ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add check to disable xformers on newer GPUs (#4342)

Browse source 
Disable xformers for GPUs with compute capability >= 12 to ensure compatibility with newer hardware.
This commit is contained in:
pluesclues 2026-03-17 01:42:38 -04:00 • committed by GitHub
commit f5e1f52b48
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 7 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

7
unsloth/utils/attention_dispatch.py
View file

@ -33,6 +33,13 @@ from ..utils.packing import (
if HAS_FLASH_ATTENTION:
from flash_attn import flash_attn_func, flash_attn_varlen_func
HAS_XFORMERS = xformers is not None
# xformers kernels (FA3, FA2, cutlass) only support compute capability <= 9.0.
# Disable xformers on newer GPUs (e.g. RTX 5070 Ti / sm_120) and fall back to SDPA.
if HAS_XFORMERS and torch.cuda.is_available():
_cc = torch.cuda.get_device_capability()
if _cc[0] >= 12:
HAS_XFORMERS = False
SDPA_HAS_GQA = "enable_gqa" in (scaled_dot_product_attention.__doc__ or "")
FLASH_VARLEN = "flash_varlen"