Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fixbug

游雁f57b3788f2c9
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 047e652a9..bbaf65735 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -212,7 +212,6 @@ class AutoModel:
deep_update(model_conf, kwargs.get("model_conf", {}))
deep_update(model_conf, kwargs)
model = model_class(**model_conf, vocab_size=vocab_size)
- model.to(device)
# init_param
init_param = kwargs.get("init_param", None)
@@ -235,6 +234,7 @@ class AutoModel:
model.to(torch.float16)
elif kwargs.get("bf16", False):
model.to(torch.bfloat16)
+ model.to(device)
return model, kwargs
def __call__(self, *args, **cfg):
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.py
index 66c416b3d..dd806cf66 100644
--- a/funasr/models/llm_asr/model.py
+++ b/funasr/models/llm_asr/model.py
@@ -687,10 +687,8 @@ class LLMASR2(nn.Module):
# fp16
if kwargs.get("fp16", False):
speech = speech.to(torch.float16)
- encoder_out_lens = encoder_out_lens.to(torch.float16)
elif kwargs.get("bf16", False):
speech = speech.to(torch.bfloat16)
- encoder_out_lens = encoder_out_lens.to(torch.bfloat16)
encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
# audio_adaptor