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

Fun-ASR-Nano vLLM inference produces degraded/garbage transcription in
fp16 due to numerical overflow in the audio embedding path (reproduced:
degenerate repeated tokens). bf16 and fp32 are correct. Add a clear
warning recommending bf16 (or fp32 on V100-class GPUs without bf16) so
users do not hit silent garbage output.

Refs #2976.

Co-authored-by: LauraGPT <LauraGPT@users.noreply.github.com>

Author: LauraGPT

funasr/models/fun_asr_nano/inference_vllm.py

@@ -175,6 +175,13 @@ def __init__(
         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)