153 wires and counting

$ follow PyTorch

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.

$ status

wire 2026-11-13
stories 153

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

TorchTitan picks up HybridEP and non-SP vision encoders; tensordict patches a segfault

By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology

TorchTitan's Muse Spark path gained a third expert-parallel backend and its checkpoint reader stopped trusting fsspec's listing cache, while tensordict fixes a consolidate() call that took down the interpreter.

[BugFix] consolidate() raises NotImplementedError for a quantized tensor instead of ending the process pytorch/tensordict

by peterdsharpe

consolidate() on a tensordict holding a quantized tensor used to segfault the whole Python process, no exception and no traceback, and with filename=... the file at that path was already replaced before the crash. It now raises NotImplementedError instead, so a bad call costs you an exception rather than a dead worker.

[muse_spark] Support HybridEP expert parallelism (#6032) pytorch/torchtitan

by Chien-Chin Huang

Muse Spark supported only AllToAll and Dist-MoE, the first slow and the second unusable on Hopper; torchtitan now enables HybridEP. Because non-blocking HybridEP never writes rows past the last expert offset, output_postprocess needs torch-specific handling to keep the pre-combine RMSNorm correct.

[checkpoint] Read remote checkpoint folders without fsspec's listing cache (#6061) pytorch/torchtitan

by Zhixiang Li

With checkpointer.folder on a remote fsspec path and async_mode="async_with_pinned_mem", keep_latest_k never deleted anything: tools/filesystem.py listed through fsspec's cache, which only sees writes from the same filesystem instance, while the save subprocess wrote through its own. Remote checkpoint folders are now read without that cache, and on fsspec older than 2026.6.0 each thread kept it...

GreenLight: linear review policy with broad rules, and eligibility for any merge-rule approver (#9417) pytorch/test-infra

by Jean Schmidt

GreenLight's review policy in .claude/skills/greenlight-review/SKILL.md is rewritten as one linear decision procedure with broader generic rules, and eligibility widens to every approver merge_rules.yaml names, on any path. All 12 open GreenLight feedback reports would get the opposite verdict under the new rules.

[td scorer] Daily refit, gate and publish to generated-stats (#9427) pytorch/test-infra

by Jean Schmidt

A new daily td-scorer.yml workflow fits each test config's per-file start-up from the last 7 days of CI shards, builds any missing day shards, fits the scorer and the risk score, then gates before publishing to generated-stats. The longer tail: executorch added ETDump profiling for CpuBackend and serialized node.meta["custom"] in exir, while test-infra removed obsolete pre-review machinery.

Quick answers

What shipped in PyTorch on November 13, 2026?
TorchTitan's Muse Spark path gained a third expert-parallel backend and its checkpoint reader stopped trusting fsspec's listing cache, while tensordict fixes a consolidate() call that took down the interpreter. In total, 88 commits, 62 pull requests, and 3 releases landed.
Who contributed to PyTorch on November 13, 2026?
18 developers shipped this update, including Chien-Chin Huang, Animesh Jain, Zhixiang Li, Driss Guessous, _githubsgi, Baris, pssrawat, and JP, and 10 more.
What were the notable PyTorch updates?
[BugFix] consolidate() raises NotImplementedError for a quantized tensor instead of ending the process, [muse_spark] Support HybridEP expert parallelism (#6032), and [checkpoint] Read remote checkpoint folders without fsspec's listing cache (#6061).