What were the notable PyTorch updates?

Revert "Remove named tensor (#173895)", Fix scoped Chromium event reset (#184973), and [library] Improve infer_schema error message when future annotations cause NameError (#185694).

@pytorch

PyTorch and the broader machine-learning ecosystem

github ↗

Pick a date

Topics: Python AI / ML Full archive →

The Wire · Showcase

PYTORCH ROLLS BACK NAMED TENSOR REMOVAL, FIXES CHROMIUM PROFILER LEAK

By RepoJournal · Filed 06:02 UTC on June 1, 2026 · About PyTorch

PyTorch reverted a breaking change that tanked internal signals, while the core team patched a profiler bug that was nuking outer event context.

The named tensor removal [1] is back after it broke internal infrastructure, a significant signal that the community still depends on this API. In parallel, the profiler team fixed a scoped Chromium event bug [2] where exiting a nested profiling context was wiping the entire thread-local event stack, collapsing outer events in the process. On the schema front, error messages for future annotations now surface the actual NameError [3] instead of swallowing it behind a bare exception handler, making debugging type resolution issues dramatically less painful. Inductor's accumulate_grad op got a state modeling fix [4] that makes functionalization explicit: the op now takes current grad as input and returns updated grad, killing hidden Tensor.grad reads that were breaking compiled autograd. The XPU team also standardized their oneDNN integration [5], swapping C API calls for C++ equivalents in DnnlExt.h to eliminate the inconsistent mixing of styles. TensorRT's dependency lockfile got a routine update.

Action items

→ Test internal code paths that use named tensors before next release cycle pytorch/pytorch [plan]
→ Review any custom profiling contexts wrapping chromium_event_timed calls pytorch/pytorch [monitor]
→ Validate accumulate_grad behavior in compiled autograd workflows pytorch/pytorch [plan]

References

[1] Revert "Remove named tensor (#173895)" pytorch/pytorch
[2] Fix scoped Chromium event reset (#184973) pytorch/pytorch
[3] [library] Improve infer_schema error message when future annotations cause NameError (#185694) pytorch/pytorch
[4] Fix functional accumulate_grad state modeling (#184082) pytorch/pytorch
[5] [xpu] Refactor OneDNN C API to C++ API (#184486) pytorch/pytorch

Quick answers

What shipped in PyTorch on June 1, 2026?: PyTorch reverted a breaking change that tanked internal signals, while the core team patched a profiler bug that was nuking outer event context. In total, 39 commits landed.
What were the notable PyTorch updates?: Revert "Remove named tensor (#173895)", Fix scoped Chromium event reset (#184973), and [library] Improve infer_schema error message when future annotations cause NameError (#185694).

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.