Who contributed to PyTorch on August 27, 2026?

12 developers shipped this update, including Gasoonjia, RJ Ascani, Anthony Shoumikhin, mcremon-meta, quinnarnold, gtnv, Yuanyuan Chen, and albanD, and 4 more.

What were the notable PyTorch updates?

Remove the deprecated MPS backend (#22181), Fix issues in the export recipe framework (#22158), and Cortex-M: lower subtraction, and stop miscompiling add with alpha (#22153).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

EXECUTORCH STRIPS MPS, PYTORCH DROPS UNUSED DEPS

By RepoJournal · Filed 06:59 UTC on August 27, 2026 · About PyTorch

12 people shipped this

Anthony Shoumikhin @shoumikhin 3 cited

RJ Ascani @rascani 2 cited

Yuanyuan Chen @cyyever 2 cited

Fierynx @Fierynx 1 cited

jinsooihm @jinsooihm 1 cited

Gasoonjia @Gasoonjia

mcremon-meta @mcremon-meta

quinnarnold @quinnarnold

gtnv @gtnv

albanD @albanD

Pian Pawakapan @pianpwk

wwwjn @wwwjn

ExecuTorch finally deletes the deprecated MPS backend and PyTorch cuts dead weight in its build, while the rest of the stack quietly hardens.

ExecuTorch removed the MPS backend as promised in its deprecation notice [1], shipping in 1.5.0 after being overdue. Meanwhile, the project fixed a subtle graph copy bug that wasted 68 MB of memory on a four-method model [2], and fixed a Cortex-M kernel miscompilation that produced incorrect results for quantized add with an alpha multiplier [3]. On PyTorch core, the build no longer links the unused cuRAND host library [4] and generate_code.py lost a flag that hadn't done anything since 2021 [5]. The gradcheck fix addresses a long-standing issue where non-differentiable outputs could corrupt projections [6]. Torchtitan's MFU calculation was overcounting FLOPS for sliding window attention models, sometimes exceeding 100% [7]. TensorRT's Windows CI was failing 170 of 300 jobs due to a malformed artifact name [8], and ExecuTorch's packaged runner now registers the device copy kernels it was missing [9].

Action items

→ Update ExecuTorch to remove MPS backend dependencies pytorch/executorch [plan]
→ Review quantized add alpha handling on Cortex-M pytorch/executorch [monitor]
→ Verify gradcheck behavior with non-differentiable outputs pytorch/pytorch [plan]

References

[1] Remove the deprecated MPS backend (#22181) pytorch/executorch
[2] Fix issues in the export recipe framework (#22158) pytorch/executorch
[3] Cortex-M: lower subtraction, and stop miscompiling add with alpha (#22153) pytorch/executorch
[4] Stop linking the cuRAND host library (#194850) pytorch/pytorch
[5] Drop the dead force_schema_registration flag from generate_code.py (#194843) pytorch/pytorch
[6] Fix fast gradcheck when non-differentiable outputs precede differentiable ones (#194222) pytorch/pytorch
[7] Fix mfu calculation for muse glimmer and gpt_oss ↗ pytorch/torchtitan
[8] fix(ci): build the Windows download artifact name in the right order ↗ pytorch/TensorRT
[9] fix(executorch): register device copy kernels in the packaged runner ↗ pytorch/TensorRT

Quick answers

What shipped in PyTorch on August 27, 2026?: ExecuTorch finally deletes the deprecated MPS backend and PyTorch cuts dead weight in its build, while the rest of the stack quietly hardens. In total, 103 commits and 50 pull requests landed.
Who contributed to PyTorch on August 27, 2026?: 12 developers shipped this update, including Gasoonjia, RJ Ascani, Anthony Shoumikhin, mcremon-meta, quinnarnold, gtnv, Yuanyuan Chen, and albanD, and 4 more.
What were the notable PyTorch updates?: Remove the deprecated MPS backend (#22181), Fix issues in the export recipe framework (#22158), and Cortex-M: lower subtraction, and stop miscompiling add with alpha (#22153).

EXECUTORCH CUTS GRAPH REPLAYS, ARM KILLS NO-OP RESIZES

Two executorch optimizations landed overnight that could shave real runtime on device, while PyTorch's CI finally silences the Node-20 roar.

python 110 shipped

@pytorch Aug 24

FSDP2 CRASH FIXED, EXECUTORCH BUILD PIPELINE REPAIRED AFTER COMMIT PROMOTION BUG

PyTorch's distributed training guard catches an AttributeError that was silently breaking gradient accumulation in sharded models, while ExecuTorch's nightly builds restart after a scheduling bug locked out half of all commits.