[V1][BugFix] Free encoder cache for aborted requests (#12545)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

vllm/v1/core/encoder_cache_manager.py

@@ -38,7 +38,8 @@ class EncoderCacheManager:
     def get_cached_input_ids(self, request: Request) -> Set[int]:
         return self.cached.get(request.request_id, set())
 
-    def free(self, request: Request, input_id: int) -> None:
+    def free_encoder_input(self, request: Request, input_id: int) -> None:
+        """Free a single encoder input id for the request."""
         req_id = request.request_id
         if req_id not in self.cached:
             return
@@ -49,6 +50,12 @@ class EncoderCacheManager:
         self.num_free_slots += request.get_num_encoder_tokens(input_id)
         self.freed.append((req_id, input_id))
 
+    def free(self, request: Request) -> None:
+        """Free all cached input ids for the request."""
+        input_ids = self.get_cached_input_ids(request)
+        for input_id in input_ids:
+            self.free_encoder_input(request, input_id)
+
     def get_freed_ids(self) -> List[Tuple[str, int]]:
         freed = self.freed
         self.freed = []