4 Commits

Author SHA1 Message Date
Jarvik7
dce37a234c Update sage_attention_patch.py
Fix inappropriate assert on Blackwell (SM120) that broke sage attention.
Tested with Torch2.9 nightly. Saves avg 2s compared to sdpa.
2025-09-05 10:19:13 -03:00
WildAi
0b9f8a06f0 SageAttention support, fixes 2025-09-03 11:44:05 +03:00
drbaph
f565f123c6 Transformers 4.56+ Compatibility & Force Offload Fix 2025-09-01 19:26:59 +01:00
WildAi
4056f54f86 init 2025-08-27 15:51:44 +03:00