fix(hosted-vllm): preserve explicit transcription content-type

litellm/llms/hosted_vllm/transcriptions/transformation.py

@@ -56,6 +56,10 @@ class HostedVLLMAudioTranscriptionConfig(OpenAIWhisperAudioTranscriptionConfig):
         litellm_params: dict,
     ) -> AudioTranscriptionRequestData:
         processed_audio = process_audio_file(audio_file)
+        explicit_content_type = (
+            audio_file[2] if isinstance(audio_file, tuple) and len(audio_file) >= 3 and audio_file[2] else None
+        )
+        content_type = explicit_content_type or processed_audio.content_type
         extra_body = optional_params.get("extra_body") or {}
         data = {
             "model": model,
@@ -66,7 +70,7 @@ class HostedVLLMAudioTranscriptionConfig(OpenAIWhisperAudioTranscriptionConfig):
             "file": (
                 processed_audio.filename,
                 processed_audio.file_content,
-                processed_audio.content_type,
+                content_type,
             )
         }
 

tests/test_litellm/llms/hosted_vllm/transcriptions/test_hosted_vllm_audio_transcription.py

@@ -4,6 +4,9 @@ import httpx
 import pytest
 
 import litellm
+from litellm.llms.hosted_vllm.transcriptions.transformation import (
+    HostedVLLMAudioTranscriptionConfig,
+)
 
 
 def _transcription_response() -> httpx.Response:
@@ -57,3 +60,30 @@ async def test_transcription_passes_custom_ca_to_async_http_client() -> None:
     request = client.post.call_args.kwargs
     assert request["data"] == {"model": "whisper-1"}
     assert request["files"] == {"file": ("audio.wav", b"audio", "audio/wav")}
+
+
+def test_transform_request_preserves_explicit_content_type() -> None:
+    config = HostedVLLMAudioTranscriptionConfig()
+
+    request_data = config.transform_audio_transcription_request(
+        model="whisper-1",
+        audio_file=("recording", b"audio", "audio/ogg"),
+        optional_params={"language": "en", "extra_body": {"temperature": 0.1}},
+        litellm_params={},
+    )
+
+    assert request_data.data == {"model": "whisper-1", "language": "en", "temperature": 0.1}
+    assert request_data.files == {"file": ("recording", b"audio", "audio/ogg")}
+
+
+def test_transform_request_derives_content_type_when_not_supplied() -> None:
+    config = HostedVLLMAudioTranscriptionConfig()
+
+    request_data = config.transform_audio_transcription_request(
+        model="whisper-1",
+        audio_file=("audio.mp3", b"audio"),
+        optional_params={},
+        litellm_params={},
+    )
+
+    assert request_data.files == {"file": ("audio.mp3", b"audio", "audio/mpeg")}