apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-30 14:22:27 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[XPU]Fixed the issue of performance degradation caused by enabling ENABLE_V1_KVCACHE_SCHEDULER (#3900)

Browse Source 
* fix bug

* fix bug

* update

* udpate

* update
This commit is contained in:
yinwei
2025-09-05 19:17:25 +08:00
committed by GitHub
parent 473cde779f
commit 77c1bd0813
4 changed files with 20 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
fastdeploy/engine/args_utils.py
View File

@@ -19,6 +19,8 @@ from dataclasses import asdict, dataclass
from dataclasses import fields as dataclass_fields
from typing import Any, Dict, List, Optional
import paddle
from fastdeploy import envs
from fastdeploy.config import (
CacheConfig,
@@ -1006,7 +1008,10 @@ class EngineArgs:
if self.max_num_batched_tokens is None:
if int(envs.ENABLE_V1_KVCACHE_SCHEDULER):
self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
if paddle.is_compiled_with_xpu():
self.max_num_batched_tokens = self.max_model_len
else:
self.max_num_batched_tokens = 8192 # if set to max_model_len, it's easy to be OOM
else:
if self.enable_chunked_prefill:
self.max_num_batched_tokens = 2048