Fix chat completion url for OpenAI compatibility

src/huggingface_hub/inference/_client.py

@@ -816,13 +816,16 @@ def chat_completion(
         # First, resolve the model chat completions URL
         if model == self.base_url:
             # base_url passed => add server route
-            model_url = model + "/v1/chat/completions"
+            model_url = model.rstrip("/")
+            if not model_url.endswith("/v1"):
+                model_url += "/v1"
+            model_url += "/chat/completions"
         elif is_url:
             # model is a URL => use it directly
             model_url = model
         else:
             # model is a model ID => resolve it + add server route
-            model_url = self._resolve_url(model) + "/v1/chat/completions"
+            model_url = self._resolve_url(model).rstrip("/") + "/v1/chat/completions"
 
         # `model` is sent in the payload. Not used by the server but can be useful for debugging/routing.
         # If it's a ID on the Hub => use it. Otherwise, we use a random string.

src/huggingface_hub/inference/_generated/_async_client.py

@@ -822,13 +822,16 @@ async def chat_completion(
         # First, resolve the model chat completions URL
         if model == self.base_url:
             # base_url passed => add server route
-            model_url = model + "/v1/chat/completions"
+            model_url = model.rstrip("/")
+            if not model_url.endswith("/v1"):
+                model_url += "/v1"
+            model_url += "/chat/completions"
         elif is_url:
             # model is a URL => use it directly
             model_url = model
         else:
             # model is a model ID => resolve it + add server route
-            model_url = self._resolve_url(model) + "/v1/chat/completions"
+            model_url = self._resolve_url(model).rstrip("/") + "/v1/chat/completions"
 
         # `model` is sent in the payload. Not used by the server but can be useful for debugging/routing.
         # If it's a ID on the Hub => use it. Otherwise, we use a random string.

tests/test_inference_client.py

@@ -916,6 +916,27 @@ def test_model_and_base_url_mutually_exclusive(self):
             InferenceClient(model="meta-llama/Meta-Llama-3-8B-Instruct", base_url="http://127.0.0.1:8000")
 
 
+@pytest.mark.parametrize(
+    "base_url",
+    [
+        "http://0.0.0.0:8080/v1",  # expected from OpenAI client
+        "http://0.0.0.0:8080",  # but not mandatory
+        "http://0.0.0.0:8080/v1/",  # ok with trailing '/' as well
+        "http://0.0.0.0:8080/",  # ok with trailing '/' as well
+    ],
+)
+def test_chat_completion_base_url_works_with_v1(base_url: str):
+    """Test that `/v1/chat/completions` is correctly appended to the base URL.
+
+    This is a regression test for https://github.com/huggingface/huggingface_hub/issues/2414
+    """
+    with patch("huggingface_hub.inference._client.InferenceClient.post") as post_mock:
+        client = InferenceClient(base_url=base_url)
+        post_mock.return_value = "{}"
+        client.chat_completion(messages=CHAT_COMPLETION_MESSAGES, stream=False)
+    assert post_mock.call_args_list[0].kwargs["model"] == "http://0.0.0.0:8080/v1/chat/completions"
+
+
 def test_stream_text_generation_response():
     data = [
         b'data: {"index":1,"token":{"id":4560,"text":" trying","logprob":-2.078125,"special":false},"generated_text":null,"details":null}',