Who contributed to PyTorch on August 28, 2026?

15 developers shipped this update, including adabeyta, dnikolaev-amd, gabrielfruet, PyTorch MergeBot, Zidong Yang, Yuanyuan Chen, Jason Ansel, and shoumikhin, and 7 more.

What were the notable PyTorch updates?

Consolidate CUDA helpers and remove the legacy _C extension in place of Stable-ABI, Consolidate CUDA helpers and remove the legacy _C extension in place of Stable-ABI (#9633), and Revert "Drop unused includes in c10 (#194959)".

Aug 28, 2026

$ the-wire · showcase

PyTorch Kills the Legacy C Extension

By RepoJournal · Filed 06:59 UTC on August 28, 2026 · About PyTorch

15 people shipped this

shoumikhin @shoumikhin 3 cited

adabeyta @adabeyta 2 cited

github-actions[bot] @github-actions[bot] 2 cited

gabrielfruet @gabrielfruet 1 cited

PyTorch MergeBot @pytorchmergebot 1 cited

Yuanyuan Chen @cyyever 1 cited

Jason Ansel @jansel 1 cited

can-gaa-hou @can-gaa-hou 1 cited

dnikolaev-amd @dnikolaev-amd

Zidong Yang @zidongyang

Yufeng Shi @YufengShi-dudu

mansnils @mansnils

rascani @rascani

jithunnair-amd @jithunnair-amd

huydhn @huydhn

The unstable `_C` extension is gone, replaced by a stable ABI that will make every vision build cleaner and safer.

The biggest story this morning is in pytorch/vision, where the final op has been ported out of the legacy `_C` extension and everything now builds into `_C_stable` [1] [2]. This eliminates the dual build and the hand-maintained source list, a long-awaited cleanup. Meanwhile, pytorch/pytorch is wrestling with a revert of the c10 include cleanup [3] that conflicted with another revert, though the original change [4] should be safe to rebase and remerge. Dynamo got three CPython-faithfulness fixes for `DequeVariable` [5], and a stale float specialization bug in `tensorify_python_scalars` was patched [6]. On the TensorRT desk, a complex decomposition fix keeps rotary attention compiling to a single engine [7], and a separate fix keeps `x & False` and `x | True` out of TensorRT [8]. Test-infra shipped two releases [9] [10], and the queue_time metric for multi-job workflows is now accurate [11]. The release notes for the nightly wheels in executorch note that new features land on main and reach users through the nightly build before the next stable release [12].

Action items

→ Update to a vision build with the stable ABI removal and verify your custom CUDA sources still compile pytorch/vision [immediate]
→ Track the c10 include cleanup rebase and remerge; avoid cherry-picking around it pytorch/pytorch [monitor]
→ Install ExecuTorch nightly wheels if you need the latest features before the next stable release pytorch/executorch [plan]
→ Verify TensorRT dynamo results for rotary attention and bitwise ops after these fixes pytorch/TensorRT [monitor]

References

[1] Consolidate CUDA helpers and remove the legacy _C extension in place of Stable-ABI ↗ pytorch/vision
[2] Consolidate CUDA helpers and remove the legacy _C extension in place of Stable-ABI (#9633) pytorch/vision
[3] Revert "Drop unused includes in c10 (#194959)" pytorch/pytorch
[4] Drop unused includes in c10 (#194959) pytorch/pytorch
[5] [dynamo] Fix DequeVariable empty popleft, maxlen on repeat, and delitem index wrapping (#195010) pytorch/pytorch
[6] Fix stale float specialization in tensorify_python_scalars (#194976) (#195040) pytorch/pytorch
[7] fix(dynamo): keep the complex decomposition inside the TRT op set ↗ pytorch/TensorRT
[8] fix(dynamo): keep `x & False` and `x | True` out of TensorRT ↗ pytorch/TensorRT
[9] v20260827-192339 ↗ pytorch/test-infra
[10] v20260827-145407 ↗ pytorch/test-infra
[11] [CRCR] Implement workflow-level queue_time tracking and update Redis helper for job dependencies ↗ pytorch/test-infra
[12] Document nightly wheels and linking macOS backends from C++ (#22240) pytorch/executorch