fix(hosted_vllm): honor custom_llm_provider and /v1 api_base on transcriptions

atranscription/aspeech ignored custom_llm_provider, so unprefixed proxy
models failed with LLM Provider NOT provided. Hosted vLLM URL joining
also doubled /v1 when api_base already ended with it.

Co-authored-by: Cursor <cursoragent@cursor.com>

litellm/llms/hosted_vllm/transcriptions/transformation.py

@@ -39,13 +39,19 @@ class HostedVLLMAudioTranscriptionConfig(OpenAIWhisperAudioTranscriptionConfig):
         litellm_params: dict,
         stream: bool | None = None,
     ) -> str:
-        if api_base:
-            # Remove trailing slashes and ensure clean base URL
-            api_base = api_base.rstrip("/")
-            if not api_base.endswith("/v1/audio/transcriptions"):
-                api_base = f"{api_base}/v1/audio/transcriptions"
-            return api_base
-        raise ValueError("api_base must be provided for Hosted VLLM rerank")
+        if not api_base:
+            raise ValueError("api_base must be provided for Hosted VLLM transcriptions")
+        normalized_base: Final = api_base.rstrip("/")
+        if normalized_base.endswith("/v1/audio/transcriptions"):
+            return normalized_base
+        return super().get_complete_url(
+            api_base=normalized_base,
+            api_key=api_key,
+            model=model,
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            stream=stream,
+        )
 
     def transform_audio_transcription_request(
         self,

litellm/main.py

@@ -7541,7 +7541,7 @@ async def atranscription(*args, **kwargs) -> TranscriptionResponse:
     ### PASS ARGS TO Image Generation ###
     kwargs["atranscription"] = True
     file: Final = kwargs.get("file", None)
-    custom_llm_provider = None
+    custom_llm_provider = kwargs.get("custom_llm_provider", None)
     try:
         # Use a partial function to pass your keyword arguments
         func: Final = partial(transcription, *args, **kwargs)
@@ -7550,7 +7550,11 @@ async def atranscription(*args, **kwargs) -> TranscriptionResponse:
         ctx: Final = contextvars.copy_context()
         func_with_context: Final = partial(ctx.run, func)
 
-        _, custom_llm_provider, _, _ = get_llm_provider(model=model, api_base=kwargs.get("api_base", None))
+        _, custom_llm_provider, _, _ = get_llm_provider(
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            api_base=kwargs.get("api_base", None),
+        )
 
         # Await normally
         init_response: Final = await loop.run_in_executor(None, func_with_context)
@@ -7869,7 +7873,11 @@ async def aspeech(*args, **kwargs) -> HttpxBinaryResponseContent:
         ctx: Final = contextvars.copy_context()
         func_with_context: Final = partial(ctx.run, func)
 
-        _, custom_llm_provider, _, _ = get_llm_provider(model=model, api_base=kwargs.get("api_base", None))
+        _, custom_llm_provider, _, _ = get_llm_provider(
+            model=model,
+            custom_llm_provider=custom_llm_provider,
+            api_base=kwargs.get("api_base", None),
+        )
 
         # Await normally
         init_response: Final = await loop.run_in_executor(None, func_with_context)

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

@@ -0,0 +1,107 @@
+import os
+import sys
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+sys.path.insert(0, os.path.abspath("../../../../../.."))
+
+import litellm
+from litellm.llms.hosted_vllm.transcriptions.transformation import (
+    HostedVLLMAudioTranscriptionConfig,
+)
+from litellm.types.utils import TranscriptionResponse
+
+
+def _complete_url(api_base: str | None) -> str:
+    return HostedVLLMAudioTranscriptionConfig().get_complete_url(
+        api_base=api_base,
+        api_key=None,
+        model="whisper-1",
+        optional_params={},
+        litellm_params={},
+    )
+
+
+class TestHostedVLLMTranscriptionUrl:
+    @pytest.mark.parametrize(
+        "api_base,expected",
+        [
+            (
+                "http://vllm.example.com",
+                "http://vllm.example.com/v1/audio/transcriptions",
+            ),
+            (
+                "http://vllm.example.com/",
+                "http://vllm.example.com/v1/audio/transcriptions",
+            ),
+            (
+                "http://vllm.example.com/v1",
+                "http://vllm.example.com/v1/audio/transcriptions",
+            ),
+            (
+                "http://vllm.example.com/v1/",
+                "http://vllm.example.com/v1/audio/transcriptions",
+            ),
+            (
+                "http://proxy.example.com/qwen3-asr/v1",
+                "http://proxy.example.com/qwen3-asr/v1/audio/transcriptions",
+            ),
+            (
+                "http://vllm.example.com/v1/audio/transcriptions",
+                "http://vllm.example.com/v1/audio/transcriptions",
+            ),
+        ],
+    )
+    def test_get_complete_url(self, api_base: str, expected: str) -> None:
+        assert _complete_url(api_base) == expected
+
+    def test_get_complete_url_requires_api_base(self) -> None:
+        with pytest.raises(ValueError, match="api_base must be provided"):
+            _complete_url(None)
+
+
+class TestAtranscriptionCustomLlmProvider:
+    @pytest.mark.asyncio
+    async def test_unprefixed_model_uses_custom_llm_provider(self) -> None:
+        """
+        Proxy/router deployments often register model=qwen3-asr-0.6b with
+        custom_llm_provider=hosted_vllm (no hosted_vllm/ prefix). Chat honors
+        that field; atranscription used to ignore it and raise
+        'LLM Provider NOT provided'.
+        """
+        with patch(
+            "litellm.main.transcription",
+            return_value=TranscriptionResponse(text="hello"),
+        ) as mock_transcription:
+            response = await litellm.atranscription(
+                model="qwen3-asr-0.6b",
+                file=b"fake-audio",
+                custom_llm_provider="hosted_vllm",
+                api_base="http://vllm.example.com/v1",
+            )
+
+        assert response.text == "hello"
+        mock_transcription.assert_called()
+
+    @pytest.mark.asyncio
+    async def test_unprefixed_model_without_provider_still_fails(self) -> None:
+        with pytest.raises(Exception, match="LLM Provider NOT provided"):
+            await litellm.atranscription(
+                model="qwen3-asr-0.6b",
+                file=b"fake-audio",
+                api_base="http://vllm.example.com/v1",
+            )
+
+    @pytest.mark.asyncio
+    async def test_aspeech_unprefixed_model_uses_custom_llm_provider(self) -> None:
+        with patch("litellm.main.speech", return_value=MagicMock()) as mock_speech:
+            await litellm.aspeech(
+                model="tts-model",
+                input="hello",
+                voice="alloy",
+                custom_llm_provider="hosted_vllm",
+                api_base="http://vllm.example.com/v1",
+            )
+
+        mock_speech.assert_called()