$ ls pytorch/reviews/ # the step-back reads
PyTorch in review
One review per week (and month): what shipped, what broke, what to act on.
Nov 2–8, 2026 weekly [plan] 8 stories
Inductor gains a flag to turn off FMA contraction
ExecuTorch fixes a two-thread deadlock, mixed-dtype cat, and host thread oversubscription, while mem-efficient SDPA moves to int64 key/va...
Oct 26–Nov 1, 2026 weekly [act] 8 stories
PyTorch 2.15.0 ships with rewritten NCCL backend
The release carries backwards incompatible changes and deprecations, and the week's other work was dominated by correctness and supply-ch...
Oct 19–25, 2026 weekly [plan] 8 stories
TensorDict stops silently dropping data on nested and lazy-stack paths
Six tensordict fixes close silent data loss in create_nested, sub-tensordict in-place ops and nested lazy stacks, alongside four new depr...
Oct 12–18, 2026 weekly [act] 8 stories
TensorDict 0.15.0 goes pure Python as a 690 KB wheel
The release carries 581 pull requests, and torchtitan's RL recipes cut rollout memory, worker imports, and controller GC pressure in the ...
Oct 5–11, 2026 weekly [act] 8 stories
torch.compile dtype fix, two Arm CVEs patched
Executorch closed CWE-125 in the Arm VGF decoder and CWE-22 in the Arm backend during the same week TorchTitan's RL losses turned out to ...
Oct 1–31, 2026 monthly [act] 8 stories
PyTorch 2.15.0 lands with rewritten NCCL backend
TensorDict 0.15.0 also shipped as a pure Python wheel, 1.77x faster on everyday operations, with removals across the release.
Sep 28–Oct 4, 2026 weekly [plan] 8 stories
Inductor correctness sweep lands with nextafter and index_put fixes
TorchTitan also drops graph-based EP chunking and ties Qwen3.5 embeddings, while ExecuTorch fixes transposed-conv BatchNorm fusion.
Sep 21–27, 2026 weekly [plan] 8 stories
TorchTitan drops pure TP on routed experts
ExecuTorch stops publishing CUDA 12.6 wheels and retires the callable torch.compiler.precompile API the same week, so both are migration ...
Sep 14–20, 2026 weekly 8 stories
Helion autotunes Metal, ending default 16x16x16 tiles
TorchRL fixed reward leakage and constructor device caching, while Dynamo and TorchTitan corrected silent correctness bugs.
Sep 7–13, 2026 weekly [act] 8 stories
TorchRL 0.14 lands with DreamerV3 and async collectors
CUDA 12.6 nightly wheels stop at PyTorch 2.15, which also drops Maxwell, Pascal, and Volta GPU support.
Sep 1–30, 2026 monthly [act] 8 stories
PyTorch 2.14.0 ships, TorchRL 0.14 lands
TensorDict 0.14.1 follows as a correctness patch, and SECURITY.md now extends the distributed trust model to checkpoints.
Aug 31–Sep 6, 2026 weekly [act]
PyTorch 2.14.0 ships, security docs clarified
MPS complex std/var fixes and CUDA graph lifecycle improvements also landed across the ecosystem.
Aug 24–30, 2026 weekly
Executorch stabilizes its build and strips legacy backends
Executorch overhauls its CI promotion and removes deprecated backends, while core PyTorch and its ecosystem libraries ship a steady strea...
Aug 1–31, 2026 monthly
ExecuTorch MoE quantization, Helion runtime splits, and a broader cleanup wave
August's work centered on ExecuTorch's MoE quantization pipeline, Helion's systematic runtime reorganization, and a broad cleanup of lega...
Keep up with PyTorch in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.