fix: add repetition_penalty=1.3 to all vLLM services (fixes #2930 text repetition)

- funasr-server (_server_app.py): add repetition_penalty=1.3 to generate()
- serve_vllm.py: add repetition_penalty=1.3 to gen_kwargs
- Bump version to 1.3.6

Author: LauraGPT

examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py

@@ -107,7 +107,7 @@ def process_audio(audio_data, sr=16000, language=None, hotwords=None,
         return {"text": "", "segments": [], "duration": len(audio_data) / sr}
 
     # vLLM batch ASR
-    gen_kwargs = {"max_new_tokens": 500}
+    gen_kwargs = {"max_new_tokens": 500, "repetition_penalty": 1.3}
     if language:
         gen_kwargs["language"] = language
     if hotwords:

funasr/bin/_server_app.py

@@ -86,6 +86,7 @@ async def transcribe(
             kwargs = {"input": tmp_path, "batch_size": 1}
             if language:
                 kwargs["language"] = language
+            kwargs["repetition_penalty"] = 1.3
             result = asr.generate(**kwargs)
             text = re.sub(r'<\|[^|]*\|>', '', result[0]["text"]).strip()
             if response_format == "verbose_json":

funasr/version.txt

@@ -1 +1 @@
-1.3.5
+1.3.6