Who contributed to PyTorch on August 18, 2026?

13 developers shipped this update, including zhangxiao, Kaue Oliveira, drisspg, Frank Seide, Ivy Zhou, Chien-Chin Huang, jinsooihm, and github-actions[bot], and 5 more.

What were the notable PyTorch updates?

[CI] Run B200 operator microbenchmarks per operator (#193739), [qnnpack] Add 1..4x8 dynamic-quant aarch64 q8gemm microkernel for small batch (#193595) (#193595), and Add Inductor kernel benchmark dashboard.

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

PYTORCH SLASHES B200 BENCH TIME 5X

By RepoJournal · Filed 06:59 UTC on August 18, 2026 · About PyTorch

13 people shipped this

Ivy Zhou @ivy-zhou 2 cited

github-actions[bot] @github-actions[bot] 2 cited

drisspg @drisspg 1 cited

Frank Seide @frankseide 1 cited

Chien-Chin Huang @fegin 1 cited

jeanschmidt @jeanschmidt 1 cited

karthickai @karthickai 1 cited

theap06 @theap06 1 cited

gtnv @gtnv 1 cited

zhangxiao @littlexiaozhang

Kaue Oliveira @kyuubyN

jinsooihm @jinsooihm

morrison-turnansky @morrison-turnansky

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.

PyTorch's own CI flag is the rare internal win that pays off externally: the B200 operator microbenchmark workflow now runs per-operator in parallel instead of one 7.5-hour serial job, cutting wall time to roughly 143 minutes [1]. That frees up GPU time and makes regression detection faster for everyone. On the kernel side, a new aarch64 Q8GEMM microkernel targets batch sizes 1 to 4 for qlinear_dynamic, and in the words of the PR, aim is to 'unblock shipping an update to Hypernova EMG Handwriting' [2] by eliminating the 5/8 compute waste at batch size 3. Meanwhile, test-infra ships a fresh Inductor kernel benchmark dashboard defaulting to CUDA on B200 with lower-is-better latency policies [3], so you can watch these kernel changes land. Across torchtitan, the team is telegraphing a CLI purge: 'We plan to remove most CLI options, especially the options related to models, optimization, and parallelism, in the future' [4], and they're shepherding the checkpointing manager toward a clean torch_checkpointing path [5][6]. Over in rl, DreamerV3's training semantics got a real fix, and the replay buffer now rejects never-written slots, which will save you from silent corruption [7][8]. Two test-infra releases (v20260817-220016, v20260818-012344) are out, and the offline sweep now spares live pet runners from deregistration [9][10][11].

Action items

→ Upgrade to the latest test-infra release (v20260818-012344) to get the pet-runner sweep fix pytorch/test-infra [immediate]
→ Watch the Inductor kernel dashboard for regressions on B200 after recent kernel changes pytorch/test-infra [monitor]
→ Plan for the torchtitan CLI removal by migrating configs to Trainer.Config pytorch/torchtitan [plan]
→ Pull in the DreamerV3 and replay-buffer fixes in torchrl if you train with those features pytorch/rl [plan]

References

[1] [CI] Run B200 operator microbenchmarks per operator (#193739) pytorch/pytorch
[2] [qnnpack] Add 1..4x8 dynamic-quant aarch64 q8gemm microkernel for small batch (#193595) (#193595) pytorch/pytorch
[3] Add Inductor kernel benchmark dashboard ↗ pytorch/test-infra
[4] [RFC][Config] Add torchtitan_configs for full training configurations (#4114) pytorch/torchtitan
[5] Add torch_checkpointing save config plumbing (#4058) pytorch/torchtitan
[6] Move the checkpointing-disabled guard into BaseCheckpointManager (#4173) pytorch/torchtitan
[7] [BugFix] [Performance] fix dreamer v3 training semantics ↗ pytorch/rl
[8] [BugFix] Reject invalid ReplayBuffer generations ↗ pytorch/rl
[9] v20260818-012344 ↗ pytorch/test-infra
[10] v20260817-220016 ↗ pytorch/test-infra
[11] Skip offline pet runners during scale-down offline sweep ↗ pytorch/test-infra

Quick answers

What shipped in PyTorch on August 18, 2026?: 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. In total, 63 commits, 27 pull requests, and 2 releases landed.
Who contributed to PyTorch on August 18, 2026?: 13 developers shipped this update, including zhangxiao, Kaue Oliveira, drisspg, Frank Seide, Ivy Zhou, Chien-Chin Huang, jinsooihm, and github-actions[bot], and 5 more.
What were the notable PyTorch updates?: [CI] Run B200 operator microbenchmarks per operator (#193739), [qnnpack] Add 1..4x8 dynamic-quant aarch64 q8gemm microkernel for small batch (#193595) (#193595), and Add Inductor kernel benchmark dashboard.

Pytorch Fixes XPU Dispatch Gap and CI OOMs

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

+10

python 85 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

@pytorch Aug 15

TENSORDICT 0.14 LANDS WITH BREAKING API DEFAULTS, TORCHTITAN GAINS ASYNC-TP BACKEND

TensorDict shipped 0.14.0 with backward-incompatible changes to `to_module()` and `copy_at_()` defaults, forcing TorchRL to patch every call site before its own 0.14 release.

+12

python 84 shipped

Elsewhere on the wire

AI Agents Aug 19

CLINE DESKTOP GOES NATIVE WITH BACKGROUND ALERTS

Cline's desktop app just became a tool you can leave running while you grab coffee, and LangChain is quietly fixing the token math that makes DeepSeek caching actually count.

ai-agents 97 shipped

@CachyOS Aug 19

CACHYOS PUSHES KERNEL 7.1.8 AND SCHEDULER FIXES ACROSS THE STACK

CachyOS is shipping a new kernel meanwhile patching its scheduler stack, and that combination is exactly what you want before your next deploy.

infra 58 shipped

Elixir & Phoenix Aug 19

LiveView memory leak fix leads overnight Elixir cleanup

A handful of pull requests across Elixir and Phoenix are quietly fixing leaks and edge cases that have been nagging at production apps.

elixir 19 shipped

@dotnet Aug 19

MAUI Overhauls NuGet Publishing as FSharp Preps for VS 18.11

MAUI is replacing its custom NuGet publishing with Microsoft's 1ES.PublishNuget, and FSharp is rolling Microsoft.Build deps forward to VS 18.11, two moves that reshape release engineering across the .NET ecosystem.