RepoJournal
PyTorch

@pytorch

PyTorch and the broader machine-learning ecosystem

Keep up with PyTorch in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

One email a day. Unsubscribe in one click.

Pick a date

  1. Tue 7 jul
  2. Wed 8 jul
  3. Thu 9 jul
  4. Fri 10 jul
  5. Sat 11 jul
  6. Sun 12 jul
  7. Mon 13 jul
  8. Tue 14 jul
  9. Thu 16 jul
  10. Sat 18 jul
  11. Sun 19 jul
  12. Mon 20 jul
  13. Tue 21 jul
  14. Fri 24 jul
  15. Sat 25 jul
  16. Sun 26 jul
  17. Mon 27 jul
  18. Wed 29 jul
  19. Sun 2 aug
  20. Mon 3 aug
  21. Sun 9 aug
  22. Mon 10 aug
  23. Sat 15 aug
  24. Sun 16 aug
  25. Mon 17 aug
  26. Tue 18 aug
  27. Wed 19 aug
  28. Thu 20 aug
  29. Sat 22 aug
  30. Today 23 aug
Topics: Python AI / ML Full archive →

The Wire · Showcase

PyTorch's Caffe2 cleanup hits a wall as internal dependencies force three reverts

By RepoJournal · Filed · About PyTorch

PyTorch MergeBot taking-lying-flat Animesh Jain shoumikhin apbose Bob Ren Joseph Loftin
7 people shipped this
PyTorch MergeBot @pytorchmergebot 3 cited taking-lying-flat @taking-lying-flat 3 cited Animesh Jain @anijain2305 1 cited shoumikhin @shoumikhin 1 cited apbose @apbose 1 cited Bob Ren @bobrenjc93 Joseph Loftin @jloftin-nv

PyTorch's attempt to shed its Caffe2 legacy just got blocked by internal dependencies, and the resulting reverts ripple across the cleanup effort.

Three reverts in a row [1][2][3] pulled back placeholder removals, with the merge bot citing that "lots of internal code still depends on this header" [1] and similar issues for timer.h [2]. Meanwhile, an Inductor patch tightens the serialization boundary , a fake tensor caching fix targets 325 uncached dispatches on Qwen3-8B [4], TensorRT adds a pin-drift detector across eleven files [5], and a probe loop fix handles a c10::Error change in PyTorch 2.13 [6]. Torchtitan flags that zero-valid-token RL batches still step the optimizer [7], and adds validation for MRoPE and activation memory budgets [8][9]. The Caffe2 reverts need coordination, but the core caching and validation fixes are safe to take.

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] Revert "Remove the caffe2/core/macros.h placeholder (#194230)" pytorch/pytorch
  2. [2] Revert "Remove caffe2/core/timer.h (#194228)" pytorch/pytorch
  3. [3] Revert "Remove unused caffe2/utils/string_utils (#185125)" pytorch/pytorch
  4. [4] [fake tensor] Cache _unsafe_view dispatch (#193765) pytorch/pytorch
  5. [5] Detect ExecuTorch pin drift across the repository ↗ pytorch/TensorRT
  6. [6] [fix] MD-TRT: handle c10::Error in bind_nccl_comm() probe loop ↗ pytorch/TensorRT
  7. [7] [Bug][RL] Zero-valid-token batches still step optimizer and advance policy version ↗ pytorch/torchtitan
  8. [8] [Bug] Validate MRoPE dimension contracts ↗ pytorch/torchtitan
  9. [9] [Bug] Validate activation memory budget (#4291) pytorch/torchtitan

Quick answers

What shipped in PyTorch on August 23, 2026?
PyTorch's attempt to shed its Caffe2 legacy just got blocked by internal dependencies, and the resulting reverts ripple across the cleanup effort. In total, 42 commits and 16 pull requests landed.
Who contributed to PyTorch on August 23, 2026?
7 developers shipped this update, including PyTorch MergeBot, Bob Ren, Animesh Jain, shoumikhin, apbose, Joseph Loftin, and taking-lying-flat.
What were the notable PyTorch updates?
Revert "Remove the caffe2/core/macros.h placeholder (#194230)", Revert "Remove caffe2/core/timer.h (#194228)", and Revert "Remove unused caffe2/utils/string_utils (#185125)".

More from @pytorch

PyTorch @pytorch

Pallas HBM Windows Open on Any Tensor Axis

Pallas stagers just gained the ability to align HBM windows on any tensor axis, not just the last one, and a critical autotuning crash is fixed.

AmesingFlank kitzbergerg ethche unography +9
python 81 shipped
PyTorch @pytorch

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.

Ivy Zhou cherry77-cloud Chien-Chin Huang felipemello1 +10
python 160 shipped
PyTorch @pytorch

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.

Ryan Zhang Jean Schmidt Yu, Guangye msnliu +10
python 85 shipped
PyTorch @pytorch

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.

zhangxiao Kaue Oliveira drisspg Frank Seide +9
python 92 shipped

Daily updates, in your inbox

Follow PyTorch

Keep up with PyTorch in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

One email a day. Unsubscribe in one click. Read a past issue →

Elsewhere on the wire

AI Agents AI Agents

Cline hubs now drain without dropping work

Cline's new hub drain and upgrade flow means you can restart a hub mid-session and lose nothing, thanks to durable event replay.

ump45nose neubig Krishna github-actions[bot] +1
ai-agents 18 shipped
Local LLMs Local LLMs

SGLang Fixes Pixtral 500s as vLLM Cuts TTFT 25%

The biggest news overnight: SGLang is patching a production-breaking Pixtral bug, and vLLM engineers found a way to shave up to 25% off time-to-first-token on Mamba models.

CISC ngxson Kartik Sirohi superm1 +7
llms 134 shipped
CachyOS @CachyOS

AIRE: ANANICY RULES FIX SESSION LAUNCHER SCHEDULING

ly-dm was being pinned to the idle scheduler, killing your login session's responsiveness.

github-actions Matias alexsmolya Peter Jung +2
infra 74 shipped
Elixir & Phoenix Elixir & Phoenix

JOSE VALIM FIXES THE CLASUE BUG

Jose Valim shipped two refactors plus the actual fix for that `none` clause compiler issue, and a UTC fast path in DateTime could save you a TZ DB lookup.

josevalim reisub
elixir 6 shipped
Want every project, not just this one?