[Export] support gpu inference

funasr/export/export_model.py

@@ -19,6 +19,7 @@ class ModelExport:
         self,
         cache_dir: Union[Path, str] = None,
         onnx: bool = True,
+        device: str = "cpu",
         quant: bool = True,
         fallback_num: int = 0,
         audio_in: str = None,
@@ -36,6 +37,7 @@ class ModelExport:
         )
         print("output dir: {}".format(self.cache_dir))
         self.onnx = onnx
+        self.device = device
         self.quant = quant
         self.fallback_num = fallback_num
         self.frontend = None
@@ -112,6 +114,10 @@ class ModelExport:
         else:
             dummy_input = model.get_dummy_inputs()
 
+        if self.device == 'cuda':
+            model = model.cuda()
+            dummy_input = tuple([i.cuda() for i in dummy_input])
+
         # model_script = torch.jit.script(model)
         model_script = torch.jit.trace(model, dummy_input)
         model_script.save(os.path.join(path, f'{model.model_name}.torchscripts'))
@@ -234,6 +240,7 @@ if __name__ == '__main__':
     parser.add_argument('--model-name', type=str, required=True)
     parser.add_argument('--export-dir', type=str, required=True)
     parser.add_argument('--type', type=str, default='onnx', help='["onnx", "torch"]')
+    parser.add_argument('--device', type=str, default='cpu', help='["cpu", "cuda"]')
     parser.add_argument('--quantize', type=str2bool, default=False, help='export quantized model')
     parser.add_argument('--fallback-num', type=int, default=0, help='amp fallback number')
     parser.add_argument('--audio_in', type=str, default=None, help='["wav", "wav.scp"]')
@@ -243,6 +250,7 @@ if __name__ == '__main__':
     export_model = ModelExport(
         cache_dir=args.export_dir,
         onnx=args.type == 'onnx',
+        device=args.device,
         quant=args.quantize,
         fallback_num=args.fallback_num,
         audio_in=args.audio_in,

funasr/runtime/python/libtorch/demo.py

@@ -2,9 +2,9 @@
 from funasr_torch import Paraformer
 
 model_dir = "/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
-model = Paraformer(model_dir, batch_size=1)
+model = Paraformer(model_dir, batch_size=1, device_id=0)
 
 wav_path = ['/nfs/zhifu.gzf/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']
 
 result = model(wav_path)
-print(result)
+print(result)

funasr/runtime/python/libtorch/funasr_torch/paraformer_bin.py

@@ -58,6 +58,9 @@ class Paraformer():
             end_idx = min(waveform_nums, beg_idx + self.batch_size)
             feats, feats_len = self.extract_feat(waveform_list[beg_idx:end_idx])
             try:
+                if int(device_id) != -1:
+                    feats = feats.cuda()
+                    feats_len = feats_len.cuda()
                 outputs = self.ort_infer(feats, feats_len)
                 am_scores, valid_token_lens = outputs[0], outputs[1]
                 if len(outputs) == 4:

funasr/runtime/python/onnxruntime/demo.py

@@ -7,9 +7,9 @@ model_dir = "/Users/shixian/code/funasr/export/damo/speech_paraformer-tiny-comma
 
 # if you use paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch, you should set pred_bias=0
 # plot_timestamp_to works only when using speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch
-model = Paraformer(model_dir, batch_size=2, plot_timestamp_to="./", pred_bias=0) 
+model = Paraformer(model_dir, batch_size=2, plot_timestamp_to="./", pred_bias=0, device_id=0)
 
 wav_path = "/Users/shixian/code/funasr/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch/example/asr_example.wav"
 
 result = model(wav_path)
-print(result)
+print(result)