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

Fix fa3 window size setup (#5316)

Browse Source
This commit is contained in:
Qingquan Song
2025-04-11 23:37:52 -07:00
committed by GitHub
parent e4155e96d0
commit aea98512a8
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/attention/flashattention_backend.py
View File

@@ -523,7 +523,7 @@ class FlashAttentionBackend(AttentionBackend):
# here is two side inclusive # here is two side inclusive
window_size = ( window_size = (
(layer.sliding_window_size, 0) (layer.sliding_window_size, 0)
if layer.sliding_window_size is not None if layer.sliding_window_size is not None and layer.sliding_window_size > -1
else (-1, -1) else (-1, -1)
) )
k_descale, v_descale = None, None k_descale, v_descale = None, None
@@ -664,7 +664,7 @@ class FlashAttentionBackend(AttentionBackend):
# here is two side inclusive # here is two side inclusive
window_size = ( window_size = (
(layer.sliding_window_size, 0) (layer.sliding_window_size, 0)
if layer.sliding_window_size is not None if layer.sliding_window_size is not None and layer.sliding_window_size > -1
else (-1, -1) else (-1, -1)
) )
causal = not layer.is_cross_attention causal = not layer.is_cross_attention