RepoJournal

$ cat pytorch/week/2026-09-07.log

PyTorch

PyTorch

the week in review · Sep 7 – Sep 13, 2026

TorchRL 0.14 lands with DreamerV3 and async collectors

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

CUDA 12.6 nightly wheels stop at PyTorch 2.15, which also drops Maxwell, Pascal, and Volta GPU support.

471 commits 267 PRs merged 15 releases 7 briefings covered

all pytorch reviews →

TorchRL 0.14: DreamerV3, Async at Scale, and Composable Replay pytorch/rl

by vmoens

The major feature release spans model-based RL, distributed training and inference, replay-buffer composition, checkpointing, on-policy optimization, vision-language-action workflows, and simulation tooling, while tightening several contracts. TorchRL users should check the tightened contracts before upgrading: this is flagged as breaking.

[CD] Remove CUDA 12.6 nightly binary builds (#196249) pytorch/pytorch

by Andrey Talman

The notice says CUDA 12.6 wheels will no longer be published from PyTorch 2.15, which also drops Maxwell, Pascal, and Volta. The companion test-infra change removes only the builds themselves, with the underlying code cleaned up later.

[Performance] Batch AsyncBatchedCollector shared-memory coordination pytorch/rl

by vmoens

The collector now coordinates shared-memory environments from one thread, batches ready requests, and drains inference slots fairly, and it returns an exact final frame batch even when total_frames is not divisible.

[Feature] Add compile-safe pause to AsyncBatchedCollector pytorch/rl

by vmoens

A new AsyncBatchedCollector.pause() context manager establishes a quiescent boundary with a per-pause coordinator barrier, finishing in-flight policy and environment work before entering and resuming every coordinator on normal or exceptional exit.

[Feature] Add CPU affinity controls to AsyncEnvPool pytorch/rl

by vmoens

AsyncEnvPool gains per-worker Linux CPU affinity masks applied before environment construction, plus a driver affinity mask for coordinator and inference-server execution. The docs cover cpuset, CFS quota, and Kubernetes CPU Manager interactions.

Fix stale autotune Future reuse across compilations (#196148) pytorch/pytorch

by Aaron Orenstein

AsyncAutotuner.choice_hash_to_future is a process-global dict keyed by choice hash plus inputs key, which did not identify the compilation it belonged to, so start() skipped submitting when the key was already present. The fix scopes the Future to the compilation.

[BugFix] Bound AsyncEnvPool queue shared mappings pytorch/rl

by vmoens

Per-environment queue results are copied out of multiprocessing shared memory before being returned, with lazy-stacked results copied while dense stacks that already own their storage avoid an extra copy.

[MPS] Remove svd_kernel_mps small-input CPU gate (#195950) pytorch/pytorch

by Nikita Shulga

The small-input gate that routed tensors with numel < 8192 to CPU is gone; linalg.svd has no _ex variant and always runs its error check, so the gate was a net loss for an input already on the MPS device.

$ 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 →