diff --git a/litellm/llms/oci/chat/transformation.py b/litellm/llms/oci/chat/transformation.py index d487d223bf4..99baa497ac0 100644 --- a/litellm/llms/oci/chat/transformation.py +++ b/litellm/llms/oci/chat/transformation.py @@ -357,7 +357,13 @@ class OCIChatConfig(BaseConfig): for key, value in {**non_default_params, **optional_params}.items(): alias = param_map.get(key) - if alias is False: +if alias is False: + # max_retries is a litellm-level control param (litellm applies + # retries itself); it is never a generation param OCI accepts, so + # drop it silently. The litellm proxy injects it on every request, + # which otherwise 500s OCI calls unless drop_params is set. + if key == "max_retries": + continue # n=1 (or None) is the OpenAI default: a single generation, which # every OCI model produces anyway. Drop it silently so standard # clients that always send n=1 (e.g. the MLflow gateway) are not diff --git a/tests/integration/test_oci_proxy_integration.py b/tests/integration/test_oci_proxy_integration.py index b2ad4eb844c..f41e4826a04 100644 --- a/tests/integration/test_oci_proxy_integration.py +++ b/tests/integration/test_oci_proxy_integration.py @@ -108,38 +108,35 @@ def _oci_env_from_profile() -> dict[str, str]: } -@pytest.fixture(scope="module") -def proxy_url() -> Iterator[str]: - oci_env = _oci_env_from_profile() - - port = _free_port() - base_url = f"http://127.0.0.1:{port}" - +def _serve(config_path: str) -> Iterator[str]: + """Boot the litellm proxy with the given config and yield its base URL.""" env = os.environ.copy() - env.update(oci_env) + env.update(_oci_env_from_profile()) # Avoid pulling in DB-backed features for this lightweight smoke run. env.pop("DATABASE_URL", None) env["STORE_MODEL_IN_DB"] = "False" + port = _free_port() + base_url = f"http://127.0.0.1:{port}" + # Prefer the `litellm` console script that lives next to the active # Python so we inherit the test virtualenv. Fall back to PATH. cli = Path(sys.executable).parent / "litellm" if not cli.exists(): cli = "litellm" - cmd = [ - str(cli), - "--config", - str(CONFIG_PATH), - "--port", - str(port), - "--host", - "127.0.0.1", - "--num_workers", - "1", - ] proc = subprocess.Popen( - cmd, + [ + str(cli), + "--config", + config_path, + "--port", + str(port), + "--host", + "127.0.0.1", + "--num_workers", + "1", + ], env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, @@ -157,6 +154,27 @@ def proxy_url() -> Iterator[str]: proc.wait(timeout=5) +@pytest.fixture(scope="module") +def proxy_url() -> Iterator[str]: + yield from _serve(str(CONFIG_PATH)) + + +@pytest.fixture(scope="module") +def proxy_url_no_drop_params(tmp_path_factory) -> Iterator[str]: + """A proxy WITHOUT drop_params, to prove benign params the proxy injects + (e.g. max_retries) don't break OCI calls.""" + cfg = tmp_path_factory.mktemp("oci_nodrop") / "config.yaml" + cfg.write_text( + "model_list:\n" + " - model_name: oci-cohere-command\n" + " litellm_params:\n" + " model: oci/cohere.command-latest\n" + "general_settings:\n" + f" master_key: {MASTER_KEY}\n" + ) + yield from _serve(str(cfg)) + + # --------------------------------------------------------------------------- # Helpers # --------------------------------------------------------------------------- @@ -273,6 +291,25 @@ def test_model_list_advertises_oci_models(proxy_url: str) -> None: assert expected in advertised, f"{expected} missing from /v1/models: {advertised}" +def test_chat_completion_no_drop_params(proxy_url_no_drop_params: str) -> None: + """A plain chat completion succeeds through a proxy without drop_params. + + Regression for the HTTP 500 ``param `max_retries` is not supported on OCI``: + the proxy injects max_retries on every request, so without this fix any OCI + call through the proxy failed unless drop_params was set. + """ + r = httpx.post( + f"{proxy_url_no_drop_params}/v1/chat/completions", + headers=_auth_headers(), + json=_chat_payload("oci-cohere-command"), + timeout=REQUEST_TIMEOUT_S, + ) + assert r.status_code == 200, f"no-drop_params -> {r.status_code}: {r.text}" + body = r.json() + assert body["object"] == "chat.completion" + assert body["choices"][0]["message"].get("content") is not None + + def test_cohere_default_n_via_proxy(proxy_url: str) -> None: """A Cohere request carrying the default n=1 succeeds through the gateway. diff --git a/tests/test_litellm/llms/oci/chat/test_oci_chat_transformation.py b/tests/test_litellm/llms/oci/chat/test_oci_chat_transformation.py index 49f9dfa416e..53c9e4b207c 100644 --- a/tests/test_litellm/llms/oci/chat/test_oci_chat_transformation.py +++ b/tests/test_litellm/llms/oci/chat/test_oci_chat_transformation.py @@ -1304,8 +1304,7 @@ class TestOCIStreamingSignedBody: When signed_json_body is provided, the POST must use that exact bytes object, not json.dumps(data) — otherwise the RSA-SHA256 signature is invalid. """ - import httpx - from unittest.mock import MagicMock, patch + from unittest.mock import MagicMock config = OCIChatConfig() signed_bytes = b'{"signed": true}' @@ -1464,6 +1463,21 @@ class TestOCIChatConfigErrorPaths: ) assert "audio" not in result + @pytest.mark.parametrize("model", ["cohere.command-latest", "xai.grok-4"]) + def test_map_openai_params_max_retries_dropped_without_drop_params(self, model): + """max_retries is a litellm control param, not a generation param. It + must be dropped silently (no raise) even when drop_params is False, so + the litellm proxy (which injects max_retries on every request) does not + 500 every OCI call. + """ + config = OCIChatConfig() + result = config.map_openai_params( + non_default_params={"max_retries": 3}, + optional_params={}, + model=model, + drop_params=False, + ) + assert "max_retries" not in result def test_map_openai_params_cohere_n_default_dropped(self): """Cohere has no numGenerations field, but n=1 (and None) is the OpenAI default single-generation request. It must be dropped silently rather