apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Revert "Revert "[Optim] Remove limitation of number of kvcache blocks (#5612)…"

Browse Source 
This reverts commit c1aa66df02.
This commit is contained in:
Jiang-Jia-Jun
2025-12-23 15:42:10 +08:00
committed by GitHub
parent c1aa66df02
commit 6250c686cc
7 changed files with 20 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/worker/worker_process.py
View File

@@ -530,11 +530,9 @@ class PaddleDisWorkerProc:
# 2. Calculate the appropriate number of blocks
model_block_memory_used = self.worker.cal_theortical_kvcache()
num_blocks_local = int(available_kv_cache_memory // model_block_memory_used)
# NOTE(liuzichang): Too many block will lead to illegal memory access
# We will develop dynamic limits in future.
if num_blocks_local > 40000:
logger.info(f"------- Reset num_blocks_local {num_blocks_local} to 40000")
num_blocks_local = min(40000, num_blocks_local)
if envs.FD_MAX_KVCACHE_BLOCKS > 0 and num_blocks_local > envs.FD_MAX_KVCACHE_BLOCKS:
logger.info(f"------- Reset num_blocks_local {num_blocks_local} to {envs.FD_MAX_KVCACHE_BLOCKS}")
num_blocks_local = envs.FD_MAX_KVCACHE_BLOCKS
logger.info(f"------- model_block_memory_used:{model_block_memory_used / 1024**3} GB --------")
logger.info(f"------- num_blocks_local:{num_blocks_local} --------")