Who contributed to PyTorch on May 25, 2026?

2 developers shipped this update, including roman-janik-nxp and Xia-Weiwen.

What were the notable PyTorch updates?

[dynamo] mimic tp_richcompare handling (#182759), [Test] Make TestFP8Matmul device-generic for out-of-tree accelerators (#184771), and [TEST]Make test_modes.py device-generic: Replace GPU_TYPE constant with torch.accelerator API (#179893).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

TORCH.COMPILE GAINS PYTHON COMPARISON OPERATORS, TESTS GO DEVICE-AGNOSTIC ACROSS THREE REPOS

By RepoJournal · Filed 06:02 UTC on May 25, 2026 · About PyTorch

2 people shipped this

roman-janik-nxp @roman-janik-nxp 1 cited

Xia-Weiwen @Xia-Weiwen 1 cited

Dynamo now handles Python's richcompare protocol, closing a critical gap in torch.compile's operator support while the ecosystem pivots hard toward device-generic testing.

The biggest move overnight: PyTorch core shipped richcompare support [1] by implementing CPython's full do_richcompare dispatch on VariableTracker, meaning torch.compile can now handle all comparison operators the way Python does. This fixes a class of compile failures that hit when models use comparison chains or custom __eq__ implementations.

In parallel, the test infrastructure is reshaping around device generics. TestFP8Matmul [2] got device-agnostic by replacing 37 hardcoded cuda references with a device parameter, and test_modes.py [3] ditched GPU_TYPE constants for torch.accelerator APIs. This matters: out-of-tree accelerators can now run quantization and dynamo tests without forking test files. ExecutorTorch's NXP backend [6] is already moving in lockstep with Neutron MLIR flow tests for tensor ops.

Two infrastructure cleanups landed: dead THNN/UCC/TH CMake references [4] got removed, and graph partition buffer signatures [5] now preserve globals correctly. The ao repo [7] added a defensive check to skip PT2E inductor registration when mkldnn isn't available, preventing silent failures on CPU builds.

Action items

→ Update your torch.compile tests if they mock comparison operators; richcompare dispatch may change behavior [ref:3] pytorch/pytorch [plan]
→ Audit custom test decorators; device-generic migration is rolling out and hardcoded device checks will collide [ref:4][ref:6] pytorch/pytorch [plan]
→ Check your CPU quantization pipelines; ao's mkldnn guard prevents bad registrations [ref:1] pytorch/ao [monitor]

References

[1] [dynamo] mimic tp_richcompare handling (#182759) pytorch/pytorch
[2] [Test] Make TestFP8Matmul device-generic for out-of-tree accelerators (#184771) pytorch/pytorch
[3] [TEST]Make test_modes.py device-generic: Replace GPU_TYPE constant with torch.accelerator API (#179893) pytorch/pytorch
[4] Remove dead THNN/UCC/TH references from CMakeLists (#185062) pytorch/pytorch
[5] Fix graph partition removed buffer signatures (#184004) pytorch/pytorch
[6] NXP backend: Enable Add Tensor with new Neutron flow ↗ pytorch/executorch
[7] [CPU] skip registration of PT2E inductor passes if mkldnn not available ↗ pytorch/ao

Quick answers

What shipped in PyTorch on May 25, 2026?: Dynamo now handles Python's richcompare protocol, closing a critical gap in torch.compile's operator support while the ecosystem pivots hard toward device-generic testing. In total, 31 commits and 2 pull requests landed.
Who contributed to PyTorch on May 25, 2026?: 2 developers shipped this update, including roman-janik-nxp and Xia-Weiwen.
What were the notable PyTorch updates?: [dynamo] mimic tp_richcompare handling (#182759), [Test] Make TestFP8Matmul device-generic for out-of-tree accelerators (#184771), and [TEST]Make test_modes.py device-generic: Replace GPU_TYPE constant with torch.accelerator API (#179893).

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.