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): warn when vLLM dtype=fp16 (degraded output) (#2980)

zhifu gao8107365f7e79
diff --git a/funasr/models/fun_asr_nano/inference_vllm.py b/funasr/models/fun_asr_nano/inference_vllm.py
index f83df875b..6c5763105 100644
--- a/funasr/models/fun_asr_nano/inference_vllm.py
+++ b/funasr/models/fun_asr_nano/inference_vllm.py
@@ -175,6 +175,13 @@ class FunASRNanoVLLM:
self.device = device
self.dtype = dtype
self.torch_dtype = dtype_map.get(dtype, torch.bfloat16)
+ if self.torch_dtype == torch.float16:
+ logger.warning(
+ "dtype='fp16' can produce degraded or garbage transcription for "
+ "Fun-ASR-Nano (numerical overflow in the audio embedding path). "
+ "Use dtype='bf16' (recommended) or dtype='fp32'. On GPUs without "
+ "bfloat16 support (e.g. NVIDIA V100), use 'fp32'."
+ )
self.model_dir = model_dir
# Step 1: Prepare LLM weights for vLLM (extract from model.pt if needed)