transformers
Fix silent SDPA math-kernel fallback for GQA when key/value head_dim > 256 or differ
#46960
Open
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Overview
Commits
9
Changes
View On
GitHub
Fix silent SDPA math-kernel fallback for GQA when key/value head_dim > 256 or differ
#46960
Butterfingrz
wants to merge 9 commits into
huggingface:main
from
Butterfingrz:fix/sdpa-gqa-large-head-dim-fallback
vasqu
commented on 2026-06-29
Fix silent SDPA math-kernel fallback for GQA with head_dim > 256
b41db1f8
Butterfingrz
force pushed
from
e5a98ff3
to
b41db1f8
2 days ago
Fix silent SDPA math-kernel fallback for GQA when key/value head_dim …
f572ef56
Butterfingrz
requested a review
from
vasqu
2 days ago
vasqu
commented on 2026-06-30
Fix silent SDPA math-kernel fallback for GQA when key/value head_dim …
8a4bd6c8
Butterfingrz
requested a review
from
vasqu
1 day ago
Butterfingrz
changed the title
Fix silent SDPA math-kernel fallback for GQA with head_dim > 256
Fix silent SDPA math-kernel fallback for GQA when key/value head_dim > 256 or differ
1 day ago
vasqu
approved these changes on 2026-07-01
vasqu
commented on 2026-07-01
Test SDPA GQA stays off the math kernel
32390c44
Butterfingrz
requested a review
from
vasqu
1 day ago
vasqu
approved these changes on 2026-07-02
Consolidate SDPA GQA tests into one forced-backend test
0d4332d4
Butterfingrz
requested a review
from
vasqu
3 hours ago
vasqu
commented on 2026-07-02
Parametrize the SDPA GQA test by expected backend
b89dac94
Butterfingrz
requested a review
from
vasqu
2 hours ago
vasqu
approved these changes on 2026-07-02
Merge branch 'main' into fix/sdpa-gqa-large-head-dim-fallback
ab5ee8ff
vasqu
enabled auto-merge
2 hours ago
Merge branch 'main' into fix/sdpa-gqa-large-head-dim-fallback
84563cac
disabled auto-merge
1 hour ago
Manually disabled by user
vasqu
enabled auto-merge
25 minutes ago
Merge branch 'main' into fix/sdpa-gqa-large-head-dim-fallback
56178db9
Login to write a write a comment.
Login via GitHub
Reviewers
vasqu
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub