diff --git a/verl/workers/rollout/vllm_rollout/vllm_async_server.py b/verl/workers/rollout/vllm_rollout/vllm_async_server.py
index d4194e1ce55..3b24590410e 100644
--- a/verl/workers/rollout/vllm_rollout/vllm_async_server.py
+++ b/verl/workers/rollout/vllm_rollout/vllm_async_server.py
@@ -641,6 +641,11 @@ async def clear_kv_cache(self):
             # is a no-op success, so we can pass it unconditionally.
             await self.engine.reset_prefix_cache(**_RESET_PREFIX_CACHE_KWARGS)
 
+            if _VLLM_VERSION >= version.parse("0.9.0"):
+                await self.engine.reset_mm_cache()
+            if _VLLM_VERSION >= version.parse("0.16.0"):
+                await self.engine.reset_encoder_cache()
+
     async def release_kv_cache(self):
         """Release only kv_cache GPU memory, keeping model weights intact.
         # TODO: support true release of kv_cache
diff --git a/verl/workers/rollout/vllm_rollout/vllm_rollout.py b/verl/workers/rollout/vllm_rollout/vllm_rollout.py
index fffe57565d0..052abaf8664 100644
--- a/verl/workers/rollout/vllm_rollout/vllm_rollout.py
+++ b/verl/workers/rollout/vllm_rollout/vllm_rollout.py
@@ -190,7 +190,7 @@ async def update_weights(
         if future is not None:
             await future
 
-        # reset prefix cache after updating weights
+        # reset caches after updating weights
         if self.rollout_rank == 0:
             await self.server_handle.clear_kv_cache.remote()
             if global_steps is not None: