RepoJournal

$ cat pytorch/week/2026-11-16.log

PyTorch

PyTorch

the week in review · Nov 16 – Nov 22, 2026

Helion lands 26 tracing and masking fixes

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

Most of the fixes correct silent wrong results in device code; DTensor's sharded any/all now returns a partial instead of all-gathering.

408 commits 248 PRs merged 4 releases 7 briefings covered

all pytorch reviews →

[compiler] Fix in-place updates, dead-code elimination, barriers and scans (#4453) pytorch/helion

by Jason Ansel

Thirteen fixes to how device code is traced and lowered, each shipped with a test. Most address silent wrong results: an in-place method inside a loop or branch was dropped, and a host statement with side effects was eliminated as dead.

[compiler] Mask padded hl.arange lanes and partial tiles; fix config slots (#4454) pytorch/helion

by Jason Ansel

Another thirteen fixes, seven of them masking bugs where lanes past the end of a padded hl.arange or a partial tile took part in means, reductions, loop carries, loads and stores. If you write reductions over padded shapes, your numbers may have been wrong without an error.

[DTensor] Reduce sharded any/all locally with safe partials (#204895) pytorch/pytorch

by Aditya Venkataraman

The all-gather before a sharded any/all is gone, so the op now produces a partial instead of a replicated tensor. The PR flags this as potentially BC-breaking, so check any code that relies on the old replicated output.

[executorch][cuda] Route 2-D depthwise and few-channel grouped convolutions to a Triton direct-stencil kernel pytorch/executorch

by pytorchbot

Two-dimensional depthwise and few-channel grouped convolutions on the CUDA backend now route to a Triton direct-stencil kernel. Previously those shapes were likely reaching a less specific path.

Fold weight transposes at export time in to_executorch pytorch/executorch

by shoumikhin

nn.Linear exports as a weight transpose plus a matmul, and on the portable path that transpose stayed a runtime op. Folding it at export time cuts planned memory, with a small 3 layer MLP as the measured case.

Add CUDA 13.5 to the nightly build and validation matrix (Linux) pytorch/test-infra

by atalman

Nightlies now build and validate against CUDA 13.5 on Linux x86_64 and aarch64, with cu135 wheels already published. Verify your extension and kernel builds on the new toolkit before it reaches a stable release.

portable: return early from amax/amin on an empty input pytorch/executorch

by AxelNoun

amax.out and amin.out skip their fast path on an empty input, then build a plan whose constructor aborts the process. A model with a dynamic batch dimension that received an empty batch was killed; the fix returns early.

$ ls pytorch/week/ # the briefings behind this review

Keep up with PyTorch in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

all pytorch reviews →