Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
fix(vllm): honor fp32 decoder fallback
LauraGPT63d7098e7045
diff --git a/funasr/models/fun_asr_nano/inference_vllm.py b/funasr/models/fun_asr_nano/inference_vllm.pyindex 3da3eecf1..331ba90cf 100644--- a/funasr/models/fun_asr_nano/inference_vllm.py+++ b/funasr/models/fun_asr_nano/inference_vllm.py@@ -47,10 +47,11 @@ def _resolve_vllm_dtype(dtype: str) -> str:"Fun-ASR-Nano's Qwen3 language model is numerically unstable in ""float16; vLLM will use bfloat16 while audio components remain in ""float16. This keeps the same memory footprint and avoids degraded "- "or repetitive transcription."+ "or repetitive transcription. On GPUs without bfloat16 support, "+ "use dtype='fp32' so both audio and vLLM compute use float32.")return "bfloat16"- return {"bf16": "bfloat16", "fp32": "auto"}.get(dtype, dtype)+ return {"bf16": "bfloat16", "fp32": "float32"}.get(dtype, dtype)def prepare_vllm_model_dir(model_dir: str, output_dir: str = None) -> str:diff --git a/tests/test_fun_asr_nano_vllm_dtype.py b/tests/test_fun_asr_nano_vllm_dtype.pyindex 48403ad85..0c83fd096 100644--- a/tests/test_fun_asr_nano_vllm_dtype.py+++ b/tests/test_fun_asr_nano_vllm_dtype.py@@ -67,12 +67,13 @@ def test_fp16_keeps_audio_compute_but_promotes_vllm_to_bf16(monkeypatch, caplog)assert streaming.torch_dtype is torch.float16assert [call["dtype"] for call in _FakeLLM.calls] == ["bfloat16", "bfloat16"]assert "audio components remain in float16" in caplog.text+ assert "dtype='fp32'" in caplog.textdef test_vllm_dtype_mapping_preserves_supported_values(caplog):with caplog.at_level(logging.WARNING):assert inference_vllm._resolve_vllm_dtype("bf16") == "bfloat16"- assert inference_vllm._resolve_vllm_dtype("fp32") == "auto"+ assert inference_vllm._resolve_vllm_dtype("fp32") == "float32"assert inference_vllm._resolve_vllm_dtype("custom") == "custom"assert not caplog.text