From e95ae2f0652dc1ee4e72aa71613dcdc102a924db Mon Sep 17 00:00:00 2001 From: ryan-crabbe-berri Date: Tue, 27 Oct 2026 10:16:20 -0700 Subject: [PATCH] fix(huggingface): pass the caller's timeout to the embedding model lookup too Without input_type the handler first fetches the model's pipeline tag from huggingface.co, and that GET still used the default timeout. Thread the caller's timeout through both lookup helpers and the transform step so every request the embedding makes honors it. Refs #47299 --- litellm/llms/huggingface/embedding/handler.py | 32 ++++++++++++++----- .../test_huggingface_embedding_handler.py | 21 ++++++------ 2 files changed, 36 insertions(+), 17 deletions(-) diff --git a/litellm/llms/huggingface/embedding/handler.py b/litellm/llms/huggingface/embedding/handler.py index 813615ff06f..11fb74abcf2 100644 --- a/litellm/llms/huggingface/embedding/handler.py +++ b/litellm/llms/huggingface/embedding/handler.py @@ -38,7 +38,9 @@ class _SupportsTokenEncode(Protocol): def encode(self, text: str) -> list[int]: ... -def get_hf_task_embedding_for_model(model: str, task_type: str | None, api_base: str) -> object: +def get_hf_task_embedding_for_model( + model: str, task_type: str | None, api_base: str, timeout: float | httpx.Timeout | None = None +) -> object: if task_type is not None: if task_type in get_args(hf_tasks_embeddings): return task_type @@ -46,7 +48,7 @@ def get_hf_task_embedding_for_model(model: str, task_type: str | None, api_base: raise Exception(f"Invalid task_type={task_type}. Expected one of={hf_tasks_embeddings}") http_client: Final = HTTPHandler(concurrent_limit=1) - model_info: Final = http_client.get(url=f"{api_base}/api/models/{model}") + model_info: Final = http_client.get(url=f"{api_base}/api/models/{model}", timeout=timeout) model_info_dict: Final = decoded_json_object(_DECODED_JSON.validate_python(model_info.json())) @@ -55,7 +57,9 @@ def get_hf_task_embedding_for_model(model: str, task_type: str | None, api_base: return pipeline_tag -async def async_get_hf_task_embedding_for_model(model: str, task_type: str | None, api_base: str) -> object: +async def async_get_hf_task_embedding_for_model( + model: str, task_type: str | None, api_base: str, timeout: float | httpx.Timeout | None = None +) -> object: if task_type is not None: if task_type in get_args(hf_tasks_embeddings): return task_type @@ -65,7 +69,7 @@ async def async_get_hf_task_embedding_for_model(model: str, task_type: str | Non llm_provider=litellm.LlmProviders.HUGGINGFACE, ) - model_info: Final = await http_client.get(url=f"{api_base}/api/models/{model}") + model_info: Final = await http_client.get(url=f"{api_base}/api/models/{model}", timeout=timeout) model_info_dict: Final = decoded_json_object(_DECODED_JSON.validate_python(model_info.json())) @@ -107,8 +111,11 @@ class HuggingFaceEmbedding(BaseLLM): embed_url: str, input: list, optional_params: dict, + timeout: float | httpx.Timeout | None = None, ) -> dict: - hf_task = await async_get_hf_task_embedding_for_model(model=model, task_type=task_type, api_base=HF_HUB_URL) + hf_task = await async_get_hf_task_embedding_for_model( + model=model, task_type=task_type, api_base=HF_HUB_URL, timeout=timeout + ) data: Final = self._transform_input_on_pipeline_tag(input=input, pipeline_tag=hf_task) @@ -148,6 +155,7 @@ class HuggingFaceEmbedding(BaseLLM): call_type: Literal["sync", "async"], optional_params: dict, embed_url: str, + timeout: float | httpx.Timeout | None = None, ) -> dict: data: dict = {} @@ -165,9 +173,13 @@ class HuggingFaceEmbedding(BaseLLM): task_type: Final = optional_params.pop("input_type", None) if call_type == "sync": - hf_task: Final = get_hf_task_embedding_for_model(model=model, task_type=task_type, api_base=HF_HUB_URL) + hf_task: Final = get_hf_task_embedding_for_model( + model=model, task_type=task_type, api_base=HF_HUB_URL, timeout=timeout + ) elif call_type == "async": - return self._async_transform_input(model=model, task_type=task_type, embed_url=embed_url, input=input) + return self._async_transform_input( + model=model, task_type=task_type, embed_url=embed_url, input=input, timeout=timeout + ) data = self._transform_input_on_pipeline_tag(input=input, pipeline_tag=hf_task) @@ -253,6 +265,7 @@ class HuggingFaceEmbedding(BaseLLM): call_type="sync", optional_params=optional_params, embed_url=api_base, + timeout=timeout, ) ## LOGGING @@ -322,7 +335,9 @@ class HuggingFaceEmbedding(BaseLLM): litellm_params=litellm_params, ) task_type: Final = optional_params.get("input_type", None) - task: Final = get_hf_task_embedding_for_model(model=model, task_type=task_type, api_base=HF_HUB_URL) + task: Final = get_hf_task_embedding_for_model( + model=model, task_type=task_type, api_base=HF_HUB_URL, timeout=timeout + ) # print_verbose(f"{model}, {task}") embed_url = "" if model.startswith(("http://", "https://")): @@ -360,6 +375,7 @@ class HuggingFaceEmbedding(BaseLLM): call_type="sync", optional_params=optional_params, embed_url=embed_url, + timeout=timeout, ) ## LOGGING diff --git a/tests/unit/llms/huggingface/embedding/test_huggingface_embedding_handler.py b/tests/unit/llms/huggingface/embedding/test_huggingface_embedding_handler.py index 21a2b4519c2..7040fbe5fcc 100644 --- a/tests/unit/llms/huggingface/embedding/test_huggingface_embedding_handler.py +++ b/tests/unit/llms/huggingface/embedding/test_huggingface_embedding_handler.py @@ -52,12 +52,10 @@ def mock_embedding_async_http_handler(): class TestHuggingFaceEmbedding: @pytest.fixture(autouse=True) def setup(self, mock_embedding_http_handler, mock_embedding_async_http_handler, mock_hf_config_fetch): - self.mock_get_task_patcher = patch( - "litellm.llms.huggingface.embedding.handler.get_hf_task_embedding_for_model" - ) + self.mock_get_task_patcher = patch("litellm.llms.huggingface.embedding.handler.get_hf_task_embedding_for_model") self.mock_get_task = self.mock_get_task_patcher.start() - def mock_get_task_side_effect(model, task_type, api_base): + def mock_get_task_side_effect(model, task_type, api_base, timeout=None): if task_type is not None: return task_type return "sentence-similarity" @@ -180,7 +178,9 @@ async def test_hf_embedding_sentence_sim_sends_one_similarity_request( monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) monkeypatch.setenv("HUGGINGFACE_API_KEY", "hf-test-token") with respx.mock(assert_all_called=False) as hf: - hf.get(url__regex=r"https://huggingface\.co/api/models/.*").respond(json={"pipeline_tag": "sentence-similarity"}) + hf.get(url__regex=r"https://huggingface\.co/api/models/.*").respond( + json={"pipeline_tag": "sentence-similarity"} + ) hf.get(url__regex=r"https://huggingface\.co/.*/config\.json").respond(json={"max_position_embeddings": 512}) inference: Final = hf.post( "https://router.huggingface.co/hf-inference/pipeline/sentence-similarity/sentence-transformers/TaylorAI/bge-micro-v2" @@ -199,7 +199,8 @@ async def test_hf_embedding_sentence_sim_sends_one_similarity_request( assert response.data == [{"object": "embedding", "index": 0, "embedding": 0.7708950042724609}] -HF_EMBED_URL: Final = "https://router.huggingface.co/hf-inference/pipeline/embed/BAAI/bge-m3" +HF_MODEL_LOOKUP_URL: Final = "https://huggingface.co/api/models/BAAI/bge-m3" +HF_EMBED_URL: Final = "https://router.huggingface.co/hf-inference/pipeline/feature-extraction/BAAI/bge-m3" @pytest.mark.respx(assert_all_called=False) @@ -209,18 +210,20 @@ async def test_huggingface_embedding_sends_the_caller_timeout( ) -> None: monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) respx_mock.get(url__regex=r"https://huggingface\.co/.*/config\.json").respond(json={"max_position_embeddings": 512}) - route: Final = respx_mock.post(HF_EMBED_URL).mock( - return_value=httpx.Response(200, json=MOCK_EMBEDDING_RESPONSE) + lookup: Final = respx_mock.get(HF_MODEL_LOOKUP_URL).mock( + return_value=httpx.Response(200, json={"pipeline_tag": "feature-extraction"}) ) + route: Final = respx_mock.post(HF_EMBED_URL).mock(return_value=httpx.Response(200, json=MOCK_EMBEDDING_RESPONSE)) request: Final = { "model": "huggingface/BAAI/bge-m3", "input": ["hello"], - "input_type": "embed", "api_key": "hf_test", "timeout": 45, } response: Final = litellm.embedding(**request) if sync_mode else await litellm.aembedding(**request) + assert all(call.request.extensions["timeout"]["read"] == 45 for call in lookup.calls) + assert lookup.called assert route.calls.last.request.extensions["timeout"]["read"] == 45 assert response.data[0]["embedding"] == MOCK_EMBEDDING_RESPONSE[0]