fix: close audio health-check file after transcription

litellm/litellm_core_utils/health_check_helpers.py

@@ -193,6 +193,13 @@ class HealthCheckHelpers:
         from litellm.litellm_core_utils.health_check_utils import _filter_model_params
         from litellm.realtime_api.main import _realtime_health_check
 
+        async def audio_transcription_health_check():
+            with get_audio_file_for_health_check() as audio_file:
+                return await litellm.atranscription(
+                    **_filter_model_params(model_params=model_params),
+                    file=audio_file,
+                )
+
         return {
             "chat": lambda: litellm.acompletion(
                 **model_params,
@@ -212,10 +219,7 @@ class HealthCheckHelpers:
                 },
                 input=prompt or "test",
             ),
-            "audio_transcription": lambda: litellm.atranscription(
-                **_filter_model_params(model_params=model_params),
-                file=get_audio_file_for_health_check(),
-            ),
+            "audio_transcription": audio_transcription_health_check,
             "image_generation": lambda: litellm.aimage_generation(
                 **_filter_model_params(model_params=model_params),
                 prompt=prompt,

tests/unit/litellm_core_utils/test_health_check_helpers.py

@@ -37,6 +37,36 @@ def _distinct_rgb_colors(png: bytes) -> set[bytes]:
     return {bytes(row[i : i + 3]) for row in rows for i in range(1, row_size, 3)}
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("raises", [False, True])
+async def test_audio_transcription_health_check_closes_file(raises: bool):
+    file_handles = []
+
+    async def transcription(**kwargs):
+        file = kwargs["file"]
+        file_handles.append(file)
+        assert not file.closed
+        assert file.read(4) == b"RIFF"
+        if raises:
+            raise RuntimeError("transcription failed")
+        return {"text": "healthy"}
+
+    handlers = HealthCheckHelpers.get_mode_handlers(
+        model="whisper-1",
+        custom_llm_provider="openai",
+        model_params={"model": "openai/whisper-1", "api_key": "sk-test"},
+    )
+    with patch("litellm.atranscription", side_effect=transcription):
+        if raises:
+            with pytest.raises(RuntimeError, match="transcription failed"):
+                await handlers["audio_transcription"]()
+        else:
+            assert await handlers["audio_transcription"]() == {"text": "healthy"}
+
+    assert len(file_handles) == 1
+    assert file_handles[0].closed
+
+
 @pytest.mark.asyncio
 async def test_image_edit_health_check_handler_uses_descriptive_prompt_and_multicolor_png():
     model_params = {"model": "openai/gpt-image-1", "api_key": "sk-test"}