diff --git a/docs/index.rst b/docs/index.rst
index 85a5da22155..cb3f7227613 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -102,6 +102,7 @@ verl is fast with:
    perf/dpsk.md
    perf/best_practices
    perf/perf_tuning
+   perf/rollout_kv_offload.md
    README_vllm0.8.md
    perf/device_tuning
    perf/verl_profiler_system.md
diff --git a/docs/perf/rollout_kv_offload.md b/docs/perf/rollout_kv_offload.md
new file mode 100644
index 00000000000..4dd1c6b1ddb
--- /dev/null
+++ b/docs/perf/rollout_kv_offload.md
@@ -0,0 +1,57 @@
+# Rollout KV Cache Offload via Mooncake-Store
+
+Last updated: 05/27/2026.
+
+Offload prefix KV blocks from the vLLM rollout engine to a shared
+[Mooncake](https://github.com/kvcache-ai/Mooncake) store so long shared
+prefixes (system prompt, agentic tool history, `rollout.n` samples per prompt)
+get deduplicated across requests and rollout replicas. This also helps
+long-tail load balancing: when work migrates to idle rollout replicas, shared
+prefix KV reduces the re-prefill cost.
+
+## Setup Mooncake + vLLM
+
+Use vLLM 0.22 or newer; earlier vLLM versions do not provide the full
+MooncakeStoreConnector hard-reset behavior required by this integration.
+
+Follow vLLM's official guide for installing the Mooncake client, starting a
+master, and writing the JSON config:
+**<https://docs.vllm.ai/en/latest/features/mooncake_store_connector_usage/>**
+
+The verl side only consumes whatever that doc produces — no extra steps.
+
+## Enable in verl
+
+verl forwards `engine_kwargs.vllm.*` straight to `vllm serve` as CLI flags.
+To attach the Mooncake connector, set `kv_transfer_config`:
+
+```yaml
+actor_rollout_ref:
+  rollout:
+    engine_kwargs:
+      vllm:
+        kv_transfer_config: |-
+          {
+            "kv_connector": "MooncakeStoreConnector",
+            "kv_role": "kv_both",
+            "kv_connector_extra_config": {
+              "mooncake_config_path": "/path/to/mooncake_config.json"
+            }
+          }
+```
+
+Or as a Hydra CLI override:
+
+```bash
++actor_rollout_ref.rollout.engine_kwargs.vllm.kv_transfer_config.kv_connector=MooncakeStoreConnector \
++actor_rollout_ref.rollout.engine_kwargs.vllm.kv_transfer_config.kv_role=kv_both \
++actor_rollout_ref.rollout.engine_kwargs.vllm.kv_transfer_config.kv_connector_extra_config.mooncake_config_path=/path/to/mooncake_config.json
+```
+
+## RL correctness: hard reset on every weight update
+
+verl clears both local and Mooncake KV caches at every weight update boundary
+to avoid reusing KV from the previous policy.
+
+**Required vLLM version**: use vLLM 0.22 or newer. Older builds may leave stale
+KV in the Mooncake master after a weight update.
diff --git a/verl/workers/rollout/vllm_rollout/vllm_async_server.py b/verl/workers/rollout/vllm_rollout/vllm_async_server.py
index fee46046cc3..dadd9a4243d 100644
--- a/verl/workers/rollout/vllm_rollout/vllm_async_server.py
+++ b/verl/workers/rollout/vllm_rollout/vllm_async_server.py
@@ -607,11 +607,15 @@ async def wake_up(self, tags: list[str] | None = None):
             # processes across all DP shards (unlike collective_rpc which only reaches
             # TP workers within a single shard).
             await self.engine.wake_up(tags=tags or self._get_wake_up_tags())
-            await self.engine.reset_prefix_cache()
+            await self.engine.reset_prefix_cache(reset_connector=True)
         elif self.rollout_mode == RolloutMode.COLOCATED:
             # Directly call engine to wake up without sync weights.
             await self.engine.wake_up(tags=self._get_wake_up_tags())
-            await self.engine.reset_prefix_cache()
+            # reset_connector=True drops any attached external KV store
+            # (e.g. MooncakeStoreConnector) whose entries were computed
+            # against the previous weights. No-op success when no connector
+            # is configured (vLLM scheduler treats it as such).
+            await self.engine.reset_prefix_cache(reset_connector=True)
         elif self.rollout_mode == RolloutMode.STANDALONE:
             logger.info("skip wake_up in standalone mode")
 
@@ -628,7 +632,11 @@ async def sleep(self):
 
     async def clear_kv_cache(self):
         if self.node_rank == 0:
-            await self.engine.reset_prefix_cache()
+            # reset_connector=True drops any attached external KV store
+            # (e.g. MooncakeStoreConnector) whose entries were computed
+            # against the previous model weights. With no connector it
+            # is a no-op success, so we can pass it unconditionally.
+            await self.engine.reset_prefix_cache(reset_connector=True)
 
     async def release_kv_cache(self):
         """Release only kv_cache GPU memory, keeping model weights intact.
@@ -689,7 +697,12 @@ async def abort_all_requests(self, reset_prefix_cache: bool = True) -> dict[str,
                 # 1. Set engine to paused state (blocks new generate calls)
                 # 2. Abort all in-flight requests
                 # 3. Wait for requests to drain
-                # 4. Clear prefix and mm caches if clear_cache=True
+                # 4. Clear prefix and mm caches if clear_cache=True.
+                #    EngineCore._reset_caches defaults reset_connector=True
+                #    on this path, so any attached external KV store (e.g.
+                #    MooncakeStoreConnector) is invalidated along with the
+                #    local prefix cache — RL-correct hard-reset at every
+                #    weight update boundary, no extra kwargs needed.
                 await self.engine.pause_generation(
                     wait_for_inflight_requests=False,
                     clear_cache=reset_prefix_cache,