$ the-wire · showcase
Faster RL transport, quieter graph capture
By RepoJournal · Filed · About PyTorch · Composed from the cited sources · methodology
torchtitan's RL path now reads staged pieces packed and defaults to ModelStateDictTransport, while the cudagraph stack drops a racy empty_cache() call before capture.
[rl] Default to ModelStateDictTransport (#6125) pytorch/torchtitan
ModelStateDictTransport, which uses one-sided torch.distributed._transport reads and replaced TorchStore for cpu_prefetch, was opt-in; the author writes that it "beat TorchStore on every measured axis". The stack is ghstacked with #6127 on top, so this is a default flip rather than a new code path.
[rl] Read staged pieces packed instead of restaging on the trainer (#6097) pytorch/torchtitan
Staged DTensors are now read packed instead of being restaged on the trainer: previously each staged piece triggered a redistribute collective per tensor whenever a staged value was needed. Cuts the collective traffic in the RL trainer path.
[cudagraph] Remove the unsynchronized empty_cache() before accelerator graph capture (#6131) pytorch/torchtitan
The unsynchronized empty_cache() before accelerator graph capture is gone, closing a race for users who do not call a sync after step() and call empty_cache() before the next step(). The summary notes TorchTitan main already syncs at the end of each step, so the exposure is in caller code that skips it.
[library] Avoid loading unrelated modules during fake registration (#204241) pytorch/pytorch
Fake kernel registration no longer fails when inspect loads unrelated modules while reading diagnostic metadata from its caller's frame. Tests cover filename and line number, stack levels, named and anonymous callers, and the existing torchvision exception, with no intended public API change.
[Release Portal] Show 1.3, 1.7, 2.10, 2.11, 4.8, 5.3 as information; drop 3.9, 5.2; check nightlies (1.5) and release-only changes (1.11) (#9547) pytorch/test-infra
Also across the stack: the Release Portal gained a per-release versions field for triton, torchvision and torchaudio plus checks on the triton release branch and PyPI, demoted steps 1.3, 1.7, 2.10, 2.11, 4.8 and 5.3 to informational, dropped 3.9 and 5.2, and began checking nightly binary builds; test-infra now mirrors nvidia-nvshmem-cu13 from PyPI so 3.8.0 reaches download.pytorch.org, where py...