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

The Wire · Showcase

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

By RepoJournal · Filed · About PyTorch

vmoens theap06 anijain2305 choijon5 Sunghyun Choi 1fanwang Kurt Mohler
16 people shipped this
vmoens @vmoens 2 cited theap06 @theap06 1 cited anijain2305 @anijain2305 1 cited choijon5 @choijon5 1 cited Sunghyun Choi @froggy-hyun 1 cited 1fanwang @1fanwang 1 cited Kurt Mohler @kurtamohler 1 cited gtnv @gtnv Jay Prajapati @coder-jayp weifengpy @weifengpy Pian Pawakapan @pianpwk tarinduj @tarinduj AmesingFlank @AmesingFlank peterdsharpe @peterdsharpe Rahul Agrawal @rahul342 ethanwee1 @ethanwee1

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.

TensorDict 0.14.0 flipped `to_module()` to preserve module parameters by default [1], a breaking change that required TorchRL to add explicit `preserve_module_state=False` parameters across its functional param-swap machinery [2]. The new version also ships Zarr-backed storage, direct `backward()` through TensorDict, and tightened memmap pickle semantics [1]. In parallel, TorchRL 0.14.0 is pinning TensorDict 0.14 while landing a critical fix for replay buffer checkpointing that now persists prefetch queue state across `state_dict` round-trips and pickle, keeping sample order and RNG state aligned [3]. On the training infrastructure front, TorchTitan relanded async-TP backend support without compiler overhead, folding tensor-parallel all-gather and reduce-scatter into adjacent GEMMs rather than running them sequentially [4], while Helion upgraded CuTe DSL to 4.7.0 with no backend code changes required and posted fresh B200 attention benchmarks [5]. Core PyTorch shipped MPS native int8 matrix multiplication support [6], fixed a weight_norm CPU kernel null-pointer crash on empty tensors [7], and patched MPSProfiler backward compatibility after a recent `stream` argument broke torchvision [8].

One email a day. Unsubscribe in one click.

Action items

References

  1. [1] TensorDict v0.14.0 ↗ pytorch/tensordict
  2. [2] [Versioning] Bump 0.14.0 ↗ pytorch/rl
  3. [3] [BugFix] Fix checkpointing of replay buffer prefetch state ↗ pytorch/rl
  4. [4] [Reland] Add dist GEMM attention and FFN backend ↗ pytorch/torchtitan
  5. [5] Upgrade CuTe DSL to 4.7.0 ↗ pytorch/helion
  6. [6] [MPS] Add native _int_mm support (#193153) pytorch/pytorch
  7. [7] Fix nan g gradient and null pointer in weight_norm CPU kernel for empty input (#192179) pytorch/pytorch
  8. [8] [MPS] Fix MPSProfiler `stream` arg backward compatibility (#193611) pytorch/pytorch

Quick answers

What shipped in PyTorch on August 15, 2026?
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. In total, 65 commits, 18 pull requests, and 1 releases landed.
Who contributed to PyTorch on August 15, 2026?
16 developers shipped this update, including theap06, vmoens, gtnv, Jay Prajapati, anijain2305, weifengpy, Pian Pawakapan, and choijon5, and 8 more.
What were the notable PyTorch updates?
TensorDict v0.14.0, [Versioning] Bump 0.14.0, and [BugFix] Fix checkpointing of replay buffer prefetch state.

More from @pytorch

PyTorch @pytorch

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.

aswanth-07 dependabot Rohit Singh Rathaur Animesh Jain +6
python 31 shipped
PyTorch @pytorch

DYNAMO TIGHTENS PYTHON SEMANTICS AS PYTORCH SHIPS TEST REFACTOR WAVE

PyTorch's compiler caught three attribute-handling bugs in a single commit batch, closing loopholes where graph breaks leaked through collections.deque writes and unbound comparison operations.

Piotr Bielak Jason Ansel Guilherme Leobas Gasoonjia
python 23 shipped
PyTorch @pytorch

PYTORCH FUSES MATMUL REDUCTION INTO TRITON, FIXES HYDRA 1.4 COMPATIBILITY

PyTorch's distributed matmul just got faster with a kernel fusion that eliminates the final memory-bandwidth bottleneck, while TorchRL and TensorDict harden their infrastructure across the stack.

theap06 Animesh Jain Wei Wang KarhouTam +4
python 30 shipped
PyTorch @pytorch

EXECUTORCH BACKENDS PATCH METADATA BUG, PYTORCH FIXES TEST BLOAT

Qualcomm and Arm backends shipped critical fixes to restore partition metadata and tame test runtime bloat across GPU CI.

winskuo-quic bdemirb DannyYuyang-quic Huy Do +3
python 17 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

OPENHANDS PINS CLIENT DEPENDENCY, ADDS STRUCTURED ERROR HANDLING

OpenHands locked its TypeScript client to exact version 1.37.0 and merged canvas error outcome improvements in a single commit cycle.

neubig
ai-agents 2 shipped
Local LLMs Local LLMs

LLAMA.CPP OVERHAULS INFERENCE FLAGS, VLLM PATCHES DEEPSEEK SPARSE MLA

llama.cpp deprecates mmap flags in favor of unified load-mode, while vllm fixes critical issues blocking DeepSeek-V4-Flash across all decode paths.

github-actions[bot] dhiltgen Fathi Boudra Eric Zhang +11
llms 136 shipped
CachyOS @CachyOS

DASHBOARD SEO OVERHAUL SHIPS AS PACKAGE TRAIN ROLLS FORWARD

CachyOS public-dashboard refactored routes with canonical metadata and improved crawler visibility [ref:18] while 61 commits across five repos kept the AUR pipeline humming.

github-actions witchlliee Jinra azdanov +1
infra 72 shipped
Elixir & Phoenix Elixir & Phoenix

ELIXIR CORE CUTS ALLOCATIONS ACROSS STRING AND AST OPERATIONS

Thomas Cioppettini landed four performance wins overnight that eliminate unnecessary allocations in List.myers_difference, String case conversion, and Macro tree walkers.

Thomas Cioppettini LostKobrakai
elixir 17 shipped
Want every project, not just this one?