Who contributed to PyTorch on May 18, 2026?

1 developer shipped this update, including jeanschmidt.

What were the notable PyTorch updates?

[Inductor] Fix NaN output in LayerNorm CPU by guarding Welford variance. (#173989), [bugfix] [inductor] add meta registration for `div_` kernel to enforce dtype check (#183859), and [inductor] Fix tail reduction suffix width (#183699).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

INDUCTOR STACK FIXES CRITICAL NANOUTPUTS AND DTYPE BUGS WHILE CI INFRA PIVOTS TO IPV6

By RepoJournal · Filed 06:03 UTC on May 18, 2026 · About PyTorch

1 person shipped this

jeanschmidt @jeanschmidt 1 cited

PyTorch's compiler pipeline landed three critical fixes overnight that prevent NaN outputs in LayerNorm, enforce dtype safety in in-place division, and repair tail reduction logic, while infrastructure teams are destroying and recreating all OSDC clusters for IPv6 migration.

The inductor desk closed out a rough week. LayerNorm on CPU with float16 was silently producing NaN when inputs contained Inf values [1], a numerical collapse in Welford variance that now gets guarded with a where() check. Separately, the `div_` kernel was type-promoting incorrectly on int/long inputs, mismatching eager semantics [2], fixed by pulling it out of the generic binop handler and adding explicit dtype validation. The tail reduction suffix width bug [3] that was corrupting vector stores into reduction buffers is also closed. These three PRs unblock compiles that were either silent-wrong or crashing entirely. On the runtime side, JIT stack handling got tightened [4] by replacing the error-prone last()+drop() pattern with pop(), reducing subtle ordering bugs. Meanwhile, the CI infrastructure team is executing a major shift: all OSDC EKS clusters (staging and production) are moving from IPv4 to IPv6-only pod networking [5], a high-risk change that requires full cluster destroy/recreate since the EKS `ip_family` parameter is immutable post-creation. The migration touches the entire stack: VPC subnets, CNI prefix delegation, and every component that binds sockets or resolves DNS. Fresh cluster deploys are also getting fixed [6] with Alpine util-linux pin bumps and cross-arch Docker build support. Test infrastructure landed performance wins on the autorevert metrics page [7], parallelizing GitHub fan-out that was serializing 100-300s of latency behind ClickHouse queries, plus a new killswitch window category [8] to surface human reverts during autorevert blackouts as non-false-negatives.

Action items

→ Review and merge the three inductor fixes (LayerNorm NaN guard, div_ dtype enforcement, tail reduction suffix) before your next CPU compile deploy pytorch/pytorch [immediate]
→ Coordinate OSDC cluster IPv6 migration with your DevOps team; plan for full cluster destroy/recreate and test dual-stack resolution across all workloads pytorch/ci-infra [immediate]
→ Monitor the ARC Helm chart bump (0.14.1-jeanschmidt.10) for HUD API fallback behavior on your runner scale sets pytorch/ci-infra [monitor]
→ Validate autorevert metrics load times after the FP verification parallelization lands pytorch/test-infra [plan]

References

[1] [Inductor] Fix NaN output in LayerNorm CPU by guarding Welford variance. (#173989) pytorch/pytorch
[2] [bugfix] [inductor] add meta registration for `div_` kernel to enforce dtype check (#183859) pytorch/pytorch
[3] [inductor] Fix tail reduction suffix width (#183699) pytorch/pytorch
[4] Use pop in place of last() + drop() in JIT runtime (#184063) pytorch/pytorch
[5] Add IPv6-only pod networking to EKS clusters and bump runner-container-hooks to v0.8.13 ↗ pytorch/ci-infra
[6] Fix cross-arch image-cache-janitor build for fresh cluster deploys (#575) pytorch/ci-infra
[7] autorevert metrics: parallelize FP verification + default to 30d window (#8090) pytorch/test-infra
[8] autorevert metrics: third FN category for killswitch-active windows (#8089) pytorch/test-infra

Quick answers

What shipped in PyTorch on May 18, 2026?: PyTorch's compiler pipeline landed three critical fixes overnight that prevent NaN outputs in LayerNorm, enforce dtype safety in in-place division, and repair tail reduction logic, while infrastructure teams are destroying and recreating all OSDC clusters for IPv6 migration. In total, 28 commits and 5 pull requests landed.
Who contributed to PyTorch on May 18, 2026?: 1 developer shipped this update, including jeanschmidt.
What were the notable PyTorch updates?: [Inductor] Fix NaN output in LayerNorm CPU by guarding Welford variance. (#173989), [bugfix] [inductor] add meta registration for `div_` kernel to enforce dtype check (#183859), and [inductor] Fix tail reduction suffix width (#183699).

CRITICAL OIDC INJECTION IN DOCS PREVIEW WORKFLOW PATCHED

PyTorch's docs-preview CI trusted fork-controlled artifacts in a context with token-write permissions, exposing the entire build pipeline to code injection.

python 66 shipped 1-min read

@pytorch 1 day ago

PYTORCH AUTOGRAD GETS 7% FASTER, AOTI FIXES SILENT FAILURES

Interned attribute names in autograd.Function shaved microseconds off the hot path while AOTI's scatter operations now properly report errors instead of silently corrupting results.

python 64 shipped 1-min read

@pytorch 4 days ago

DYNAMO REVERTS BREAKING CHANGE, EXECUTORCH CLEANS UP DEPRECATED TYPES

PyTorch reverted a Dynamo optimization that broke internal tests, while ExecutorTorch is aggressively deprecating c10 shims in favor of standard library types.

python 91 shipped 1-min read

@pytorch 5 days ago

PYTORCH SHIPS BUILD FIX WHILE HELION TUNES H100 KERNELS TO DEFAULT

A critical build regression in cusparselt.cpp is now patched, while the kernel autotuner promotes its pointwise seed heuristic to production defaults on H100 and B200.

python 36 shipped 1-min read

Elsewhere on the wire

AI Agents about 9 hours ago

CLAUDE OPUS 5 LANDS ACROSS THE STACK

The newest Anthropic model is now live in langchain, Cline, and llama-index, with native support for extended reasoning and 1M context windows.

ai-agents 28 shipped 1-min read

Local LLMs about 9 hours ago

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.