From 941a11ada3fcc301958ffe7d36c6c7b3f478aac4 Mon Sep 17 00:00:00 2001
From: shuwenn <47200617+alphabetc1@users.noreply.github.com>
Date: Fri, 19 Jun 2026 17:46:04 +0800
Subject: [PATCH] [HiCache] refactor: remove unused transfer buffer (#26003)

---
 .../sglang/srt/managers/cache_controller.py   | 54 +------------------
 1 file changed, 2 insertions(+), 52 deletions(-)

diff --git a/python/sglang/srt/managers/cache_controller.py b/python/sglang/srt/managers/cache_controller.py
index d5e782aa1..5d552abe4 100644
--- a/python/sglang/srt/managers/cache_controller.py
+++ b/python/sglang/srt/managers/cache_controller.py
@@ -16,7 +16,7 @@ limitations under the License.
 import logging
 import threading
 import time
-from queue import Empty, Full, Queue
+from queue import Empty, Queue
 from typing import TYPE_CHECKING, List, NamedTuple, Optional
 
 import torch
@@ -149,45 +149,6 @@ class HiCacheAck(NamedTuple):
     node_ids: List[int]
 
 
-class TransferBuffer:
-    """
-    Overlapping buffer preparation and transfer operations to improve throughput.
-    """
-
-    def __init__(self, stop_event, buffer_count: int = 3) -> None:
-        self.stop_event = stop_event
-        self.buffers = Queue(maxsize=buffer_count)
-
-    def full(self) -> bool:
-        return self.buffers.full()
-
-    def empty(self) -> bool:
-        return self.buffers.empty()
-
-    def put(self, item, block=True, timeout=1) -> None:
-        while not self.stop_event.is_set():
-            try:
-                self.buffers.put(item, block=block, timeout=timeout)
-                break
-            except Full:
-                if not block:
-                    break
-                continue
-            except Exception as e:
-                logger.error(e)
-
-    def get(self, block=True, timeout=1) -> Optional[CacheOperation]:
-        try:
-            return self.buffers.get(block=block, timeout=timeout)
-        except Empty:
-            return None
-        except Exception as e:
-            logger.error(e)
-
-    def clear(self):
-        self.buffers.queue.clear()
-
-
 class StorageOperation:
     counter = 0
 
@@ -298,9 +259,6 @@ class HiCacheController:
         self.page_set_func = self._generic_page_set
 
         # Dedicated stop event for storage background threads (prefetch/backup).
-        # NOTE: Do NOT reuse `self.stop_event` here since it also guards core HiCache
-        # transfer buffers (CPU<->GPU). We want to allow runtime attach/detach of
-        # storage without stopping the whole controller.
         self.storage_stop_event = threading.Event()
 
         self.device = self.mem_pool_device.device
@@ -321,10 +279,6 @@ class HiCacheController:
         self.ack_load_queue: List[HiCacheAck] = []
         self.ack_write_queue: List[HiCacheAck] = []
 
-        self.stop_event = threading.Event()
-        self.write_buffer = TransferBuffer(self.stop_event)
-        self.load_buffer = TransferBuffer(self.stop_event, buffer_count=10)
-
         self.write_stream = device_module.Stream()
         self.load_stream = device_module.Stream()
 
@@ -419,7 +373,7 @@ class HiCacheController:
         # Always request stop. This is safe even when storage is already disabled,
         # and makes detach truly idempotent (previous partial detach may have left
         # threads alive).
-        # NOTE: do NOT clear stop_event unless threads have fully stopped; otherwise
+        # NOTE: do NOT clear storage_stop_event unless threads have fully stopped; otherwise
         # a still-alive thread may resume and touch released state.
         self.storage_stop_event.set()
 
@@ -671,13 +625,10 @@ class HiCacheController:
         )
 
     def reset(self):
-        self.stop_event.set()
         self.storage_stop_event.set()
 
         self.write_queue.clear()
         self.load_queue.clear()
-        self.write_buffer.clear()
-        self.load_buffer.clear()
         self.ack_write_queue.clear()
         self.ack_load_queue.clear()
         if self.enable_storage:
@@ -690,7 +641,6 @@ class HiCacheController:
             self.host_mem_release_queue.queue.clear()
             self.prefetch_tokens_occupied = 0
 
-        self.stop_event.clear()
         self.storage_stop_event.clear()
 
         if self.enable_storage: