$ cat pytorch/week/2026-09-21.log
the week in review · Sep 21 – Sep 27, 2026
TorchTitan drops pure TP on routed experts
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
ExecuTorch stops publishing CUDA 12.6 wheels and retires the callable torch.compiler.precompile API the same week, so both are migration work.
[MoE] Deprecate pure TP on routed experts pytorch/torchtitan
MoE expert parallelism must now be at least the dense tensor-parallel degree, and routed-expert activations stay sequence-sharded under EP so routed-expert computation is not duplicated. If your TorchTitan config still sets pure TP across routed experts, it will be rejected; shared experts and other dense layers keep the old path.
Use stacked layout for fused feed-forward projections pytorch/torchtitan
W13 now stores W1 and W3 in a stacked layout rather than interleaved rows in a [2F, D] tensor. This is a reland: the original PR was accidentally merged into its ghstack base branch and never reached main, so the same reviewed change is now rebased onto current main.
[precompile] Make torch.compiler.precompile a module and retire the callable API (#197993) pytorch/pytorch
torch.compiler.precompile is now a module and the callable entry point is gone; it exports only types (Capture, MakeFxTracer, PrecompiledRunnable, PrecompileSummary, and a re-exported symbol). The landing note warns the commit is not usable on its own, so wait for the follow-up before porting.
Stop publishing CUDA 12.6 wheels (#23015) pytorch/executorch
PyTorch stopped offering CUDA 12.6 on its nightly channel, so no ExecuTorch nightly has built that train since early September; the newest published cu126 nightly is 1.5.0.dev20260906, while cpu, cu130, and cu132 keep building. Pin a supported train before your next nightly pull.
[GreenLight] Adds a triviality definition and requirement for GL policy (#8814) pytorch/test-infra
GreenLight now requires a PR to be both safe and trivial to land, with triviality defined as a claim about the review. Expect your automated merge candidates to be held back more often.
Route all triton_meta device properties through a CooR-aware helper (#196149) (#196149) pytorch/pytorch
compile_on_one_rank requires byte-identical generated code across ranks, so triton_meta must not carry a rank-specific device index. That dropping only existed in TritonKernel; every triton_meta device property now goes through a CooR-aware helper, which matters if you compile on one rank and replay elsewhere.
Added TorchAO NVFP4 weight-only quantization + Linear and FLUX examples pytorch/TensorRT
NVFP4 weight-only quantization (NVFP4WeightOnlyConfig) now runs on the Dynamo path, with torchao_trt.dequantize_nvfp4 mapped to a two-level TensorRT IDequantizeLayer: packed FP4 E2M1 plus swizzled FP8 E4M3 block scales at block size 16 and an FP32 per-tensor scale. Linear and FLUX examples are included; the MXFP4 recipe, which has no TorchAO weight-only config of its own, landed alongside it.
Qualcomm AI Engine Direct - HF LLM Optimization Stage 1 pytorch/executorch
Qualcomm's AI Engine Direct path takes a first step toward matching Static LLM accuracy and performance for HF models, scoped to Llama3.2 1B decode mode, with KV annotation among the work done. It is stage one of a series, so treat the numbers as a starting point.
$ ls pytorch/week/ # the briefings behind this review
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.