PyTorch: The Great Runner Migration, and a Distributed Rollback
The infrastructure team finished ripping the EC2 test path and "use-arc" flag out of core CI workflows, completing the shift to the OSDC/ARC fleet, while two distributed process-group PRs got reverted for breaking trunk tests.
Duration: PT2M42S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-07-08T13:01:08Z
- Audio duration: PT2M42S
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 8th, 2026, and today's codebase activity centers on one big cleanup finishing, and one distributed feature rollback still settling.
The lead story is Huy Dhn's multi-PR stack finishing off the OSDC and ARC migration. PR 189171 removes the "use-arc" input entirely from the Linux build, test, and docs workflows, because every EC2 path is now gone -- OSDC is the only option left. That builds on 189170 and 189169, which deleted the legacy EC2 jobs…
Second theme: distributed process groups had a rough day. Three separate reverts landed -- "Support out-of-order ranks in split group" (189090), its CPU-only gloo counterpart (189103), and the PyObject Dispatch change (187949) -- all backed out after breaking trunk tests. Ezyang's PR 189231 is the follow-up, fixing…
Smaller but notable: MPS users get real wins today. Isalia20's LU factor and LU solve migrations to Metal kernels show speedups up to two orders of magnitude on batched inputs. And qflen's fix for catastrophic cancellation in gelu and normal CDF calculations corrects silently wrong results for negative inputs --…
What's next: watch for the TPU workflow to stabilize, and expect the split-group revert to…
Nearby episodes from PyTorch
- Distributed Cleanup and CI Infrastructure Overhaul
- Correctness Fixes Sweep the Compiler Stack
- Weekly Recap - Release Hardening and Correctness Cleanup
- Determinism and Correctness Sweep
- Silent Correctness Bugs Take Center Stage
- FlexAttention's Rough Week
- Accelerator Backends and Memory Management
- Weekly Recap - Release Stabilization & Core Improvements