Who contributed to PyTorch on July 25, 2026?

9 developers shipped this update, including atalman, Cynthia Yang, Michael Lazos, Akshay Madavalappil Ramesh, Benjamin Glass, acisseJZhong, jinsooihm, and github-actions[bot], and 1 more.

What were the notable PyTorch updates?

[docs-preview] Stop sourcing untrusted artifact metadata (OIDC/shell injection) (#191017), [fx] _rename_unbacked_to: unify instead of assert on unbacked dest (#190083) (#190083), and [inductor][NVGEMM] Enable scaled-GEMM subprocess precompile (#189841).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

CRITICAL OIDC INJECTION IN DOCS PREVIEW WORKFLOW PATCHED

By RepoJournal · Filed 06:59 UTC on July 25, 2026 · About PyTorch

9 people shipped this

atalman @atalman 2 cited

Cynthia Yang @ZixinYang 1 cited

Michael Lazos @mlazos 1 cited

acisseJZhong @acisseJZhong 1 cited

jinsooihm @jinsooihm 1 cited

izaitsevfb @izaitsevfb 1 cited

Akshay Madavalappil Ramesh @AMR5210

Benjamin Glass @benjaminglass1

github-actions[bot] @github-actions[bot]

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

The `upload-docs-preview.yml` workflow ran with `id-token: write` and directly sourced an artifact produced by a fork-triggerable job, meaning untrusted PR code could inject shell commands and forge OIDC tokens [1]. Fix landed. Separately, the compiler team unified unbacked symbol handling in FX to stop crashing on re-exports of already-lowered programs [2], while Inductor unblocked scaled-GEMM subprocess precompile by fixing two RecursionErrors in quantization constraints [3]. On the infra side, test-infra added per-trunk-commit workflow-duration KPI tracking to HUD [6], then reverted and re-applied numpy preview-build support for Python 3.15 wheels [7]. TorchTitan shipped window FIFO scheduling for RL training to prevent head-of-line blocking in rollout buffers [4], and lifted PTRR load balancer restrictions so flex-attention models with multiple named masks can now run [5]. Activity this period: 56 commits, 9 PRs, 1 release across pytorch, torchtitan, and test-infra.

Action items

→ Review docs-preview workflow changes; verify no fork-triggered workflows source untrusted artifacts pytorch/pytorch [immediate]
→ Update to latest FX unbacked-symbol handling if you re-export lowered programs pytorch/pytorch [plan]
→ Monitor TorchTitan RL tests; window FIFO changes scheduler behavior pytorch/torchtitan [monitor]

References

[1] [docs-preview] Stop sourcing untrusted artifact metadata (OIDC/shell injection) (#191017) pytorch/pytorch
[2] [fx] _rename_unbacked_to: unify instead of assert on unbacked dest (#190083) (#190083) pytorch/pytorch
[3] [inductor][NVGEMM] Enable scaled-GEMM subprocess precompile (#189841) pytorch/pytorch
[4] [rl] add window FIFO scheduling ↗ pytorch/torchtitan
[5] [cp] Allow PTRR load balancer to select a mask from a dict (#3972) pytorch/torchtitan
[6] Add "pull workflow duration per trunk commit" KPI to HUD ↗ pytorch/test-infra
[7] Build numpy for preview Python (3.15/3.15t) and host on download.pytorch.org ↗ pytorch/test-infra

Quick answers

What shipped in PyTorch on July 25, 2026?: PyTorch's docs-preview CI trusted fork-controlled artifacts in a context with token-write permissions, exposing the entire build pipeline to code injection. In total, 56 commits, 9 pull requests, and 1 releases landed.
Who contributed to PyTorch on July 25, 2026?: 9 developers shipped this update, including atalman, Cynthia Yang, Michael Lazos, Akshay Madavalappil Ramesh, Benjamin Glass, acisseJZhong, jinsooihm, and github-actions[bot], and 1 more.
What were the notable PyTorch updates?: [docs-preview] Stop sourcing untrusted artifact metadata (OIDC/shell injection) (#191017), [fx] _rename_unbacked_to: unify instead of assert on unbacked dest (#190083) (#190083), and [inductor][NVGEMM] Enable scaled-GEMM subprocess precompile (#189841).

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

@pytorch 6 days ago

HELION BACKEND REFACTOR UNLOCKS MULTI-DEVICE COMPILATION; DYNAMO FIXES FORWARD-AD TANGENT BUG

Helion shipped a four-part backend architecture overhaul that finally separates printer, lowering, and backend registration logic by device, eliminating the need to edit 17 core files for every new backend.

python 51 shipped 2-min read

Elsewhere on the wire

AI Agents about 5 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 5 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.