From f3bf61105486730699660fac20aed50a92d87272 Mon Sep 17 00:00:00 2001
From: Junrong Lin <33685709+ocss884@users.noreply.github.com>
Date: Thu, 15 May 2025 10:15:02 +0800
Subject: [PATCH] feat: add flush cache to EngineBase and
 HttpServerEngineAdapter (#6009)

---
 python/sglang/srt/entrypoints/EngineBase.py         | 5 +++++
 python/sglang/srt/entrypoints/http_server_engine.py | 3 +++
 2 files changed, 8 insertions(+)

diff --git a/python/sglang/srt/entrypoints/EngineBase.py b/python/sglang/srt/entrypoints/EngineBase.py
index 894a6f065..c7dfafd41 100644
--- a/python/sglang/srt/entrypoints/EngineBase.py
+++ b/python/sglang/srt/entrypoints/EngineBase.py
@@ -27,6 +27,11 @@ class EngineBase(ABC):
         """Generate outputs based on given inputs."""
         pass
 
+    @abstractmethod
+    def flush_cache(self):
+        """Flush the cache of the engine."""
+        pass
+
     @abstractmethod
     def update_weights_from_tensor(
         self,
diff --git a/python/sglang/srt/entrypoints/http_server_engine.py b/python/sglang/srt/entrypoints/http_server_engine.py
index f4d81a417..f50e13f3c 100644
--- a/python/sglang/srt/entrypoints/http_server_engine.py
+++ b/python/sglang/srt/entrypoints/http_server_engine.py
@@ -140,3 +140,6 @@ class HttpServerEngineAdapter(EngineBase):
 
     def resume_memory_occupation(self):
         return self._make_request("resume_memory_occupation")
+
+    def flush_cache(self):
+        return self._make_request("flush_cache")