_129")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1700 = sub(x = candidate_kv_5, y = old_kv_5)[name = string("op_1700")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1703 = mul(x = var_1700, y = var_1098)[name = string("op_1703")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1705 = add(x = old_kv_5, y = var_1703)[name = string("op_1705")]; |
| tensor<fp32, [1, 4]> var_1707 = sub(x = candidate_scale_5, y = old_scale_5)[name = string("op_1707")]; |
| tensor<fp32, [1, 4]> var_1708 = mul(x = var_1707, y = ingest_vec)[name = string("op_1708")]; |
| tensor<fp32, [1, 4]> var_1710 = add(x = old_scale_5, y = var_1708)[name = string("op_1710")]; |
| tensor<fp32, [1, 15, 256]> candidate_conv_5 = transpose(perm = candidate_conv_5_perm_0, x = var_1626)[name = string("transpose_35")]; |
| tensor<fp32, [1, 15, 256]> var_1712 = sub(x = candidate_conv_5, y = old_cache_5)[name = string("op_1712")]; |
| tensor<fp32, [1, 15, 256]> var_1713 = mul(x = var_1712, y = ingest_scalar)[name = string("op_1713")]; |
| tensor<fp32, [1, 15, 256]> var_1715 = add(x = old_cache_5, y = var_1713)[name = string("op_1715")]; |
| tensor<int32, [5]> old_kv_7_begin_0 = const()[name = string("old_kv_7_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_7_end_0 = const()[name = string("old_kv_7_end_0"), val = tensor<int32, [5]>([4, 1, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_7_end_mask_0 = const()[name = string("old_kv_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_7_squeeze_mask_0 = const()[name = string("old_kv_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, 64, 64]> old_kv_7 = slice_by_index(begin = old_kv_7_begin_0, end = old_kv_7_end_0, end_mask = old_kv_7_end_mask_0, squeeze_mask = old_kv_7_squeeze_mask_0, x = enc_ret_kv)[name = string("old_kv_7")]; |
| tensor<int32, [3]> old_scale_7_begin_0 = const()[name = string("old_scale_7_begin_0"), val = tensor<int32, [3]>([3, 0, 0])]; |
| tensor<int32, [3]> old_scale_7_end_0 = const()[name = string("old_scale_7_end_0"), val = tensor<int32, [3]>([4, 1, 4])]; |
| tensor<bool, [3]> old_scale_7_end_mask_0 = const()[name = string("old_scale_7_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_7_squeeze_mask_0 = const()[name = string("old_scale_7_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [1, 4]> old_scale_7 = slice_by_index(begin = old_scale_7_begin_0, end = old_scale_7_end_0, end_mask = old_scale_7_end_mask_0, squeeze_mask = old_scale_7_squeeze_mask_0, x = enc_ret_scale)[name = string("old_scale_7")]; |
| tensor<int32, [4]> old_cache_begin_0 = const()[name = string("old_cache_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])]; |
| tensor<int32, [4]> old_cache_end_0 = const()[name = string("old_cache_end_0"), val = tensor<int32, [4]>([4, 1, 15, 256])]; |
| tensor<bool, [4]> old_cache_end_mask_0 = const()[name = string("old_cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])]; |
| tensor<bool, [4]> old_cache_squeeze_mask_0 = const()[name = string("old_cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])]; |
| tensor<fp32, [1, 15, 256]> old_cache = slice_by_index(begin = old_cache_begin_0, end = old_cache_end_0, end_mask = old_cache_end_mask_0, squeeze_mask = old_cache_squeeze_mask_0, x = enc_conv_cache)[name = string("old_cache")]; |
| fp32 var_1726 = const()[name = string("op_1726"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_131 = layer_norm(axes = input_131_axes_0, beta = model_enc_encoder_layers_3_sequential_0_module_sequential_0_bias, epsilon = var_1726, gamma = model_enc_encoder_layers_3_sequential_0_module_sequential_0_weight, x = input_129)[name = string("input_131")]; |
| tensor<fp32, [1, 1, 1024]> inputs_25 = linear(bias = model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_3_sequential_0_module_sequential_1_linear_weight, x = input_131)[name = string("linear_28")]; |
| tensor<fp32, [1, 1, 1024]> input_133 = silu(x = inputs_25)[name = string("input_133")]; |
| tensor<fp32, [1, 1, 256]> input_137 = linear(bias = model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_3_sequential_0_module_sequential_4_linear_weight, x = input_133)[name = string("linear_29")]; |
| fp32 var_1750 = const()[name = string("op_1750"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1751 = mul(x = input_137, y = var_1750)[name = string("op_1751")]; |
| tensor<fp32, [1, 1, 256]> input_139 = add(x = var_1751, y = input_129)[name = string("input_139")]; |
| fp32 var_1757 = const()[name = string("op_1757"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_13_axes_0 = const()[name = string("x_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = model_enc_encoder_layers_3_sequential_1_module_layer_norm_bias, epsilon = var_1757, gamma = model_enc_encoder_layers_3_sequential_1_module_layer_norm_weight, x = input_139)[name = string("x_13")]; |
| tensor<fp32, [1, 1, 256]> q_13 = linear(bias = model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_bias, weight = model_enc_encoder_layers_3_sequential_1_module_self_attn_q_proj_weight, x = x_13)[name = string("linear_30")]; |
| tensor<fp32, [1, 1, 256]> k_19 = linear(bias = model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_bias, weight = model_enc_encoder_layers_3_sequential_1_module_self_attn_k_proj_weight, x = x_13)[name = string("linear_31")]; |
| tensor<fp32, [1, 1, 256]> v_13 = linear(bias = model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_bias, weight = model_enc_encoder_layers_3_sequential_1_module_self_attn_v_proj_weight, x = x_13)[name = string("linear_32")]; |
| tensor<fp32, [1, 1, 256]> input_143 = linear(bias = model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_bias, weight = model_enc_encoder_layers_3_sequential_1_module_self_attn_g_proj_weight, x = x_13)[name = string("linear_33")]; |
| fp32 var_1788 = const()[name = string("op_1788"), val = fp32(0x1p-3)]; |
| tensor<fp32, [1, 1, 256]> k_21 = mul(x = k_19, y = var_1788)[name = string("k_21")]; |
| tensor<int32, [4]> var_1792 = const()[name = string("op_1792"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1793 = reshape(shape = var_1792, x = q_13)[name = string("op_1793")]; |
| tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1799 = const()[name = string("op_1799"), val = tensor<int32, [4]>([1, 1, 4, 64])]; |
| tensor<fp32, [1, 1, 4, 64]> var_1800 = reshape(shape = var_1799, x = k_21)[name = string("op_1800")]; |
| tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_1807 = const()[name = string("op_1807"), val = tensor<int32, [4]>([1, 4, 64, 1])]; |
| tensor<fp32, [1, 4, 64, 1]> v_15 = reshape(shape = var_1807, x = v_13)[name = string("v_15")]; |
| tensor<fp32, [1, 4, 1, 64]> k_23 = transpose(perm = k_23_perm_0, x = var_1800)[name = string("transpose_32")]; |
| tensor<fp32, [1, 4, 64, 64]> kv_7 = mul(x = k_23, y = v_15)[name = string("kv_7")]; |
| fp32 var_1822 = const()[name = string("op_1822"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 4]> candidate_scale_7 = add(x = old_scale_7, y = var_1822)[name = string("candidate_scale_7")]; |
| tensor<fp32, [1, 4]> var_1824 = sqrt(x = old_scale_7)[name = string("op_1824")]; |
| tensor<fp32, [1, 4]> var_1826 = sqrt(x = candidate_scale_7)[name = string("op_1826")]; |
| tensor<fp32, [1, 4]> var_1827 = real_div(x = var_1824, y = var_1826)[name = string("op_1827")]; |
| tensor<int32, [1]> var_1829_axes_0 = const()[name = string("op_1829_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1829 = expand_dims(axes = var_1829_axes_0, x = var_1827)[name = string("op_1829")]; |
| tensor<int32, [1]> blend_7_axes_0 = const()[name = string("blend_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> blend_7 = expand_dims(axes = blend_7_axes_0, x = var_1829)[name = string("blend_7")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1832 = mul(x = old_kv_7, y = blend_7)[name = string("op_1832")]; |
| tensor<int32, [1]> var_1835_axes_0 = const()[name = string("op_1835_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1]> var_1835 = expand_dims(axes = var_1835_axes_0, x = var_1826)[name = string("op_1835")]; |
| tensor<int32, [1]> var_1837_axes_0 = const()[name = string("op_1837_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 1, 1]> var_1837 = expand_dims(axes = var_1837_axes_0, x = var_1835)[name = string("op_1837")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1838 = real_div(x = kv_7, y = var_1837)[name = string("op_1838")]; |
| tensor<fp32, [1, 4, 64, 64]> candidate_kv_7 = add(x = var_1832, y = var_1838)[name = string("candidate_kv_7")]; |
| tensor<fp32, [1, 4, 1, 64]> q_15 = transpose(perm = q_15_perm_0, x = var_1793)[name = string("transpose_33")]; |
| tensor<fp32, [1, 4, 64, 64]> var_1841 = mul(x = q_15, y = candidate_kv_7)[name = string("op_1841")]; |
| tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_141_keep_dims_0 = const()[name = string("input_141_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 64]> input_141 = reduce_sum(axes = input_141_axes_0, keep_dims = input_141_keep_dims_0, x = var_1841)[name = string("input_141")]; |
| fp32 var_1848 = const()[name = string("op_1848"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_1852_axes_0 = const()[name = string("op_1852_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 4, 64]> var_1852 = layer_norm(axes = var_1852_axes_0, epsilon = var_1848, x = input_141)[name = string("op_1852")]; |
| tensor<int32, [3]> var_1854 = const()[name = string("op_1854"), val = tensor<int32, [3]>([1, 1, 256])]; |
| tensor<fp32, [1, 1, 256]> output_7 = reshape(shape = var_1854, x = var_1852)[name = string("output_7")]; |
| tensor<fp32, [1, 1, 256]> var_1856 = silu(x = input_143)[name = string("op_1856")]; |
| tensor<fp32, [1, 1, 256]> input_145 = mul(x = var_1856, y = output_7)[name = string("input_145")]; |
| tensor<fp32, [1, 1, 256]> input_147 = linear(bias = model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_bias, weight = model_enc_encoder_layers_3_sequential_1_module_self_attn_out_proj_weight, x = input_145)[name = string("linear_34")]; |
| tensor<fp32, [1, 1, 256]> input_149 = add(x = input_139, y = input_147)[name = string("input_149")]; |
| fp32 var_1867 = const()[name = string("op_1867"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_15_axes_0 = const()[name = string("x_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_15 = layer_norm(axes = x_15_axes_0, beta = model_enc_encoder_layers_3_sequential_2_module_sequential_0_bias, epsilon = var_1867, gamma = model_enc_encoder_layers_3_sequential_2_module_sequential_0_weight, x = input_149)[name = string("x_15")]; |
| tensor<int32, [3]> input_151_perm_0 = const()[name = string("input_151_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string inputs_27_pad_type_0 = const()[name = string("inputs_27_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> inputs_27_strides_0 = const()[name = string("inputs_27_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> inputs_27_pad_0 = const()[name = string("inputs_27_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> inputs_27_dilations_0 = const()[name = string("inputs_27_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 inputs_27_groups_0 = const()[name = string("inputs_27_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_151 = transpose(perm = input_151_perm_0, x = x_15)[name = string("transpose_31")]; |
| tensor<fp32, [1, 512, 1]> inputs_27 = conv(bias = model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_bias, dilations = inputs_27_dilations_0, groups = inputs_27_groups_0, pad = inputs_27_pad_0, pad_type = inputs_27_pad_type_0, strides = inputs_27_strides_0, weight = model_enc_encoder_layers_3_sequential_2_module_sequential_2_conv_weight, x = input_151)[name = string("inputs_27")]; |
| tensor<int32, [2]> var_1888_split_sizes_0 = const()[name = string("op_1888_split_sizes_0"), val = tensor<int32, [2]>([256, 256])]; |
| int32 var_1888_axis_0 = const()[name = string("op_1888_axis_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> var_1888_0, tensor<fp32, [1, 256, 1]> var_1888_1 = split(axis = var_1888_axis_0, split_sizes = var_1888_split_sizes_0, x = inputs_27)[name = string("op_1888")]; |
| tensor<fp32, [1, 256, 1]> var_1890 = sigmoid(x = var_1888_1)[name = string("op_1890")]; |
| tensor<fp32, [1, 256, 1]> current = mul(x = var_1888_0, y = var_1890)[name = string("current")]; |
| tensor<int32, [3]> cache_perm_0 = const()[name = string("cache_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| int32 var_1896 = const()[name = string("op_1896"), val = int32(2)]; |
| bool depthwise_window_interleave_0 = const()[name = string("depthwise_window_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 256, 15]> cache = transpose(perm = cache_perm_0, x = old_cache)[name = string("transpose_30")]; |
| tensor<fp32, [1, 256, 16]> depthwise_window = concat(axis = var_1896, interleave = depthwise_window_interleave_0, values = (cache, current))[name = string("depthwise_window")]; |
| string input_153_pad_type_0 = const()[name = string("input_153_pad_type_0"), val = string("valid")]; |
| int32 input_153_groups_0 = const()[name = string("input_153_groups_0"), val = int32(256)]; |
| tensor<int32, [1]> input_153_strides_0 = const()[name = string("input_153_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_153_pad_0 = const()[name = string("input_153_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_153_dilations_0 = const()[name = string("input_153_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [256, 1, 16]> const_40 = const()[name = string("const_40"), val = tensor<fp32, [256, 1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42785472)))]; |
| tensor<fp32, [256]> const_41 = const()[name = string("const_41"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42801920)))]; |
| tensor<fp32, [1, 256, 1]> inputs_29 = conv(bias = const_41, dilations = input_153_dilations_0, groups = input_153_groups_0, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = input_153_strides_0, weight = const_40, x = depthwise_window)[name = string("inputs_29")]; |
| tensor<int32, [3]> var_1928_begin_0 = const()[name = string("op_1928_begin_0"), val = tensor<int32, [3]>([0, 0, 1])]; |
| tensor<int32, [3]> var_1928_end_0 = const()[name = string("op_1928_end_0"), val = tensor<int32, [3]>([1, 256, 16])]; |
| tensor<bool, [3]> var_1928_end_mask_0 = const()[name = string("op_1928_end_mask_0"), val = tensor<bool, [3]>([true, true, true])]; |
| tensor<fp32, [1, 256, 15]> var_1928 = slice_by_index(begin = var_1928_begin_0, end = var_1928_end_0, end_mask = var_1928_end_mask_0, x = depthwise_window)[name = string("op_1928")]; |
| tensor<int32, [3]> candidate_conv_perm_0 = const()[name = string("candidate_conv_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 256, 1]> input_155 = silu(x = inputs_29)[name = string("input_155")]; |
| string input_157_pad_type_0 = const()[name = string("input_157_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> input_157_strides_0 = const()[name = string("input_157_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_157_pad_0 = const()[name = string("input_157_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_157_dilations_0 = const()[name = string("input_157_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 input_157_groups_0 = const()[name = string("input_157_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 1]> input_157 = conv(bias = model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_bias, dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = model_enc_encoder_layers_3_sequential_2_module_sequential_7_conv_weight, x = input_155)[name = string("input_157")]; |
| tensor<int32, [3]> conv_output_perm_0 = const()[name = string("conv_output_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, 1, 256]> conv_output = transpose(perm = conv_output_perm_0, x = input_157)[name = string("transpose_28")]; |
| tensor<fp32, [1, 1, 256]> input_159 = add(x = input_149, y = conv_output)[name = string("input_159")]; |
| fp32 var_1964 = const()[name = string("op_1964"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> input_161 = layer_norm(axes = input_161_axes_0, beta = model_enc_encoder_layers_3_sequential_3_module_sequential_0_bias, epsilon = var_1964, gamma = model_enc_encoder_layers_3_sequential_3_module_sequential_0_weight, x = input_159)[name = string("input_161")]; |
| tensor<fp32, [1, 1, 1024]> inputs = linear(bias = model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_bias, weight = model_enc_encoder_layers_3_sequential_3_module_sequential_1_linear_weight, x = input_161)[name = string("linear_35")]; |
| tensor<fp32, [1, 1, 1024]> input_163 = silu(x = inputs)[name = string("input_163")]; |
| tensor<fp32, [1, 1, 256]> input_167 = linear(bias = model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_bias, weight = model_enc_encoder_layers_3_sequential_3_module_sequential_4_linear_weight, x = input_163)[name = string("linear_36")]; |
| fp32 var_1988 = const()[name = string("op_1988"), val = fp32(0x1p-1)]; |
| tensor<fp32, [1, 1, 256]> var_1989 = mul(x = input_167, y = var_1988)[name = string("op_1989")]; |
| tensor<fp32, [1, 1, 256]> input_169 = add(x = var_1989, y = input_159)[name = string("input_169")]; |
| fp32 var_1995 = const()[name = string("op_1995"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 256]> x_17 = layer_norm(axes = x_17_axes_0, beta = model_enc_encoder_layers_3_sequential_4_bias, epsilon = var_1995, gamma = model_enc_encoder_layers_3_sequential_4_weight, x = input_169)[name = string("x_17")]; |
| tensor<fp32, [1, 4, 64, 64]> var_2002 = sub(x = candidate_kv_7, y = old_kv_7)[name = string("op_2002")]; |
| tensor<fp32, [1, 4, 64, 64]> var_2005 = mul(x = var_2002, y = var_1098)[name = string("op_2005")]; |
| tensor<fp32, [1, 4, 64, 64]> blended_kv = add(x = old_kv_7, y = var_2005)[name = string("blended_kv")]; |
| tensor<fp32, [1, 4]> var_2009 = sub(x = candidate_scale_7, y = old_scale_7)[name = string("op_2009")]; |
| tensor<fp32, [1, 4]> var_2010 = mul(x = var_2009, y = ingest_vec)[name = string("op_2010")]; |
| tensor<fp32, [1, 4]> blended_scale = add(x = old_scale_7, y = var_2010)[name = string("blended_scale")]; |
| tensor<fp32, [1, 15, 256]> candidate_conv = transpose(perm = candidate_conv_perm_0, x = var_1928)[name = string("transpose_29")]; |
| tensor<fp32, [1, 15, 256]> var_2014 = sub(x = candidate_conv, y = old_cache)[name = string("op_2014")]; |
| tensor<fp32, [1, 15, 256]> var_2015 = mul(x = var_2014, y = ingest_scalar)[name = string("op_2015")]; |
| tensor<fp32, [1, 15, 256]> blended_conv = add(x = old_cache, y = var_2015)[name = string("blended_conv")]; |
| tensor<fp32, [1, 1, 256]> appended_encoder_frame = mul(x = x_17, y = ingest_scalar)[name = string("appended_encoder_frame")]; |
| tensor<int32, [3]> var_2028_begin_0 = const()[name = string("op_2028_begin_0"), val = tensor<int32, [3]>([0, 1, 0])]; |
| tensor<int32, [3]> var_2028_end_0 = const()[name = string("op_2028_end_0"), val = tensor<int32, [3]>([1, 19, 256])]; |
| tensor<bool, [3]> var_2028_end_mask_0 = const()[name = string("op_2028_end_mask_0"), val = tensor<bool, [3]>([true, true, true])]; |
| tensor<fp32, [1, 18, 256]> var_2028 = slice_by_index(begin = var_2028_begin_0, end = var_2028_end_0, end_mask = var_2028_end_mask_0, x = top_buffer)[name = string("op_2028")]; |
| int32 var_2035 = const()[name = string("op_2035"), val = int32(1)]; |
| bool top_buffer_interleave_0 = const()[name = string("top_buffer_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 19, 256]> top_buffer_out = concat(axis = var_2035, interleave = top_buffer_interleave_0, values = (var_2028, appended_encoder_frame))[name = string("top_buffer")]; |
| tensor<int32, [3]> var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string var_2054_pad_type_0 = const()[name = string("op_2054_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> var_2054_strides_0 = const()[name = string("op_2054_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> var_2054_pad_0 = const()[name = string("op_2054_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> var_2054_dilations_0 = const()[name = string("op_2054_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 var_2054_groups_0 = const()[name = string("op_2054_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 256, 19]> var_2039 = transpose(perm = var_2039_perm_0, x = top_buffer_out)[name = string("transpose_27")]; |
| tensor<fp32, [1, 256, 1]> var_2054 = conv(bias = model_cnn_bias, dilations = var_2054_dilations_0, groups = var_2054_groups_0, pad = var_2054_pad_0, pad_type = var_2054_pad_type_0, strides = var_2054_strides_0, weight = model_cnn_weight, x = var_2039)[name = string("op_2054")]; |
| tensor<int32, [3]> input_171_perm_0 = const()[name = string("input_171_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [1]> var_2059 = const()[name = string("op_2059"), val = tensor<int32, [1]>([-1])]; |
| bool var_2060 = const()[name = string("op_2060"), val = bool(true)]; |
| tensor<fp32, [1, 1, 256]> input_171 = transpose(perm = input_171_perm_0, x = var_2054)[name = string("transpose_26")]; |
| tensor<fp32, [1, 1, 1]> var_2061 = reduce_l2_norm(axes = var_2059, keep_dims = var_2060, x = input_171)[name = string("op_2061")]; |
| fp32 var_2062 = const()[name = string("op_2062"), val = fp32(0x1.a36e2ep-14)]; |
| tensor<fp32, [1, 1, 1]> var_2063 = maximum(x = var_2061, y = var_2062)[name = string("op_2063")]; |
| tensor<fp32, [1, 1, 256]> x_19 = real_div(x = input_171, y = var_2063)[name = string("x_19")]; |
| tensor<fp32, [1, 1, 12, 256]> var_2074 = const()[name = string("op_2074"), val = tensor<fp32, [1, 1, 12, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42803008)))]; |
| tensor<int32, [1]> var_2080_axes_0 = const()[name = string("op_2080_axes_0"), val = tensor<int32, [1]>([2])]; |
| tensor<fp32, [1, 1, 1, 256]> var_2080 = expand_dims(axes = var_2080_axes_0, x = x_19)[name = string("op_2080")]; |
| tensor<int32, [4]> var_2085 = const()[name = string("op_2085"), val = tensor<int32, [4]>([1, 1, 12, 1])]; |
| tensor<fp32, [1, 1, 12, 256]> repeated_emb = tile(reps = var_2085, x = var_2080)[name = string("repeated_emb")]; |
| int32 var_2088 = const()[name = string("op_2088"), val = int32(-1)]; |
| bool input_173_interleave_0 = const()[name = string("input_173_interleave_0"), val = bool(false)]; |
| tensor<fp32, [1, 1, 12, 512]> input_173 = concat(axis = var_2088, interleave = input_173_interleave_0, values = (repeated_emb, var_2074))[name = string("input_173")]; |
| tensor<fp32, [1, 1, 12, 256]> src_1 = linear(bias = model_dec_convert_bias, weight = model_dec_convert_weight, x = input_173)[name = string("linear_37")]; |
| tensor<int32, [5]> old_kv_9_begin_0 = const()[name = string("old_kv_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_9_end_0 = const()[name = string("old_kv_9_end_0"), val = tensor<int32, [5]>([1, 12, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_9_end_mask_0 = const()[name = string("old_kv_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_9_squeeze_mask_0 = const()[name = string("old_kv_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [12, 4, 64, 64]> old_kv_9 = slice_by_index(begin = old_kv_9_begin_0, end = old_kv_9_end_0, end_mask = old_kv_9_end_mask_0, squeeze_mask = old_kv_9_squeeze_mask_0, x = dec_ret_kv)[name = string("old_kv_9")]; |
| tensor<int32, [3]> old_scale_9_begin_0 = const()[name = string("old_scale_9_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> old_scale_9_end_0 = const()[name = string("old_scale_9_end_0"), val = tensor<int32, [3]>([1, 12, 4])]; |
| tensor<bool, [3]> old_scale_9_end_mask_0 = const()[name = string("old_scale_9_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_9_squeeze_mask_0 = const()[name = string("old_scale_9_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [12, 4]> old_scale_9 = slice_by_index(begin = old_scale_9_begin_0, end = old_scale_9_end_0, end_mask = old_scale_9_end_mask_0, squeeze_mask = old_scale_9_squeeze_mask_0, x = dec_ret_scale)[name = string("old_scale_9")]; |
| tensor<int32, [4]> var_2125_perm_0 = const()[name = string("op_2125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2128 = const()[name = string("op_2128"), val = tensor<int32, [3]>([12, 1, 256])]; |
| tensor<fp32, [1, 12, 1, 256]> var_2125 = transpose(perm = var_2125_perm_0, x = src_1)[name = string("transpose_25")]; |
| tensor<fp32, [12, 1, 256]> x_21 = reshape(shape = var_2128, x = var_2125)[name = string("x_21")]; |
| tensor<fp32, [12, 1, 256]> q_17 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn1_q_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn1_q_proj_weight, x = x_21)[name = string("linear_38")]; |
| tensor<fp32, [12, 1, 256]> k_25 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn1_k_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn1_k_proj_weight, x = x_21)[name = string("linear_39")]; |
| tensor<fp32, [12, 1, 256]> v_17 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn1_v_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn1_v_proj_weight, x = x_21)[name = string("linear_40")]; |
| tensor<fp32, [12, 1, 256]> input_177 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn1_g_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn1_g_proj_weight, x = x_21)[name = string("linear_41")]; |
| fp32 var_2155 = const()[name = string("op_2155"), val = fp32(0x1p-3)]; |
| tensor<fp32, [12, 1, 256]> k_27 = mul(x = k_25, y = var_2155)[name = string("k_27")]; |
| tensor<int32, [4]> var_2159 = const()[name = string("op_2159"), val = tensor<int32, [4]>([12, 1, 4, 64])]; |
| tensor<fp32, [12, 1, 4, 64]> var_2160 = reshape(shape = var_2159, x = q_17)[name = string("op_2160")]; |
| tensor<int32, [4]> q_19_perm_0 = const()[name = string("q_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_2166 = const()[name = string("op_2166"), val = tensor<int32, [4]>([12, 1, 4, 64])]; |
| tensor<fp32, [12, 1, 4, 64]> var_2167 = reshape(shape = var_2166, x = k_27)[name = string("op_2167")]; |
| tensor<int32, [4]> k_29_perm_0 = const()[name = string("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_2174 = const()[name = string("op_2174"), val = tensor<int32, [4]>([12, 4, 64, 1])]; |
| tensor<fp32, [12, 4, 64, 1]> v_19 = reshape(shape = var_2174, x = v_17)[name = string("v_19")]; |
| tensor<fp32, [12, 4, 1, 64]> k_29 = transpose(perm = k_29_perm_0, x = var_2167)[name = string("transpose_23")]; |
| tensor<fp32, [12, 4, 64, 64]> kv_9 = mul(x = k_29, y = v_19)[name = string("kv_9")]; |
| fp32 var_2190 = const()[name = string("op_2190"), val = fp32(0x1p+0)]; |
| tensor<fp32, [12, 4]> candidate_scale_9 = add(x = old_scale_9, y = var_2190)[name = string("candidate_scale_9")]; |
| tensor<fp32, [12, 4]> var_2192 = sqrt(x = old_scale_9)[name = string("op_2192")]; |
| tensor<fp32, [12, 4]> var_2194 = sqrt(x = candidate_scale_9)[name = string("op_2194")]; |
| tensor<fp32, [12, 4]> var_2195 = real_div(x = var_2192, y = var_2194)[name = string("op_2195")]; |
| tensor<int32, [1]> var_2197_axes_0 = const()[name = string("op_2197_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1]> var_2197 = expand_dims(axes = var_2197_axes_0, x = var_2195)[name = string("op_2197")]; |
| tensor<int32, [1]> blend_9_axes_0 = const()[name = string("blend_9_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1, 1]> blend_9 = expand_dims(axes = blend_9_axes_0, x = var_2197)[name = string("blend_9")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2200 = mul(x = old_kv_9, y = blend_9)[name = string("op_2200")]; |
| tensor<int32, [1]> var_2203_axes_0 = const()[name = string("op_2203_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1]> var_2203 = expand_dims(axes = var_2203_axes_0, x = var_2194)[name = string("op_2203")]; |
| tensor<int32, [1]> var_2205_axes_0 = const()[name = string("op_2205_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1, 1]> var_2205 = expand_dims(axes = var_2205_axes_0, x = var_2203)[name = string("op_2205")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2206 = real_div(x = kv_9, y = var_2205)[name = string("op_2206")]; |
| tensor<fp32, [12, 4, 64, 64]> candidate_kv_9 = add(x = var_2200, y = var_2206)[name = string("candidate_kv_9")]; |
| tensor<fp32, [12, 4, 1, 64]> q_19 = transpose(perm = q_19_perm_0, x = var_2160)[name = string("transpose_24")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2209 = mul(x = q_19, y = candidate_kv_9)[name = string("op_2209")]; |
| tensor<int32, [1]> input_175_axes_0 = const()[name = string("input_175_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_175_keep_dims_0 = const()[name = string("input_175_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [12, 4, 64]> input_175 = reduce_sum(axes = input_175_axes_0, keep_dims = input_175_keep_dims_0, x = var_2209)[name = string("input_175")]; |
| fp32 var_2216 = const()[name = string("op_2216"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_2220_axes_0 = const()[name = string("op_2220_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 64]> var_2220 = layer_norm(axes = var_2220_axes_0, epsilon = var_2216, x = input_175)[name = string("op_2220")]; |
| tensor<int32, [3]> var_2222 = const()[name = string("op_2222"), val = tensor<int32, [3]>([12, 1, 256])]; |
| tensor<fp32, [12, 1, 256]> output_9 = reshape(shape = var_2222, x = var_2220)[name = string("output_9")]; |
| tensor<fp32, [12, 1, 256]> var_2224 = silu(x = input_177)[name = string("op_2224")]; |
| tensor<fp32, [12, 1, 256]> input_179 = mul(x = var_2224, y = output_9)[name = string("input_179")]; |
| tensor<fp32, [12, 1, 256]> input_181 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn1_out_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn1_out_proj_weight, x = input_179)[name = string("linear_42")]; |
| tensor<fp32, [12, 1, 256]> input_183 = add(x = x_21, y = input_181)[name = string("input_183")]; |
| fp32 var_2235 = const()[name = string("op_2235"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 1, 256]> x_23 = layer_norm(axes = x_23_axes_0, beta = model_dec_attractor_decoder_layers_0_norm11_bias, epsilon = var_2235, gamma = model_dec_attractor_decoder_layers_0_norm11_weight, x = input_183)[name = string("x_23")]; |
| tensor<int32, [4]> var_2241 = const()[name = string("op_2241"), val = tensor<int32, [4]>([1, 12, 1, 256])]; |
| tensor<fp32, [1, 12, 1, 256]> var_2242 = reshape(shape = var_2241, x = x_23)[name = string("op_2242")]; |
| tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2248 = const()[name = string("op_2248"), val = tensor<int32, [3]>([1, 12, 256])]; |
| tensor<fp32, [1, 1, 12, 256]> x_25 = transpose(perm = x_25_perm_0, x = var_2242)[name = string("transpose_22")]; |
| tensor<fp32, [1, 12, 256]> x_27 = reshape(shape = var_2248, x = x_25)[name = string("x_27")]; |
| tensor<fp32, [256, 256]> var_2273_0 = const()[name = string("op_2273_0"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42815360)))]; |
| tensor<fp32, [256, 256]> var_2273_1 = const()[name = string("op_2273_1"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43077568)))]; |
| tensor<fp32, [256, 256]> var_2273_2 = const()[name = string("op_2273_2"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43339776)))]; |
| tensor<fp32, [256]> var_2276_0 = const()[name = string("op_2276_0"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43601984)))]; |
| tensor<fp32, [256]> var_2276_1 = const()[name = string("op_2276_1"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43603072)))]; |
| tensor<fp32, [256]> var_2276_2 = const()[name = string("op_2276_2"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43604160)))]; |
| tensor<fp32, [1, 12, 256]> q_21 = linear(bias = var_2276_0, weight = var_2273_0, x = x_27)[name = string("linear_43")]; |
| tensor<fp32, [1, 12, 256]> k_31 = linear(bias = var_2276_1, weight = var_2273_1, x = x_27)[name = string("linear_44")]; |
| tensor<fp32, [1, 12, 256]> v_21 = linear(bias = var_2276_2, weight = var_2273_2, x = x_27)[name = string("linear_45")]; |
| tensor<int32, [4]> var_2283 = const()[name = string("op_2283"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2284 = reshape(shape = var_2283, x = q_21)[name = string("op_2284")]; |
| tensor<int32, [4]> var_2289 = const()[name = string("op_2289"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2290 = reshape(shape = var_2289, x = k_31)[name = string("op_2290")]; |
| tensor<int32, [4]> var_2295 = const()[name = string("op_2295"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2296 = reshape(shape = var_2295, x = v_21)[name = string("op_2296")]; |
| tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool var_2303_transpose_x_0 = const()[name = string("op_2303_transpose_x_0"), val = bool(false)]; |
| bool var_2303_transpose_y_0 = const()[name = string("op_2303_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp32, [1, 4, 64, 12]> transpose_7 = transpose(perm = transpose_7_perm_0, x = var_2290)[name = string("transpose_19")]; |
| tensor<fp32, [1, 4, 12, 64]> transpose_6 = transpose(perm = transpose_6_perm_0, x = var_2284)[name = string("transpose_20")]; |
| tensor<fp32, [1, 4, 12, 12]> var_2303 = matmul(transpose_x = var_2303_transpose_x_0, transpose_y = var_2303_transpose_y_0, x = transpose_6, y = transpose_7)[name = string("op_2303")]; |
| tensor<fp32, [1]> _inversed_attn_1_y_0 = const()[name = string("_inversed_attn_1_y_0"), val = tensor<fp32, [1]>([0x1p-3])]; |
| tensor<fp32, [1, 4, 12, 12]> _inversed_attn_1 = mul(x = var_2303, y = _inversed_attn_1_y_0)[name = string("_inversed_attn_1")]; |
| int32 var_2307 = const()[name = string("op_2307"), val = int32(-1)]; |
| tensor<fp32, [1, 4, 12, 12]> attn_3 = softmax(axis = var_2307, x = _inversed_attn_1)[name = string("attn_3")]; |
| bool out_1_transpose_x_0 = const()[name = string("out_1_transpose_x_0"), val = bool(false)]; |
| bool out_1_transpose_y_0 = const()[name = string("out_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 12, 64]> v_23 = transpose(perm = v_23_perm_0, x = var_2296)[name = string("transpose_21")]; |
| tensor<fp32, [1, 4, 12, 64]> out_1 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3, y = v_23)[name = string("out_1")]; |
| tensor<int32, [4]> var_2313_perm_0 = const()[name = string("op_2313_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2314 = const()[name = string("op_2314"), val = tensor<int32, [3]>([1, 12, 256])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2313 = transpose(perm = var_2313_perm_0, x = out_1)[name = string("transpose_18")]; |
| tensor<fp32, [1, 12, 256]> out_3 = reshape(shape = var_2314, x = var_2313)[name = string("out_3")]; |
| tensor<fp32, [1, 12, 256]> var_2316 = linear(bias = model_dec_attractor_decoder_layers_0_self_attn2_out_proj_bias, weight = model_dec_attractor_decoder_layers_0_self_attn2_out_proj_weight, x = out_3)[name = string("linear_46")]; |
| tensor<fp32, [1, 12, 256]> input_185 = add(x = x_27, y = var_2316)[name = string("input_185")]; |
| fp32 var_2320 = const()[name = string("op_2320"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 12, 256]> input_187 = layer_norm(axes = input_187_axes_0, beta = model_dec_attractor_decoder_layers_0_norm21_bias, epsilon = var_2320, gamma = model_dec_attractor_decoder_layers_0_norm21_weight, x = input_185)[name = string("input_187")]; |
| tensor<fp32, [1, 12, 2048]> input_189 = linear(bias = model_dec_attractor_decoder_layers_0_linear1_bias, weight = model_dec_attractor_decoder_layers_0_linear1_weight, x = input_187)[name = string("linear_47")]; |
| tensor<fp32, [1, 12, 2048]> input_191 = relu(x = input_189)[name = string("input_191")]; |
| tensor<fp32, [1, 12, 256]> input_195 = linear(bias = model_dec_attractor_decoder_layers_0_linear2_bias, weight = model_dec_attractor_decoder_layers_0_linear2_weight, x = input_191)[name = string("linear_48")]; |
| tensor<fp32, [1, 12, 256]> input_197 = add(x = input_187, y = input_195)[name = string("input_197")]; |
| fp32 var_2342 = const()[name = string("op_2342"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_29_axes_0 = const()[name = string("x_29_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 12, 256]> x_29 = layer_norm(axes = x_29_axes_0, beta = model_dec_attractor_decoder_layers_0_norm22_bias, epsilon = var_2342, gamma = model_dec_attractor_decoder_layers_0_norm22_weight, x = input_197)[name = string("x_29")]; |
| tensor<int32, [4]> var_2348 = const()[name = string("op_2348"), val = tensor<int32, [4]>([1, 1, 12, 256])]; |
| tensor<fp32, [1, 1, 12, 256]> src = reshape(shape = var_2348, x = x_29)[name = string("src")]; |
| tensor<int32, [5]> old_kv_begin_0 = const()[name = string("old_kv_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> old_kv_end_0 = const()[name = string("old_kv_end_0"), val = tensor<int32, [5]>([2, 12, 4, 64, 64])]; |
| tensor<bool, [5]> old_kv_end_mask_0 = const()[name = string("old_kv_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> old_kv_squeeze_mask_0 = const()[name = string("old_kv_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [12, 4, 64, 64]> old_kv = slice_by_index(begin = old_kv_begin_0, end = old_kv_end_0, end_mask = old_kv_end_mask_0, squeeze_mask = old_kv_squeeze_mask_0, x = dec_ret_kv)[name = string("old_kv")]; |
| tensor<int32, [3]> old_scale_begin_0 = const()[name = string("old_scale_begin_0"), val = tensor<int32, [3]>([1, 0, 0])]; |
| tensor<int32, [3]> old_scale_end_0 = const()[name = string("old_scale_end_0"), val = tensor<int32, [3]>([2, 12, 4])]; |
| tensor<bool, [3]> old_scale_end_mask_0 = const()[name = string("old_scale_end_mask_0"), val = tensor<bool, [3]>([false, true, true])]; |
| tensor<bool, [3]> old_scale_squeeze_mask_0 = const()[name = string("old_scale_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])]; |
| tensor<fp32, [12, 4]> old_scale = slice_by_index(begin = old_scale_begin_0, end = old_scale_end_0, end_mask = old_scale_end_mask_0, squeeze_mask = old_scale_squeeze_mask_0, x = dec_ret_scale)[name = string("old_scale")]; |
| tensor<int32, [4]> var_2382_perm_0 = const()[name = string("op_2382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2385 = const()[name = string("op_2385"), val = tensor<int32, [3]>([12, 1, 256])]; |
| tensor<fp32, [1, 12, 1, 256]> var_2382 = transpose(perm = var_2382_perm_0, x = src)[name = string("transpose_17")]; |
| tensor<fp32, [12, 1, 256]> x_31 = reshape(shape = var_2385, x = var_2382)[name = string("x_31")]; |
| tensor<fp32, [12, 1, 256]> q_25 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn1_q_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn1_q_proj_weight, x = x_31)[name = string("linear_49")]; |
| tensor<fp32, [12, 1, 256]> k_35 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn1_k_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn1_k_proj_weight, x = x_31)[name = string("linear_50")]; |
| tensor<fp32, [12, 1, 256]> v_25 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn1_v_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn1_v_proj_weight, x = x_31)[name = string("linear_51")]; |
| tensor<fp32, [12, 1, 256]> input_201 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn1_g_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn1_g_proj_weight, x = x_31)[name = string("linear_52")]; |
| fp32 var_2412 = const()[name = string("op_2412"), val = fp32(0x1p-3)]; |
| tensor<fp32, [12, 1, 256]> k_37 = mul(x = k_35, y = var_2412)[name = string("k_37")]; |
| tensor<int32, [4]> var_2416 = const()[name = string("op_2416"), val = tensor<int32, [4]>([12, 1, 4, 64])]; |
| tensor<fp32, [12, 1, 4, 64]> var_2417 = reshape(shape = var_2416, x = q_25)[name = string("op_2417")]; |
| tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_2423 = const()[name = string("op_2423"), val = tensor<int32, [4]>([12, 1, 4, 64])]; |
| tensor<fp32, [12, 1, 4, 64]> var_2424 = reshape(shape = var_2423, x = k_37)[name = string("op_2424")]; |
| tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [4]> var_2431 = const()[name = string("op_2431"), val = tensor<int32, [4]>([12, 4, 64, 1])]; |
| tensor<fp32, [12, 4, 64, 1]> v_27 = reshape(shape = var_2431, x = v_25)[name = string("v_27")]; |
| tensor<fp32, [12, 4, 1, 64]> k_39 = transpose(perm = k_39_perm_0, x = var_2424)[name = string("transpose_15")]; |
| tensor<fp32, [12, 4, 64, 64]> kv = mul(x = k_39, y = v_27)[name = string("kv")]; |
| fp32 var_2446 = const()[name = string("op_2446"), val = fp32(0x1p+0)]; |
| tensor<fp32, [12, 4]> candidate_scale = add(x = old_scale, y = var_2446)[name = string("candidate_scale")]; |
| tensor<fp32, [12, 4]> var_2448 = sqrt(x = old_scale)[name = string("op_2448")]; |
| tensor<fp32, [12, 4]> var_2450 = sqrt(x = candidate_scale)[name = string("op_2450")]; |
| tensor<fp32, [12, 4]> var_2451 = real_div(x = var_2448, y = var_2450)[name = string("op_2451")]; |
| tensor<int32, [1]> var_2453_axes_0 = const()[name = string("op_2453_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1]> var_2453 = expand_dims(axes = var_2453_axes_0, x = var_2451)[name = string("op_2453")]; |
| tensor<int32, [1]> blend_axes_0 = const()[name = string("blend_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1, 1]> blend = expand_dims(axes = blend_axes_0, x = var_2453)[name = string("blend")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2456 = mul(x = old_kv, y = blend)[name = string("op_2456")]; |
| tensor<int32, [1]> var_2459_axes_0 = const()[name = string("op_2459_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1]> var_2459 = expand_dims(axes = var_2459_axes_0, x = var_2450)[name = string("op_2459")]; |
| tensor<int32, [1]> var_2461_axes_0 = const()[name = string("op_2461_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 1, 1]> var_2461 = expand_dims(axes = var_2461_axes_0, x = var_2459)[name = string("op_2461")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2462 = real_div(x = kv, y = var_2461)[name = string("op_2462")]; |
| tensor<fp32, [12, 4, 64, 64]> candidate_kv = add(x = var_2456, y = var_2462)[name = string("candidate_kv")]; |
| tensor<fp32, [12, 4, 1, 64]> q_27 = transpose(perm = q_27_perm_0, x = var_2417)[name = string("transpose_16")]; |
| tensor<fp32, [12, 4, 64, 64]> var_2465 = mul(x = q_27, y = candidate_kv)[name = string("op_2465")]; |
| tensor<int32, [1]> input_199_axes_0 = const()[name = string("input_199_axes_0"), val = tensor<int32, [1]>([3])]; |
| bool input_199_keep_dims_0 = const()[name = string("input_199_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [12, 4, 64]> input_199 = reduce_sum(axes = input_199_axes_0, keep_dims = input_199_keep_dims_0, x = var_2465)[name = string("input_199")]; |
| fp32 var_2472 = const()[name = string("op_2472"), val = fp32(0x1.0c6f7ap-20)]; |
| tensor<int32, [1]> var_2476_axes_0 = const()[name = string("op_2476_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 4, 64]> var_2476 = layer_norm(axes = var_2476_axes_0, epsilon = var_2472, x = input_199)[name = string("op_2476")]; |
| tensor<int32, [3]> var_2478 = const()[name = string("op_2478"), val = tensor<int32, [3]>([12, 1, 256])]; |
| tensor<fp32, [12, 1, 256]> output = reshape(shape = var_2478, x = var_2476)[name = string("output")]; |
| tensor<fp32, [12, 1, 256]> var_2480 = silu(x = input_201)[name = string("op_2480")]; |
| tensor<fp32, [12, 1, 256]> input_203 = mul(x = var_2480, y = output)[name = string("input_203")]; |
| tensor<fp32, [12, 1, 256]> input_205 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn1_out_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn1_out_proj_weight, x = input_203)[name = string("linear_53")]; |
| tensor<fp32, [12, 1, 256]> input_207 = add(x = x_31, y = input_205)[name = string("input_207")]; |
| fp32 var_2491 = const()[name = string("op_2491"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_33_axes_0 = const()[name = string("x_33_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [12, 1, 256]> x_33 = layer_norm(axes = x_33_axes_0, beta = model_dec_attractor_decoder_layers_1_norm11_bias, epsilon = var_2491, gamma = model_dec_attractor_decoder_layers_1_norm11_weight, x = input_207)[name = string("x_33")]; |
| tensor<int32, [4]> var_2497 = const()[name = string("op_2497"), val = tensor<int32, [4]>([1, 12, 1, 256])]; |
| tensor<fp32, [1, 12, 1, 256]> var_2498 = reshape(shape = var_2497, x = x_33)[name = string("op_2498")]; |
| tensor<int32, [4]> x_35_perm_0 = const()[name = string("x_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2504 = const()[name = string("op_2504"), val = tensor<int32, [3]>([1, 12, 256])]; |
| tensor<fp32, [1, 1, 12, 256]> x_35 = transpose(perm = x_35_perm_0, x = var_2498)[name = string("transpose_14")]; |
| tensor<fp32, [1, 12, 256]> x_37 = reshape(shape = var_2504, x = x_35)[name = string("x_37")]; |
| tensor<fp32, [256, 256]> var_2529_0 = const()[name = string("op_2529_0"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43605248)))]; |
| tensor<fp32, [256, 256]> var_2529_1 = const()[name = string("op_2529_1"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43867456)))]; |
| tensor<fp32, [256, 256]> var_2529_2 = const()[name = string("op_2529_2"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44129664)))]; |
| tensor<fp32, [256]> var_2532_0 = const()[name = string("op_2532_0"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44391872)))]; |
| tensor<fp32, [256]> var_2532_1 = const()[name = string("op_2532_1"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44392960)))]; |
| tensor<fp32, [256]> var_2532_2 = const()[name = string("op_2532_2"), val = tensor<fp32, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44394048)))]; |
| tensor<fp32, [1, 12, 256]> q_29 = linear(bias = var_2532_0, weight = var_2529_0, x = x_37)[name = string("linear_54")]; |
| tensor<fp32, [1, 12, 256]> k_41 = linear(bias = var_2532_1, weight = var_2529_1, x = x_37)[name = string("linear_55")]; |
| tensor<fp32, [1, 12, 256]> v_29 = linear(bias = var_2532_2, weight = var_2529_2, x = x_37)[name = string("linear_56")]; |
| tensor<int32, [4]> var_2539 = const()[name = string("op_2539"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2540 = reshape(shape = var_2539, x = q_29)[name = string("op_2540")]; |
| tensor<int32, [4]> var_2545 = const()[name = string("op_2545"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2546 = reshape(shape = var_2545, x = k_41)[name = string("op_2546")]; |
| tensor<int32, [4]> var_2551 = const()[name = string("op_2551"), val = tensor<int32, [4]>([1, 12, 4, 64])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2552 = reshape(shape = var_2551, x = v_29)[name = string("op_2552")]; |
| tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool var_2559_transpose_x_0 = const()[name = string("op_2559_transpose_x_0"), val = bool(false)]; |
| bool var_2559_transpose_y_0 = const()[name = string("op_2559_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp32, [1, 4, 64, 12]> transpose_9 = transpose(perm = transpose_9_perm_0, x = var_2546)[name = string("transpose_11")]; |
| tensor<fp32, [1, 4, 12, 64]> transpose_8 = transpose(perm = transpose_8_perm_0, x = var_2540)[name = string("transpose_12")]; |
| tensor<fp32, [1, 4, 12, 12]> var_2559 = matmul(transpose_x = var_2559_transpose_x_0, transpose_y = var_2559_transpose_y_0, x = transpose_8, y = transpose_9)[name = string("op_2559")]; |
| tensor<fp32, [1]> _inversed_attn_5_y_0 = const()[name = string("_inversed_attn_5_y_0"), val = tensor<fp32, [1]>([0x1p-3])]; |
| tensor<fp32, [1, 4, 12, 12]> _inversed_attn_5 = mul(x = var_2559, y = _inversed_attn_5_y_0)[name = string("_inversed_attn_5")]; |
| int32 var_2563 = const()[name = string("op_2563"), val = int32(-1)]; |
| tensor<fp32, [1, 4, 12, 12]> attn = softmax(axis = var_2563, x = _inversed_attn_5)[name = string("attn")]; |
| bool out_5_transpose_x_0 = const()[name = string("out_5_transpose_x_0"), val = bool(false)]; |
| bool out_5_transpose_y_0 = const()[name = string("out_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [1, 4, 12, 64]> v = transpose(perm = v_perm_0, x = var_2552)[name = string("transpose_13")]; |
| tensor<fp32, [1, 4, 12, 64]> out_5 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn, y = v)[name = string("out_5")]; |
| tensor<int32, [4]> var_2569_perm_0 = const()[name = string("op_2569_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_2570 = const()[name = string("op_2570"), val = tensor<int32, [3]>([1, 12, 256])]; |
| tensor<fp32, [1, 12, 4, 64]> var_2569 = transpose(perm = var_2569_perm_0, x = out_5)[name = string("transpose_10")]; |
| tensor<fp32, [1, 12, 256]> out = reshape(shape = var_2570, x = var_2569)[name = string("out")]; |
| tensor<fp32, [1, 12, 256]> var_2572 = linear(bias = model_dec_attractor_decoder_layers_1_self_attn2_out_proj_bias, weight = model_dec_attractor_decoder_layers_1_self_attn2_out_proj_weight, x = out)[name = string("linear_57")]; |
| tensor<fp32, [1, 12, 256]> input_209 = add(x = x_37, y = var_2572)[name = string("input_209")]; |
| fp32 var_2576 = const()[name = string("op_2576"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 12, 256]> input_211 = layer_norm(axes = input_211_axes_0, beta = model_dec_attractor_decoder_layers_1_norm21_bias, epsilon = var_2576, gamma = model_dec_attractor_decoder_layers_1_norm21_weight, x = input_209)[name = string("input_211")]; |
| tensor<fp32, [1, 12, 2048]> input_213 = linear(bias = model_dec_attractor_decoder_layers_1_linear1_bias, weight = model_dec_attractor_decoder_layers_1_linear1_weight, x = input_211)[name = string("linear_58")]; |
| tensor<fp32, [1, 12, 2048]> input_215 = relu(x = input_213)[name = string("input_215")]; |
| tensor<fp32, [1, 12, 256]> input_219 = linear(bias = model_dec_attractor_decoder_layers_1_linear2_bias, weight = model_dec_attractor_decoder_layers_1_linear2_weight, x = input_215)[name = string("linear_59")]; |
| tensor<fp32, [1, 12, 256]> input_221 = add(x = input_211, y = input_219)[name = string("input_221")]; |
| fp32 var_2598 = const()[name = string("op_2598"), val = fp32(0x1.4f8b58p-17)]; |
| tensor<int32, [1]> x_axes_0 = const()[name = string("x_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 12, 256]> x = layer_norm(axes = x_axes_0, beta = model_dec_attractor_decoder_layers_1_norm22_bias, epsilon = var_2598, gamma = model_dec_attractor_decoder_layers_1_norm22_weight, x = input_221)[name = string("x")]; |
| tensor<int32, [4]> var_2604 = const()[name = string("op_2604"), val = tensor<int32, [4]>([1, 1, 12, 256])]; |
| tensor<fp32, [1, 1, 12, 256]> input = reshape(shape = var_2604, x = x)[name = string("input")]; |
| tensor<int32, [1]> var_2607 = const()[name = string("op_2607"), val = tensor<int32, [1]>([-1])]; |
| bool var_2608 = const()[name = string("op_2608"), val = bool(true)]; |
| tensor<fp32, [1, 1, 12, 1]> var_2609 = reduce_l2_norm(axes = var_2607, keep_dims = var_2608, x = input)[name = string("op_2609")]; |
| fp32 var_2610 = const()[name = string("op_2610"), val = fp32(0x1.a36e2ep-14)]; |
| tensor<fp32, [1, 1, 12, 1]> var_2611 = maximum(x = var_2609, y = var_2610)[name = string("op_2611")]; |
| tensor<fp32, [1, 1, 12, 256]> attractors = real_div(x = input, y = var_2611)[name = string("attractors")]; |
| tensor<int32, [1]> var_2614_axes_0 = const()[name = string("op_2614_axes_0"), val = tensor<int32, [1]>([-2])]; |
| tensor<fp32, [1, 1, 1, 256]> var_2614 = expand_dims(axes = var_2614_axes_0, x = x_19)[name = string("op_2614")]; |
| bool var_2618_transpose_x_1 = const()[name = string("op_2618_transpose_x_1"), val = bool(false)]; |
| bool var_2618_transpose_y_1 = const()[name = string("op_2618_transpose_y_1"), val = bool(true)]; |
| tensor<fp32, [1, 1, 1, 12]> var_2618 = matmul(transpose_x = var_2618_transpose_x_1, transpose_y = var_2618_transpose_y_1, x = var_2614, y = attractors)[name = string("op_2618")]; |
| tensor<int32, [1]> logits_axes_0 = const()[name = string("logits_axes_0"), val = tensor<int32, [1]>([-2])]; |
| tensor<fp32, [1, 1, 12]> logits = squeeze(axes = logits_axes_0, x = var_2618)[name = string("logits")]; |
| int32 candidate_dec_ret_kv_axis_0 = const()[name = string("candidate_dec_ret_kv_axis_0"), val = int32(0)]; |
| tensor<fp32, [2, 12, 4, 64, 64]> candidate_dec_ret_kv = stack(axis = candidate_dec_ret_kv_axis_0, values = (candidate_kv_9, candidate_kv))[name = string("candidate_dec_ret_kv")]; |
| int32 candidate_dec_ret_scale_axis_0 = const()[name = string("candidate_dec_ret_scale_axis_0"), val = int32(0)]; |
| tensor<fp32, [2, 12, 4]> candidate_dec_ret_scale = stack(axis = candidate_dec_ret_scale_axis_0, values = (candidate_scale_9, candidate_scale))[name = string("candidate_dec_ret_scale")]; |
| tensor<fp32, [2, 12, 4, 64, 64]> var_2628 = sub(x = candidate_dec_ret_kv, y = dec_ret_kv)[name = string("op_2628")]; |
| tensor<int32, [1]> var_2630_axes_0 = const()[name = string("op_2630_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 1, 1]> var_2630 = expand_dims(axes = var_2630_axes_0, x = decode_scalar)[name = string("op_2630")]; |
| tensor<fp32, [2, 12, 4, 64, 64]> var_2631 = mul(x = var_2628, y = var_2630)[name = string("op_2631")]; |
| tensor<fp32, [2, 12, 4, 64, 64]> dec_ret_kv_out = add(x = dec_ret_kv, y = var_2631)[name = string("op_2633")]; |
| tensor<fp32, [2, 12, 4]> var_2635 = sub(x = candidate_dec_ret_scale, y = dec_ret_scale)[name = string("op_2635")]; |
| tensor<int32, [1]> var_2637_axes_0 = const()[name = string("op_2637_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, 1, 1]> var_2637 = expand_dims(axes = var_2637_axes_0, x = decode_vec)[name = string("op_2637")]; |
| tensor<fp32, [2, 12, 4]> var_2638 = mul(x = var_2635, y = var_2637)[name = string("op_2638")]; |
| tensor<fp32, [2, 12, 4]> dec_ret_scale_out = add(x = dec_ret_scale, y = var_2638)[name = string("op_2640")]; |
| tensor<fp32, [1, 1, 12]> full_logits = mul(x = logits, y = decode_scalar)[name = string("op_2641")]; |
| int32 var_2644_axis_0 = const()[name = string("op_2644_axis_0"), val = int32(0)]; |
| tensor<fp32, [4, 1, 4, 64, 64]> enc_ret_kv_out = stack(axis = var_2644_axis_0, values = (var_1101, var_1403, var_1705, blended_kv))[name = string("op_2644")]; |
| int32 var_2647_axis_0 = const()[name = string("op_2647_axis_0"), val = int32(0)]; |
| tensor<fp32, [4, 1, 4]> enc_ret_scale_out = stack(axis = var_2647_axis_0, values = (var_1106, var_1408, var_1710, blended_scale))[name = string("op_2647")]; |
| int32 var_2650_axis_0 = const()[name = string("op_2650_axis_0"), val = int32(0)]; |
| tensor<fp32, [4, 1, 15, 256]> enc_conv_cache_out = stack(axis = var_2650_axis_0, values = (var_1111, var_1413, var_1715, blended_conv))[name = string("op_2650")]; |
| } -> (full_logits, enc_ret_kv_out, enc_ret_scale_out, enc_conv_cache_out, dec_ret_kv_out, dec_ret_scale_out, top_buffer_out); |
| } |
