[Bugfix] Backport request id validation to v0

vllm/engine/multiprocessing/client.py

@@ -576,6 +576,10 @@ async def _process_request(
         if self._errored_with is not None:
             raise ENGINE_DEAD_ERROR(self._errored_with)
 
+        # Ensure the request id is unique among running requests
+        if request_id in self.output_queues:
+            raise ValueError(f"Request {request_id} already exists")
+
         # Constructing guided decoding logits processors is expensive, so we do
         # it here to avoid contending with cpu resources and the GIL on the
         # backend process.

vllm/v1/engine/async_llm.py

@@ -142,7 +142,7 @@ async def add_request(
         """Add new request to the AsyncLLM."""
 
         if self.detokenizer.is_request_active(request_id):
-            raise KeyError(f"Request {request_id} already exists.")
+            raise ValueError(f"Request {request_id} already exists.")
 
         # 1) Create a new AsyncStream for the request.
         stream = self._add_request_to_streams(request_id)