输出词级别时间戳时（output_timestamp=True），无法批量推理 #207
Description
res = model.generate(
input=files,
cache={},
language="en", # "zh", "en", "yue", "ja", "ko", "nospeech"
use_itn=False,
batch_size=4,
output_timestamp=True
)
不添加output_timestamp=True 可以正常批量推理，添加了output_timestamp=True会报如下错误：
RuntimeError Traceback (most recent call last)
Cell In[3], line 20
6 files = '''
7 /data/training_share/l00845921/en2zh_wav/en_zh_mp3/74_food_part1_01/74_food_part1_01_5gCGBWiazhY#00-00-07_0.mp3
8 /data/training_share/l00845921/en2zh_wav/en_zh_mp3/74_food_part1_01/74_food_part1_01_6UPkwUvDk9Y#00-00-13_0.mp3
(...)
14 /data/training_share/l00845921/en2zh_wav/en_zh_mp3/74_food_part1_01/74_food_part1_01_3ucLC6aje3w#00-00-15_3.mp3
15 '''.strip().split('\n')
18 model = AutoModel(model=model_dir, trust_remote_code=True, device="cuda:0")
---> 20 res = model.generate(
21 input=files,
22 cache={},
23 language="en", # "zh", "en", "yue", "ja", "ko", "nospeech"
24 use_itn=False,
25 batch_size=4,
26 output_timestamp=True
27 )
File /usr/local/lib/python3.11/site-packages/funasr/auto/auto_model.py:260, in AutoModel.generate(self, input, input_len, **cfg)
258 def generate(self, input, input_len=None, **cfg):
259 if self.vad_model is None:
--> 260 return self.inference(input, input_len=input_len, **cfg)
262 else:
263 return self.inference_with_vad(input, input_len=input_len, **cfg)
File /usr/local/lib/python3.11/site-packages/funasr/auto/auto_model.py:302, in AutoModel.inference(self, input, input_len, model, kwargs, key, **cfg)
300 time1 = time.perf_counter()
301 with torch.no_grad():
--> 302 res = model.inference(**batch, **kwargs)
303 if isinstance(res, (list, tuple)):
304 results = res[0] if len(res) > 0 else [{"text": ""}]
File /data/xyh/SenseVoice/model.py:904, in SenseVoiceSmall.inference(self, data_in, data_lengths, key, tokenizer, frontend, **kwargs)
894 logits_speech[pred==self.blank_id, self.blank_id] = 0
896 align = ctc_forced_align(
897 logits_speech.unsqueeze(0).float(),
898 torch.Tensor(token_int[4:]).unsqueeze(0).long().to(logits_speech.device),
(...)
901 ignore_id=self.ignore_id,
902 )
--> 904 pred = groupby(align[0, :encoder_out_lens[0]])
905 _start = 0
906 token_id = 0
RuntimeError: CUDA error: device-side assert triggered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1.
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.