[Frontend] Set MAX_AUDIO_CLIP_FILESIZE_MB via env var instead of hardcoding (#21374)

Signed-off-by: Deven Labovitch <[email protected]>

Author: deven-labovitch

docs/serving/openai_compatible_server.md

@@ -351,6 +351,11 @@ you can use the [official OpenAI Python client](https://github.com/openai/openai
 Code example: <gh-file:examples/online_serving/openai_transcription_client.py>
 <!-- TODO: api enforced limits + uploading audios -->
 
+#### API Enforced Limits
+
+Set the maximum audio file size (in MB) that VLLM will accept, via the
+`VLLM_MAX_AUDIO_CLIP_FILESIZE_MB` environment variable. Default is 25 MB.
+
 #### Extra Parameters
 
 The following [sampling parameters][sampling-params] are supported.

vllm/entrypoints/openai/speech_to_text.py

@@ -11,6 +11,7 @@
 import numpy as np
 from fastapi import Request
 
+import vllm.envs as envs
 from vllm.config import ModelConfig
 from vllm.engine.protocol import EngineClient
 from vllm.entrypoints.logger import RequestLogger
@@ -38,10 +39,6 @@
 
 logger = init_logger(__name__)
 
-# As per https://platform.openai.com/docs/guides/speech-to-text#overview.
-# TODO configurable
-MAX_AUDIO_CLIP_FILESIZE_MB = 25
-
 
 class OpenAISpeechToText(OpenAIServing):
     """Base class for speech-to-text operations like transcription and 
@@ -70,6 +67,8 @@ def __init__(
         self.asr_config = self.model_cls.get_speech_to_text_config(
             model_config, task_type)
 
+        self.max_audio_filesize_mb = envs.VLLM_MAX_AUDIO_CLIP_FILESIZE_MB
+
         if self.default_sampling_params:
             logger.info(
                 "Overwriting default completion sampling param with: %s",
@@ -93,7 +92,7 @@ async def _preprocess_speech_to_text(
         lang = request.language or "en"
         self.model_cls.validate_language(lang)
 
-        if len(audio_data) / 1024**2 > MAX_AUDIO_CLIP_FILESIZE_MB:
+        if len(audio_data) / 1024**2 > self.max_audio_filesize_mb:
             raise ValueError("Maximum file size exceeded.")
 
         with io.BytesIO(audio_data) as bytes_:

vllm/envs.py

@@ -61,6 +61,7 @@
     VLLM_IMAGE_FETCH_TIMEOUT: int = 5
     VLLM_VIDEO_FETCH_TIMEOUT: int = 30
     VLLM_AUDIO_FETCH_TIMEOUT: int = 10
+    VLLM_MAX_AUDIO_CLIP_FILESIZE_MB: int = 25
     VLLM_VIDEO_LOADER_BACKEND: str = "opencv"
     VLLM_MM_INPUT_CACHE_GIB: int = 8
     VLLM_TARGET_DEVICE: str = "cuda"
@@ -519,6 +520,12 @@ def get_vllm_port() -> Optional[int]:
     "VLLM_AUDIO_FETCH_TIMEOUT":
     lambda: int(os.getenv("VLLM_AUDIO_FETCH_TIMEOUT", "10")),
 
+    # Maximum filesize in MB for a single audio file when processing
+    # speech-to-text requests. Files larger than this will be rejected.
+    # Default is 25 MB
+    "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB":
+    lambda: int(os.getenv("VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", "25")),
+
     # Backend for Video IO
     # - "opencv": Default backend that uses OpenCV stream buffered backend.
     #