COMMITS
August 25, 2026
A
N
Clean up stale TORCH_NCCL_AVOID_RECORD_STREAMS and avoidRecordStreams_ (#194701)
Natalia Gimelshein committed
A
P
Revert "[inductor] Delete the post-fusion reindex retry (#193158)"
PyTorch MergeBot committed
J
[ROCm][Inductor] Fix CKTILE float16 autotune dtype mismatch (#193902)
Jagadish Krishnamoorthy committed
Z
A
[fx] Shallow-copy `from_node` provenance in `NodeSource` instead of deepcopy (#193051) (#193051)
Andrew Pullin committed
A
Fix macOS py3.15 wheel builds: pin Cython < 3.3.0 and bump the cp315 numpy pin (#194734)
Andrey Talman committed
I
J
I
[Inductor] Scale test_bmm_large_batch_dynamic inputs to unskip it on ROCm (#184531)
iupaikov-amd committed
S
Add XPU dispatch for linalg.polar (#188253)
Slawomir Siwek committed
V
[Test] Make dynamo test_streams.py device-agnostic for out-of-tree backends (#185213)
vishalgoyal316 committed
A
[Test] Generalize `export/test_torchbind.py` tests and enable on XPU (#192030)
Artur Kłoniecki committed
P
[triton hash update] update the pinned triton hash (#194667)
PyTorch UpdateBot committed
C
T
[distributed] Run legacy-only NCCL tests with nccl-legacy (#193240)
Tristan Rice committed
T
[distributed] Configure collective timeouts through ProcessGroup (#193238)
Tristan Rice committed
T
[c10d] Bind local rank before eager NCCL2 initialization (#193237)
Tristan Rice committed
Z
J
G
[dynamo] Support types.SimpleNamespace (#193719)
gabrielfruet committed
H
[CI] Add sm80 to the clang20 CUDA pull builds' arch list (#194685)
Huy Do committed
Y
[BC breaking] Remove assert_allclose (#164560)
Yuanyuan Chen committed
August 23, 2026
T
[assoc_scan] Memory-efficient backward via associative_scan (#192761)
Thomas Bohnstingl committed
August 21, 2026
M
[inductor] Vendor and adapt QuACK symmetric GEMM (#194177)
Michael Lazos committed
August 18, 2026
X
[CI] Add gfx950 FlyDSL Inductor test shard (#193473)
XiaobingSuper committed
August 24, 2026
X
[Inductor] Add gfx950 FlyDSL GEMM and torch.mm autotuning (#190903)
XiaobingSuper committed
J
[c10d] Gate NCCL CFT tests on CFT-capable GPU/driver/CUDA stacks (#194299)
Junjie Wang committed