feat: add ability to specify params to drop

Some LLMs don't support all params that clients could send. This MR gives the administrator the ability to filter those out

Similar to https://docs.litellm.ai/docs/completion/drop_params

Signed-off-by: Max Wittig <[email protected]>

Author: max-wittig

src/vllm_router/app.py

@@ -199,6 +199,8 @@ def initialize_all(app: FastAPI, args):
     if args.callbacks:
         configure_custom_callbacks(args.callbacks, app)
 
+    app.state.drop_params = parse_comma_separated_args(args.drop_params)
+
     initialize_routing_logic(
         args.routing_logic,
         session_key=args.session_key,

src/vllm_router/parsers/parser.py

@@ -240,6 +240,16 @@ def parse_args():
         help="The request rewriter to use. Default is 'noop' (no rewriting).",
     )
 
+    # Drop params arguments
+    parser.add_argument(
+        "--drop-params",
+        type=str,
+        default=None,
+        help="Comma-separated list of OpenAI parameters to drop from requests. "
+        "This allows dropping unsupported parameters by your LLM provider. "
+        "Example: 'frequency_penalty,logit_bias'",
+    )
+
     # Batch API
     # TODO(gaocegege): Make these batch api related arguments to a separate config.
     parser.add_argument(

src/vllm_router/services/request_service/request.py

@@ -203,6 +203,13 @@ async def route_general_request(
                 status_code=400, detail="Request body is not JSON parsable."
             )
 
+    if hasattr(request.app.state, "drop_params") and request.app.state.drop_params:
+        for param in request.app.state.drop_params:
+            request_json.pop(param, None)
+            logger.debug(f"Dropped param {param} from request")
+        request_body = json.dumps(request_json)
+        update_content_length(request, request_body)
+
     service_discovery = get_service_discovery()
     endpoints = service_discovery.get_endpoint_info()
 

tests/e2e/run-static-discovery-routing-test.sh

@@ -99,7 +99,8 @@ start_router() {
         --decode-model-labels "decode" \
         --static-model-labels "prefill,decode" \
         --session-key "$SESSION_KEY" \
-        --routing-logic "$routing_logic" > "$log_file" 2>&1 &
+        --routing-logic "$routing_logic" \
+        --drop-params "test-param-to-drop" > "$log_file" 2>&1 &
 
     ROUTER_PID=$!
     print_status "Router started with PID: $ROUTER_PID"

tests/e2e/test-routing.py

@@ -157,15 +157,20 @@ def _save_routing_lines(
             )
             return False
 
-    def send_request(self, request_id: str, prompt: str) -> bool:
+    def send_request(
+        self, request_id: str, prompt: str, custom_payload: dict = None
+    ) -> bool:
         """Send a single request and track which endpoint it goes to"""
         try:
-            payload = {
-                "model": self.model,
-                "prompt": prompt,
-                "temperature": 0.7,
-                "max_tokens": 10,
-            }
+            if custom_payload is not None:
+                payload = custom_payload
+            else:
+                payload = {
+                    "model": self.model,
+                    "prompt": prompt,
+                    "temperature": 0.7,
+                    "max_tokens": 10,
+                }
 
             headers = {
                 "Content-Type": "application/json",
@@ -542,6 +547,43 @@ def test_chat_completions(self) -> bool:
             print_error(f"❌ Chat completions failed: {e} payload: {payload}")
             return False
 
+    def test_drop_params(self) -> bool:
+        """Test that the router drops specified parameters from requests"""
+        print_status("🧪 Testing drop_params functionality")
+
+        # Send a request with parameters that should be dropped
+        try:
+            # Use the existing send_request method with custom payload to test drop_params
+            custom_payload = {
+                "model": self.model,
+                "prompt": "Test prompt with parameters to drop",
+                "temperature": 0.7,
+                "max_tokens": 10,
+                "test-param-to-drop": 0.5,  # This should be dropped
+            }
+
+            # Send request using existing method with custom payload
+            if not self.send_request(
+                "test-drop-params-request",
+                "Test prompt with parameters to drop",
+                custom_payload,
+            ):
+                print_error("❌ Drop params test request failed")
+                return False
+
+            # Check router logs for evidence that parameters were dropped
+            content = self._read_log_file()
+            if content is not None:
+                if "Dropped param test-param-to-drop from request" in str(content):
+                    print_status("✅ Drop params test request completed successfully")
+                    return True
+
+            print_error("❌ Drop params test request failed")
+            return False
+        except Exception as e:
+            print_error(f"❌ Unexpected error in drop params test: {e}")
+            return False
+
     def run_test(self) -> bool:
         """Run the complete routing test"""
         try:
@@ -559,6 +601,10 @@ def run_test(self) -> bool:
             if not self.test_chat_completions():
                 return False
 
+            # Test drop_params functionality
+            if not self.test_drop_params():
+                return False
+
             # Test routing logic
             test_runners = {
                 "roundrobin": self.test_roundrobin_routing,