chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Support FlashAttention3 page_size > 1 and topk > 1 case with paged attn and spec decode (#7725)

Browse Source
This commit is contained in:
Yubo Wang
2025-11-25 19:44:41 -08:00
committed by GitHub
parent ca5c8b16f6
commit 18fb51583f
9 changed files with 706 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -1720,7 +1720,7 @@ class ServerArgs:
if (
self.speculative_eagle_topk > 1
and self.page_size > 1
and self.attention_backend != "flashinfer"
and self.attention_backend not in ["flashinfer", "fa3"]
):
raise ValueError(
"speculative_eagle_topk > 1 with page_size > 1 is unstable and produces incorrect results for paged attention backends. This combination is only supported for the 'flashinfer' backend."