📝 document logic for suppressin gabort timeouts

Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>

vllm/entrypoints/openai/rpc/client.py

@@ -336,8 +336,16 @@ async def _is_tracing_enabled_rpc(self) -> bool:
     async def abort(self, request_id: str):
         """Send an ABORT_REQUEST signal to the RPC Server"""
 
-        # Suppress timeouts as well- if the server is busy and does not ack in
-        # time we assume it got the message.
+        # Suppress timeouts as well.
+        # In cases where the server is busy processing requests and a very
+        # large volume of abort requests arrive, it is likely that the server
+        # will not be able to ack all of them in time. We have seen this when
+        # we abort 20k requests at once while another 2k are processing- many
+        # of them time out, but we see the server successfully abort all of the
+        # requests.
+        # In this case we assume that the server has received or will receive
+        # these abort requests, and ignore the timeout. This prevents a massive
+        # wall of `TimeoutError` stack traces.
         with suppress(RPCClientClosedError, TimeoutError):
             await self._send_one_way_rpc_request(
                 request=RPCAbortRequest(request_id),