What pytorch/pytorch shipped
Written by FoxPlug from the repository's own releases; not affiliated with PyTorch. Paste your own repository in the box below and FoxPlug writes the same weekly update for your product. This page is an automatic summary of the releases, pull requests and commits in github.com/pytorch/pytorch; PyTorch did not write it and does not use or endorse FoxPlug. Every line links to the change it describes.
Get a weekly update like this for your product, free
Week of October 5, 2026
What shipped
- Release-only changes were applied to the 2.15 branch, pinning GitHub actions and workflow triggers to the release branch. Pull request #199819
- The torch-xpu-ops commit pin was updated, adding mixed-precision support for fused Adam and AdamW on XPU and int64 support for upsample bilinear 2d channels-last. Pull request #200289
- Abi3 wheels are now enabled for Python 3.12 and newer in the Triton integration. Pull request #199841
- Python 3.15 wheels now depend on Triton and smoke-test torch.compile on 3.15 and XPU. Pull request #200320
- ProcessGroupNCCL.Options gained a
lazy_initsetting so creating a group no longer allocates a communicator until the device is bound. Pull request #200291 - The XPU static launcher now reports
n_spillsin dword-equivalents per lane, matching the units used by CUDA and HIP threshold checks. Pull request #200407 - XPU CI runners now reclaim workspace ownership before checkout so container-written files can be read by the runner. Pull request #200351
- The cu13.4 ptxas bundling workaround was reverted because the updated Triton pin no longer needs it. Pull request #199936
- The MPS strided softmax kernel now splits the reduction dimension across threadgroups. Pull request #200069
- Stale exp and silu
inductor_defaultxfails were removed for ROCm after the Triton 3.9 pin update. Pull request #200157
Why it matters
This week brought several fixes and updates to XPU, CD, and Triton integration, including a new torch-xpu-ops pin and Python 3.15 wheel support. The release 2.15 branch received release-only workflow changes, and several post-cut reverts brought it back in line with trunk. Users of XPU and ROCm see better CI reliability and a statically-launched kernel reporting fix, while NCCL group creation can avoid unnecessary memory allocation.
Changelog entry
- Apply release-only changes to the 2.15 branch Pull request #199819
- Update torch-xpu-ops commit pin Pull request #200289
- Enable abi3 wheels for Python 3.12+ Pull request #199841
- Depend on triton for Python 3.15 wheels and smoke-test torch.compile on 3.15 and XPU Pull request #200320
- Add ProcessGroupNCCL.Options.lazy_init Pull request #200291
- Report static launcher
n_spillsin dword-equivalents per lane on XPU Pull request #200407 - Reclaim workspace ownership before checkout on XPU runners Pull request #200351
- Revert cu13.4 ptxas bundling workaround Pull request #199936
- Split the reduction dim of strided softmax across threadgroups on MPS Pull request #200069
- Remove stale exp/silu
inductor_defaultxfails after Triton 3.9 on ROCm Pull request #200157 - Add stable dtype and device PyObject conversions and an
is_tensorprobe Pull request #200183 - Cherry-pick fix preventing pattern rewrites from dropping undeclared non-default kwargs Pull request #199988
This week: release-only changes landed on 2.15, torch-xpu-ops pin updated with fused Adam/AdamW mixed precision and int64 upsample, Triton abi3 wheels enabled for Python 3.12+, Python 3.15 wheels now pull Triton, and XPU CI and softmax fixes went in.
This week in PyTorch: release-only workflow changes were applied to the 2.15 branch, including pinned actions and release-branch triggers. The torch-xpu-ops commit pin was updated with mixed-precision fused Adam/AdamW and int64 upsample support for channels-last. Triton abi3 wheels are now enabled for Python 3.12+, and Python 3.15 wheels now declare Triton as a dependency and smoke-test torch.compile. ProcessGroupNCCL.Options gained lazy_init to avoid allocating a communicator at group creation. XPU CI now reclaims workspace ownership before checkout, and the XPU static launcher reports n_spills in consistent units. The MPS strided softmax kernel splits the reduction dimension across threadgroups. The cu13.4 ptxas bundling workaround was reverted.