Skip to content

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

fix(fun_asr_nano): make bf16/fp16 inference work instead of crashing (#2978)

zhifu gaofaffb45a331e
diff --git a/funasr/models/fun_asr_nano/model.py b/funasr/models/fun_asr_nano/model.py
index d5d61cd37..9b27fb794 100644
--- a/funasr/models/fun_asr_nano/model.py
+++ b/funasr/models/fun_asr_nano/model.py
@@ -571,11 +571,11 @@ class FunASRNano(nn.Module):
encoder_out_lens = kwargs["audio_embedding_lens"]
else:
speech_lengths = batch["speech_lengths"][:, 0]
- # fp16
- if kwargs.get("fp16", False):
- speech = speech.to(torch.float16)
- elif kwargs.get("bf16", False):
- speech = speech.to(torch.bfloat16)
+ # NOTE: the audio encoder contains fp32-only ops, so casting its
+ # input to fp16/bf16 here raises a dtype mismatch. The encoder
+ # therefore always runs in fp32; low precision (fp16/bf16) is
+ # applied to the LLM decoder only. The audio embeddings are cast
+ # to the LLM dtype automatically when written into inputs_embeds.
# audio encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)