fix

Author: hiworldwzj

lightllm/server/multi_level_kv_cache/cpu_cache_client.py

@@ -123,7 +123,7 @@ def update_pages_status_to_ready(
                     assert cur_page.ref_count > 0
                     cur_page.ref_count -= 1
                     if cur_page.ref_count == 0:
-                        # 放回 LRU 列表头部
+                        # 放回 LRU 列表尾部
                         self.page_items.add_item_to_tail(cur_page.self_index)
 
                 # 全部落盘，已落盘前缀部分会在落盘中自动剔除

lightllm/server/multi_level_kv_cache/manager.py

@@ -164,16 +164,11 @@ def _handle_group_req_multi_cache_match(self, group_req_indexes: GroupReqIndexes
                 continue
 
             finded_page_indexes: List[int] = []
-            disk_service = (
-                self.disk_cache_worker.service
-                if (self.disk_cache_worker is not None and self.disk_cache_worker.service is not None)
-                else None
-            )
             req.disk_prompt_cache_len = 0
 
             # 匹配 CPU cache
             all_pages = self._cpu_cache_match(token_hash_list)
-            if len(all_pages) == len(token_hash_list) or disk_service is None:
+            if len(all_pages) == len(token_hash_list) or self.only_cpu_cache_enable:
                 finded_page_indexes = all_pages
             else:
                 # 匹配 disk cache并load到cpu cache

lightllm/server/router/model_infer/mode_backend/multi_level_kv_cache.py

@@ -254,8 +254,6 @@ def update_cpu_cache_task_states(self):
                 self.cpu_cache_client.lock.acquire_sleep1ms()
                 # 分组update，避免不同请求的page交叉，导致disk cache hash不一致
                 for pages in page_array_list:
-                    if not pages:
-                        continue
                     self.cpu_cache_client.update_pages_status_to_ready(
                         page_list=pages, deref=True, disk_offload_enable=self.args.enable_disk_cache
                     )

lightllm/utils/envs_utils.py

@@ -208,7 +208,7 @@ def enable_diverse_mode_gqa_decode_fast_kernel() -> bool:
 
 @lru_cache(maxsize=None)
 def get_disk_cache_prompt_limit_length():
-    return int(os.getenv("LIGHTLLM_DISK_CACHE_PROMPT_LIMIT_LENGTH", 10000))
+    return int(os.getenv("LIGHTLLM_DISK_CACHE_PROMPT_LIMIT_LENGTH", 2048))
 
 
 @lru_cache(maxsize=None)