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

Expose Flex attention causal/decode masks as static methods (#26514)

Browse Source 
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Cheng Wan
2026-05-28 01:03:32 -07:00
committed by GitHub
co-authored by Claude Opus 4.7
parent e5f5d84780
commit b429a30428
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/torch_flex_backend.py
+4 -2
View File
@@ -70,10 +70,12 @@ class TorchFlexAttnBackend(AttentionBackend):
) )
) )
def _causal_mask(self, b, h, q_idx, kv_idx): @staticmethod
def _causal_mask(b, h, q_idx, kv_idx):
return q_idx >= kv_idx return q_idx >= kv_idx
def _decode_mask(self, b, h, q_idx, kv_idx): @staticmethod
def _decode_mask(b, h, q_idx, kv_idx):
return q_idx <= kv_idx return q_idx <= kv_idx
def _run_flex_forward_extend( def _run_flex_forward_extend(