COMMITS
August 25, 2026
G
[caffe2] Unblock caffe2_gpu under CUDA 13.3 (cudafe++ alias-decl reprint bug) (#194659)
Grace Cheng(Dev Infra) committed
J
[PyTorch][jit] Avoid IValue copies in unpickler (#193250) (#193250)
Joshua Su committed
P
Revert "[SDPA][CUDA] Prefer cuDNN SDPA backend on sm86 and sm89 (#192799)"
PyTorch MergeBot committed
August 24, 2026
E
[inductor] Fuse interleaved epilogues in nested reductions (#190595)
eellison committed
J
Strengthen headeronly linter (#194486)
Jane Xu committed
Y
Remove disable/enable_ubsan (#194445)
Yuanyuan Chen committed
G
Fix pdist and cdist launch failure past 2^24 outputs on ROCm (#188006)
geozhai committed
E
[inductor] Update dynamic vision_maskrcnn graph break baselines (#194630)
Elias Ellison committed
G
[SDPA][CUDA] Prefer cuDNN SDPA backend on sm86 and sm89 (#192799)
Grayson Derossi committed
S
Add sync_cache_decision_cross_ranks (#193043) (#193043)
Shuai Yang committed
G
[CI] Read re-enabled issues with git log, not git cherry (#194485)
George Hong committed
L
Reapply "Dispatch index_put_ to fill_ for scalar values (#188267 / #189105)" (#189303)
Lukas Geiger committed
D
[sdpa] Fix bottom-right efficient attention when Q exceeds K (#193632)
drisspg committed
N
Clamp default num_threads to cpuset/affinity mask (#194605)
Nikita Shulga committed
A
[S615805][S694227][PyTorch] Safely preserve tensor strides (#194251) (#194251)
Angeline Mingwei Zhang committed
E
A
[ROCm] - Filter out CK gemms for when ONLY gfx1250 arch is provided (#192399)
Andy Lugo committed
A
Fix Windows py3.15 builds: constrain Cython < 3.3.0 and bump the cp315 numpy pin (#194618)
Andrey Talman committed
A
D
E
bump tolerances for flex attn test (#194000)
eqy committed
A
Fix libtorch-extract being skipped on nightly-branch pushes (#194610)
Andrey Talman committed
P
[autograd] Expose queue_callback as torch.autograd.graph.queue_callback (#193958)
Pawel Gadzinski committed
B
[Dynamo] derive registered Stream/Event classes from DeviceInterface (#194343)
Beaver2323 committed
P
Revert "[sdpa] Fix bottom-right efficient attention when Q exceeds K (#193632)"
PyTorch MergeBot committed
E
August 17, 2026
K
[inductor] Coordinate-descent tune compile-time combo kernel blocks (#189003)
karthickai committed
K
[inductor] Make combo compile-time autotune deterministic-mode safe (#189724)
karthickai committed
August 21, 2026
J
Introduce stateless RNG APIs for randint / raw bits (#190253)
Joel Schlosser committed