RepoJournal
PyTorch @pytorch
90 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 2026-07-13
stories 14
GitHub RSS Archive Reviews Topics

© 2026 RepoJournal Home Showcase Explore How it works Privacy

$ the-wire · showcase

PYTORCH FIXES CRITICAL MPS MATH BUG WHILE TAMING CI GHOSTS

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

qflen Jason Ansel Aaron Orenstein Yang, Shuxin jewelkm89 PyTorch UpdateBot
6 people shipped this
qflen @qflen 1 cited Jason Ansel @jansel 1 cited Aaron Orenstein @aorenste 1 cited Yang, Shuxin @orrangetabby17 1 cited jewelkm89 @jewelkm89 1 cited PyTorch UpdateBot @pytorchupdatebot

PyTorch's Metal backend shipped an accurate erfc function that eliminates 100% relative error on upper tail computations, fixing a cascade of broken numerics in GELU and special functions.

The Metal Performance Shaders backend had a silent catastrophe: `c10::metal::erfc` was computing `1.0 - erf(x)`, which saturates to zero once erf hits fp32 limits around x=3.9, poisoning everything downstream [1]. That means GELU activations, gamma function completements, and `torch.special.erfc` all inherited the flush. This fix splits out a dedicated Metal implementation that holds accuracy across the entire range. In parallel, PyTorch axed a stale Dynamo skip decorator on max_pool1d tests [2] now that the CPU pooling suite no longer triggers LLVM OOM failures, restoring regression coverage that had been dormant. On the CI side, the team hardened crcr-test to handle deleted push dispatches without false L1 probe failures [3]. GitHub sends a null SHA when branches or tags get deleted during PR cleanup; the old code tried checking out ref `0000...`, which crashed. The fix adds a dedicated `push-deleted` job that validates deletion semantics and skips checkout entirely. Routine inductor work continues: fbcode xfails for numerical property mismatches [4] and XPU maxpool backward fallback tuning [5].

Action items

References

  1. [1] Add an accurate Metal erfc for MPS (eager + inductor) (#189291) pytorch/pytorch ↗
  2. [2] Remove stale Dynamo skip for max_pool1d pooling test (#186531) pytorch/pytorch ↗
  3. [3] Handle deleted push dispatches (ciflow tag cleanup) ↗ pytorch/crcr-test
  4. [4] [inductor] IS_FBCODE xfails for opinfo numeric-property tests (#189640) pytorch/pytorch ↗
  5. [5] [xpu][inductor] fix the fallback behavior of maxpool backward for performance (#187940) pytorch/pytorch ↗

$ ls pytorch/reviews/ # the step-back reads

the week in review Helion autotunes Metal, ending default 16x16x16 tiles the month in review ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave
Topics: Python AI / ML Full archive → Weekly reviews →

Quick answers

What shipped in PyTorch on July 13, 2026?
PyTorch's Metal backend shipped an accurate erfc function that eliminates 100% relative error on upper tail computations, fixing a cascade of broken numerics in GELU and special functions. In total, 13 commits and 1 pull requests landed.
Who contributed to PyTorch on July 13, 2026?
6 developers shipped this update, including qflen, Jason Ansel, PyTorch UpdateBot, Aaron Orenstein, Yang, Shuxin, and jewelkm89.
What were the notable PyTorch updates?
Add an accurate Metal erfc for MPS (eager + inductor) (#189291), Remove stale Dynamo skip for max_pool1d pooling test (#186531), and Handle deleted push dispatches (ciflow tag cleanup).
♥

$ log @pytorch --recent

torchtitan fixes multi-node HF export, torch._native opens to non-aten namespaces 57 shipped Fake-tensor hardening, ROCm graph fixes, and an XNNPACK use-after-free 180 shipped CPU caching allocator teardown race fixed, torchtitan shares one training engine 67 shipped ROCm test timeouts traced to duplicated runs, and distributed/test_store drops 15 minutes 100 shipped

$ log --all # elsewhere on the wire

@openclaw Cold-start Gateway diagnostics stop crying wolf, doctor repairs plugin drift Agentic Coding Cline desktop bundles shrink, goose unblocks its own clicks

$ ls showcases/

Rails Django Laravel Node.js Go Vercel Vue.js Kubernetes HashiCorp Arch Linux Linux Anthropic

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.