feat(openai_like): honour unsupported_params in JSON providers, declare Sail's

litellm/llms/openai_like/dynamic_config.py

@@ -94,7 +94,11 @@ def create_config_class(provider: SimpleProviderConfig):
             that don't support function calling."""
             from litellm.utils import supports_function_calling, supports_reasoning
 
-            supported_params: Final = super().get_supported_openai_params(model=model)
+            supported_params: Final = [
+                param
+                for param in super().get_supported_openai_params(model=model)
+                if param not in provider.unsupported_params
+            ]
 
             _supports_fc: Final = supports_function_calling(model=model, custom_llm_provider=provider.slug)
 

litellm/llms/openai_like/json_loader.py

@@ -22,6 +22,7 @@ class SimpleProviderConfig:
         self.constraints = data.get("constraints", {})
         self.special_handling = data.get("special_handling", {})
         self.supported_endpoints = data.get("supported_endpoints", [])
+        self.unsupported_params: Final = tuple(data.get("unsupported_params", ()))
 
 
 class JSONProviderRegistry:

litellm/llms/openai_like/providers.json

@@ -208,6 +208,7 @@
     "supported_endpoints": ["/v1/chat/completions", "/v1/responses", "/v1/messages"],
     "param_mappings": {
       "max_tokens": "max_completion_tokens"
-    }
+    },
+    "unsupported_params": ["stop", "seed", "frequency_penalty", "presence_penalty", "logit_bias", "logprobs", "top_logprobs"]
   }
 }

tests/test_litellm/llms/openai_like/test_sail_provider.py

@@ -250,6 +250,39 @@ class TestSailRequestShape:
 
         assert not route.called
 
+    @pytest.mark.respx()
+    def test_sail_unsupported_params_dropped_with_drop_params(self, respx_mock: respx.Router):
+        respx_mock.post(SAIL_CHAT_COMPLETIONS).respond(json=_chat_completion_payload())
+
+        litellm.completion(
+            model=MODEL,
+            messages=[{"role": "user", "content": "hi"}],
+            stop=["x"],
+            seed=1,
+            frequency_penalty=0.5,
+            drop_params=True,
+        )
+
+        body = json.loads(respx_mock.calls[0].request.content)
+        assert "stop" not in body
+        assert "seed" not in body
+        assert "frequency_penalty" not in body
+        assert body["model"] == "zai-org/GLM-5.3"
+        assert body["messages"] == [{"role": "user", "content": "hi"}]
+
+    @pytest.mark.respx(assert_all_called=False)
+    def test_sail_unsupported_params_raise_without_drop_params(self, respx_mock: respx.Router):
+        route = respx_mock.post(SAIL_CHAT_COMPLETIONS)
+
+        with pytest.raises(litellm.UnsupportedParamsError):
+            litellm.completion(
+                model=MODEL,
+                messages=[{"role": "user", "content": "hi"}],
+                stop=["x"],
+            )
+
+        assert not route.called
+
     @pytest.mark.asyncio
     @pytest.mark.respx()
     async def test_sail_anthropic_messages_posts_to_messages_endpoint(self, respx_mock: respx.Router):