Who contributed to PyTorch on August 22, 2026?

13 developers shipped this update, including AmesingFlank, kitzbergerg, ethche, unography, 谢翊凡, dependabot, Tomasz Szulist, and Cui, Yifeng, and 5 more.

What were the notable PyTorch updates?

[Pallas] Stage aligned HBM windows on any tensor axis, [Pallas] Lower static value slices as basic indices, and [Pallas] Stage aligned dynamic HBM windows.

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

Pallas HBM Windows Open on Any Tensor Axis

By RepoJournal · Filed 06:59 UTC on August 22, 2026 · About PyTorch

13 people shipped this

AmesingFlank @AmesingFlank 3 cited

kitzbergerg @kitzbergerg 1 cited

unography @unography 1 cited

Cui, Yifeng @CuiYifeng 1 cited

Vincent Moens @vmoens 1 cited

ethche @ethche

谢翊凡 @xyf5432

dependabot @dependabot

Tomasz Szulist @tszulist-hbn

Yuanyuan Chen @cyyever

Warren Deng @warrendeng

Frederik Gossen @frgossen

priba @priba

Pallas stagers just gained the ability to align HBM windows on any tensor axis, not just the last one, and a critical autotuning crash is fixed.

The Pallas frontend is flexing serious new muscle with a 16-PR stack from AmesingFlank that extends stage-aligned HBM windows to any tensor axis [1], covers static value slices [2], and supports aligned dynamic HBM windows [3] for projection kernels. Meanwhile, the autotuner now handles kernels with callable arguments, fixing the `TypeError: Cannot serialize type: code` that broke AOT tuning for things like flex attention's score_mod [4]. On the XPU front, a torch-xpu-ops commit pin [5] lands fixes for register spilling, group_reduce asserts, and TopK NaN behavior. "AOT tuning currently does not work for any kernel with a callable input" [4], but that's no longer the case. Elsewhere, DreamerV3 gets a focused correctness patch fixing episode-boundary semantics [6], and TensorDict deprecates `TensorDictModule.device` ahead of removal in 0.17 [7]. Buckle up, because the Pallas changes alone shift how you write tile-based kernels.

Action items

→ Review Pallas PRs #3390 and #3404 for your projection epilogues pytorch/helion [plan]
→ Update torch-xpu-ops pin to 1a3415 for XPU correctness fixes pytorch/pytorch [plan]
→ Test your AOT autotuning with callables after PR #3325 pytorch/helion [plan]
→ Watch TensorDictModule.device deprecation; prepare for removal in 0.17 pytorch/tensordict [monitor]

References

[1] [Pallas] Stage aligned HBM windows on any tensor axis ↗ pytorch/helion
[2] [Pallas] Lower static value slices as basic indices ↗ pytorch/helion
[3] [Pallas] Stage aligned dynamic HBM windows ↗ pytorch/helion
[4] [autotuner] AOT tuning for callables ↗ pytorch/helion
[5] Update torch-xpu-ops commit pin (#194185) pytorch/pytorch
[6] [BugFix] Match DreamerV3 reference semantics ↗ pytorch/rl
[7] [Deprecation] Deprecate TensorDictModule.device (#1770) pytorch/tensordict

Quick answers

What shipped in PyTorch on August 22, 2026?: Pallas stagers just gained the ability to align HBM windows on any tensor axis, not just the last one, and a critical autotuning crash is fixed. In total, 63 commits and 18 pull requests landed.
Who contributed to PyTorch on August 22, 2026?: 13 developers shipped this update, including AmesingFlank, kitzbergerg, ethche, unography, 谢翊凡, dependabot, Tomasz Szulist, and Cui, Yifeng, and 5 more.
What were the notable PyTorch updates?: [Pallas] Stage aligned HBM windows on any tensor axis, [Pallas] Lower static value slices as basic indices, and [Pallas] Stage aligned dynamic HBM windows.

NVFP4 GroupedExperts lands on Blackwell

TorchAO's new differentiable NVFP4 grouped GEMM stack brings high-throughput grouped experts to Blackwell, and it's the headline act.

+10

python 160 shipped

@pytorch Aug 19

Pytorch Fixes XPU Dispatch Gap and CI OOMs

PyTorch's XPU backend lost a critical dispatch for _safe_softmax, and CI is getting a GOMAXPROCS pin to stop GPU nodes from OOMKilling.

+10

python 85 shipped

@pytorch Aug 18

PYTORCH SLASHES B200 BENCH TIME 5X

The B200 microbenchmark suite just went from a 7.5-hour serial slog to a 143-minute parallel run, and that's only the start of what's moving across PyTorch today.

python 92 shipped

@pytorch Aug 17

PYTORCH CUTS MPS DEAD CODE, DREAMER V3 FIXES GPU MEMORY LEAK

PyTorch is aggressively pruning unused Metal Performance Shaders infrastructure while TorchRL patches a critical memory leak that kept loss tensors pinned to CUDA for entire training runs.

python 25 shipped

Elsewhere on the wire

AI Agents Aug 22

Cline hubs get drain and replay as Langchain plumb gates

Cline just shipped rollover-grade hub commands while Langchain exposes gateway internals, and OpenHands quietly broke its own monitoring contract.

ai-agents 42 shipped

Local LLMs Aug 22

SGLang ships v0.5.18 with 710 PRs and DeepSeek V4 FP4 auto-routing

The biggest release of the week just landed: SGLang v0.5.18 adds new models, fixes a silent disagg bug, and defaults DeepSeek V4 FP4 checkpoints to the right MoE runner.