Skip to content

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

fp16

游雁63e60cc43dda
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 7b5a02fff..22b1ac0ec 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -233,6 +233,8 @@ class AutoModel:
# fp16
if kwargs.get("fp16", False):
model.to(torch.float16)
+ elif kwargs.get("bf16", False):
+ model.to(torch.bfloat16)
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 5f1530992..f72b2c814 100644
--- a/funasr/models/llm_asr/model.py
+++ b/funasr/models/llm_asr/model.py
@@ -684,6 +684,13 @@ class LLMASR2(nn.Module):
# audio encoder
speech = batch["speech"]
speech_lengths = batch["speech_lengths"][:, 0]
+ # 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