apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[XPU] Fixed the issue of performance degradation caused by enabling ENABLE_V1_KVCACHE_SCHEDULER (#3393)

Browse Source 
* fix v1 schedule oom bug

* fix v1 schedule oom bug
This commit is contained in:
yinwei
2025-08-14 17:41:40 +08:00
committed by GitHub
parent 28918702c2
commit 101605869c
4 changed files with 22 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
fastdeploy/engine/config.py
View File

@@ -236,10 +236,13 @@ class Config:
if self.cache_config.enable_chunked_prefill:
self.max_num_batched_tokens = 2048
else:
if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
if not int(os.getenv("ENABLE_V1_KVCACHE_SCHEDULER", "0")):
self.max_num_batched_tokens = self.max_model_len
else:
self.max_num_batched_tokens = 8192
if paddle.is_compiled_with_xpu():
self.max_num_batched_tokens = self.max_model_len
else:
self.max_num_batched_tokens = 8192
if self.long_prefill_token_threshold == 0:
self.long_prefill_token_threshold = int(self.max_model_len * 0.04)
@@ -287,7 +290,7 @@ class Config:
)
if not self.cache_config.enable_chunked_prefill:
if not int(os.getenv('ENABLE_V1_KVCACHE_SCHEDULER', '0')):
if not int(os.getenv("ENABLE_V1_KVCACHE_SCHEDULER", "0")):
assert self.max_num_batched_tokens >= self.max_model_len, (
f"max_num_batched_tokens: {self.max_num_batched_tokens} "
f"should be larger than or equal to max_model_len: {self.max_model_len}"