EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Set schedule policy more conservative for DP attention (#2096)

Browse Source
This commit is contained in:
Ke Bao
2024-11-20 12:57:18 +08:00
committed by GitHub
parent ffd20fcd03
commit 699384cb01
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/server_args.py
View File

@@ -191,10 +191,12 @@ class ServerArgs:
self.dp_size = self.tp_size self.dp_size = self.tp_size
self.chunked_prefill_size = self.chunked_prefill_size // 2 self.chunked_prefill_size = self.chunked_prefill_size // 2
self.cuda_graph_max_bs = min(self.cuda_graph_max_bs, 96) self.cuda_graph_max_bs = min(self.cuda_graph_max_bs, 96)
self.schedule_conservativeness = self.schedule_conservativeness * 0.3
self.enable_overlap_schedule = False self.enable_overlap_schedule = False
logger.warning( logger.warning(
f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. " f"DP attention is enabled. The chunked prefill size is adjusted to {self.chunked_prefill_size} to avoid MoE kernel issues. "
f"The CUDA graph max batch size is adjusted to {self.cuda_graph_max_bs}. " f"The CUDA graph max batch size is adjusted to {self.cuda_graph_max_bs}. "
f"The schedule conservativeness is adjusted to {self.schedule_conservativeness}. "
"Data parallel size is adjusted to be the same as tensor parallel size." "Data parallel size is adjusted to be the same as tensor parallel size."
) )