PyTorch: Correctness Fixes and Reverts Take Center Stage
A cluster of correctness fixes landed across CUDA graphs, group norm, embedding bag, and RNG code, while two contentious changes—the XNNPACK update and a ROCm CI image switch—got reverted again after breaking things. Meanwhile, a multi-PR stack from Guilherme Leobas continues restructuring Dynamo's object protocol internals for maintainability.
Duration: PT2M52S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-07-17T13:00:35Z
- Audio duration: PT2M52S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's July 17th, 2026, and today's PyTorch activity has a clear throughline: correctness bugs found and fixed, and a couple of risky changes pulled back out.
Start with the reverts. XNNPACK's update to a newer github revision has now been reverted twice — PR 190282 undoes it again after internal failures, following the same pattern from PR 189871. And the ROCm CI image switch away from conda, PR 189760, got reverted after causing regressions on ROCm jobs. Two examples of…
On the correctness front, there's real signal. Benjaminglass1's stack — PRs 190243 through 190246 — cleans up native group norm, fixing precision errors from missed upcasts and null-pointer checks that were more verbose than necessary. Separately, PR 190264 fixes a real bug in embedding bag's CPU backward pass,…
RNG is also getting real attention. Guangyey's PR 190315 refactors torch.random and its docs, while jbschlosser's PR 190253 introduces a new stateless RNG API for raw bits, and weifengpy's PR 190289 adds stateful RNG replay for distributed local tensors. Three independent efforts converging on the same subsystem…
Finally, Guilherme Leobas has five stacked PRs — 190257 through 190261 — reworking…
W…
Nearby episodes from PyTorch
- One Engineer, Many Kernels
- Weekly Recap - Cleanup, Correctness, and a Random Number Overhaul
- Cleanup Sweeps and a New RNG Foundation
- Hardening Memory and Lifetime Management Across GPU Backends
- Distributed Backends and Overflow Fixes Take Center Stage
- Packaging Fixes and Bounds-Check Sweep
- Autograd Overhead Cuts and a New GEMM Backend Migration
- Hardening Boundary Checks and MPS Correctness