[Bugfix] Fix inconsistent handling of cache reset (#33481)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

docs/benchmarking/sweeps.md

@@ -82,7 +82,7 @@ vllm bench sweep serve \
     You can use `--dry-run` to preview the commands to be run.
 
     We only start the server once for each `--serve-params`, and keep it running for multiple `--bench-params`.
-    Between each benchmark run, we call the `/reset_prefix_cache` and `/reset_mm_cache` endpoints to get a clean slate for the next run.
+    Between each benchmark run, we call all `/reset_*_cache` endpoints to get a clean slate for the next run.
     In case you are using a custom `--serve-cmd`, you can override the commands used for resetting the state by setting `--after-bench-cmd`.
 
 !!! note

vllm/benchmarks/sweep/server.py

@@ -12,6 +12,12 @@ from typing_extensions import Self
 
 
 class ServerProcess:
+    VLLM_RESET_CACHE_ENDPOINTS = [
+        "/reset_prefix_cache",
+        "/reset_mm_cache",
+        "/reset_encoder_cache",
+    ]
+
     def __init__(
         self,
         server_cmd: list[str],
@@ -120,11 +126,9 @@ class ServerProcess:
             server_address = self._get_vllm_server_address()
             print(f"Resetting caches at {server_address}")
 
-            res = requests.post(f"{server_address}/reset_prefix_cache")
-            res.raise_for_status()
-
-            res = requests.post(f"{server_address}/reset_mm_cache")
-            res.raise_for_status()
+            for endpoint in self.VLLM_RESET_CACHE_ENDPOINTS:
+                res = requests.post(server_address + endpoint)
+                res.raise_for_status()
         elif server_cmd[0].endswith("infinity_emb"):
             if "--vector-disk-cache" in server_cmd:
                 raise NotImplementedError(

vllm/entrypoints/openai/engine/serving.py

@@ -286,10 +286,6 @@ class OpenAIServing:
             raise TypeError(f"{reasoning_parser_name=} has not been registered") from e
         return parser
 
-    async def reset_mm_cache(self) -> None:
-        self.input_processor.clear_mm_cache()
-        await self.engine_client.reset_mm_cache()
-
     async def beam_search(
         self,
         prompt: PromptType,

vllm/v1/engine/async_llm.py

@@ -741,6 +741,7 @@ class AsyncLLM(EngineClient):
         if clear_cache:
             await self.reset_prefix_cache()
             await self.reset_mm_cache()
+            await self.reset_encoder_cache()
 
     async def resume_generation(self) -> None:
         """Resume generation after :meth:`pause_generation`."""

vllm/v1/worker/gpu/mm/encoder_runner.py

@@ -31,6 +31,22 @@ class EncoderRunner:
         self.req_id_to_mm_features: dict[str, list[MultiModalFeatureSpec]] = {}
         self.encoder_cache: dict[str, torch.Tensor] = {}
 
+    def reset_mm_cache(self) -> None:
+        """
+        Clear the multi-modal cache that was used during profiling,
+        but no longer needed during inference.
+        """
+        # TODO: Implement MM budget for encoder dummy run
+        pass
+
+    def reset_encoder_cache(self) -> None:
+        """Clear the GPU-side encoder cache storing vision embeddings.
+
+        This should be called when model weights are updated to ensure
+        stale embeddings computed with old weights are not reused.
+        """
+        self.encoder_cache.clear()
+
     def add_request(self, req_id: str, mm_features: list[MultiModalFeatureSpec]):
         self.req_id_to_mm_features[req_id] = mm_features
 

vllm/v1/worker/gpu/model_runner.py

@@ -339,7 +339,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         gc.collect()
 
     def reset_mm_cache(self) -> None:
-        pass
+        self.encoder_runner.reset_mm_cache()
+
+    def reset_encoder_cache(self) -> None:
+        self.encoder_runner.reset_encoder_cache()
 
     def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
         # SP is not supported yet.

vllm/v1/worker/gpu_model_runner.py

@@ -717,6 +717,10 @@ class GPUModelRunner(
                 self.effective_drafter_max_model_len = self.max_model_len
 
     def reset_mm_cache(self) -> None:
+        """
+        Clear the multi-modal cache that was used during profiling,
+        but no longer needed during inference.
+        """
         if self.mm_budget:
             self.mm_budget.reset_cache()