Who contributed to PyTorch on August 19, 2026?

14 developers shipped this update, including Ryan Zhang, Jean Schmidt, Yu, Guangye, msnliu, Bryanzzy1, mathieuorhan, lanluo-nvidia, and Anthony Shoumikhin, and 6 more.

What were the notable PyTorch updates?

Update third_party/kineto submodule to 4f8caaa (#193990), Update torch-xpu-ops commit pin (#193929), and Add missing XPU: _safe_softmax_xpu dispatch for _safe_softmax (#193786).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

Pytorch Fixes XPU Dispatch Gap and CI OOMs

By RepoJournal · Filed 06:59 UTC on August 19, 2026 · About PyTorch

14 people shipped this

Ryan Zhang @ryanzhang22 1 cited

Jean Schmidt @jeanschmidt 1 cited

Yu, Guangye @guangyey 1 cited

msnliu @msnliu 1 cited

mathieuorhan @mathieuorhan 1 cited

Anthony Shoumikhin @shoumikhin 1 cited

Chien-Chin Huang @fegin 1 cited

SherlockNoMad @SherlockNoMad 1 cited

Bryanzzy1 @Bryanzzy1

lanluo-nvidia @lanluo-nvidia

Zupeng Wang @zupengwang

Tejaswin Parthasarathy @tp5uiuc

weifengpy @weifengpy

pianpwk @pianpwk

PyTorch's XPU backend lost a critical dispatch for _safe_softmax, and CI is getting a GOMAXPROCS pin to stop GPU nodes from OOMKilling.

A missing XPU dispatch line in `aten::_safe_softmax` broke the backend in 2.14, and the fix lands with this PR [3] while a matching commit pin updates torch-xpu-ops to address batchnorm accuracy and empty-batch FFT handling [2]. Over in CI-infra, dcgm-exporter was OOMKilling about 48% of 8-GPU pods because Go 1.24 ignores cgroup limits; pinning GOMAXPROCS=2 stops it dead [6]. TorchTitan is tightening its runtime: it drops Python 3.10 support entirely [5] and adds an opt-in fused MLA implementation for DeepSeek-V3, "Keeps V as a zero-copy view" [4]. The Kineto submodule bump pulls in a CUPTI buffer rejection bugfix [1], and TorchRL locks TransformedEnv specs by default so hooks can't mutate parent specs [7]. TensorRT's executorch fixes are worth a skim if you hit 0-d outputs [8].

Action items

→ Pull in the XPU dispatch fix and torch-xpu-ops pin if you're on XPU 2.14 pytorch/pytorch [immediate]
→ Apply GOMAXPROCS=2 to dcgm-exporter if you see 8-GPU OOMKills pytorch/ci-infra [immediate]
→ Plan for Python 3.11+ before the next torchtitan release pytorch/torchtitan [plan]
→ Review the fused MLA override for DeepSeek-V3 attention pytorch/torchtitan [monitor]

References

[1] Update third_party/kineto submodule to 4f8caaa (#193990) pytorch/pytorch
[2] Update torch-xpu-ops commit pin (#193929) pytorch/pytorch
[3] Add missing XPU: _safe_softmax_xpu dispatch for _safe_softmax (#193786) pytorch/pytorch
[4] Fused MLA override ↗ pytorch/torchtitan
[5] Drop Python 3.10 support, require Python 3.11+ (#4180) pytorch/torchtitan
[6] Pin GOMAXPROCS=2 on dcgm-exporter to stop 8-GPU node OOMKills ↗ pytorch/ci-infra
[7] [Performance] Lock TransformedEnv specs by default ↗ pytorch/rl
[8] fix(executorch): do not resize a 0-d delegate output (#4469) pytorch/TensorRT

Quick answers

What shipped in PyTorch on August 19, 2026?: PyTorch's XPU backend lost a critical dispatch for _safe_softmax, and CI is getting a GOMAXPROCS pin to stop GPU nodes from OOMKilling. In total, 57 commits and 28 pull requests landed.
Who contributed to PyTorch on August 19, 2026?: 14 developers shipped this update, including Ryan Zhang, Jean Schmidt, Yu, Guangye, msnliu, Bryanzzy1, mathieuorhan, lanluo-nvidia, and Anthony Shoumikhin, and 6 more.
What were the notable PyTorch updates?: Update third_party/kineto submodule to 4f8caaa (#193990), Update torch-xpu-ops commit pin (#193929), and Add missing XPU: _safe_softmax_xpu dispatch for _safe_softmax (#193786).

NVFP4 GroupedExperts lands on Blackwell

TorchAO's new differentiable NVFP4 grouped GEMM stack brings high-throughput grouped experts to Blackwell, and it's the headline act.

+10

python 160 shipped

@pytorch Aug 18

PYTORCH SLASHES B200 BENCH TIME 5X

The B200 microbenchmark suite just went from a 7.5-hour serial slog to a 143-minute parallel run, and that's only the start of what's moving across PyTorch today.

python 92 shipped

@pytorch Aug 17

PYTORCH CUTS MPS DEAD CODE, DREAMER V3 FIXES GPU MEMORY LEAK

PyTorch is aggressively pruning unused Metal Performance Shaders infrastructure while TorchRL patches a critical memory leak that kept loss tensors pinned to CUDA for entire training runs.

python 25 shipped

@pytorch Aug 16

PYTORCH CORE FIXES CUDA MEMORY LEAKS AND CHECKPOINT STORAGE SHARING

PyTorch shipped critical fixes for fake tensor leaks and checkpoint deserialization on meta devices, while the RL stack tackled sparse masked modes and routing tie-breaking across three repos.

python 31 shipped

Elsewhere on the wire

AI Agents Aug 20

LangChain ships core 1.6.0 with a security warning tucked inside

LangChain's newest core release fixes a tool-injection bug, but the fix itself widens the path for arbitrary code execution.

ai-agents 95 shipped

Local LLMs Aug 20

OLLAMA FIXES DEADLOCK THAT HUNG CHAT AND GENERATE

A mid-stream parser error in Ollama's server silently wedged chat and generate requests, leaking goroutines and hanging retries until a fix landed overnight.

+14

llms 268 shipped

@CachyOS Aug 20

LINUX-CACHYOS SHIPS 7.2.0-1

CachyOS's custom kernel just advanced to 7.2.0-1, and the sched-ext toolchain is moving in lockstep.

infra 55 shipped

Elixir & Phoenix Aug 20

LIVE VIEW HOOKS GET THE ELOQUENT TREATMENT

Phoenix Live View is shipping a passel of hook and JS-client fixes today, headlined by new beforeUpdate arguments that hand you more control over a hook's lifespan.