PyTorch: Apple Silicon Hardening and Dynamo's Global State Cleanup
Mac backend work this cycle focused on correctness under virtualization and safer exception handling, while a large Dynamo stack fixed how compiled code tracks live globals and module state. Test suite cleanup on MPS also continued, trimming redundant coverage now handled by OpInfo 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: 2026-09-13T13:00:26Z
- 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 day, and welcome to PyTorch, your developer briefing for September 13th, 2026.
The clearest throughline today is a hardening pass on the Apple GPU backend, paired with a deep Dynamo stack that fixes how compiled code tracks live global state.
Start with MPS. Nikita Shulga's PR 196881 found that virtualized macOS exposes a paravirtual device with no real GPU family, and PyTorch was still handing it Metal Performance Primitives kernels it couldn't build. The fix gates that path properly and adds an environment override so developers can force Metal 3 or…
Second theme: bobrenjc93's four-PR Dynamo stack, numbers 196895 through 196897, rethinks how compiled artifacts handle guarded globals. The core problem: compiled bytecode and its guards could disagree about the same global variable, module aliases could go stale mid-trace, and package uninstall logic sometimes…
Elsewhere, a correctness fix landed for `sinc`'s second derivative near zero, and InstanceNorm now raises consistently on channel mismatches instead of silently warning.
What's next: watch for the CUDA 13.0 docker bump to 13.0.3, which also reverts an earlier test skip now that the root toolkit mismatch is fixed.
Nearby episodes from PyTorch
- Hardening the Compile-and-Ship Pipeline
- Distributed Training Takes Control of Its Own Timing
- Guard Serialization Gets a Precision Pass
- Inductor's Reduction Fusion Overhaul
- Hardware Independence Takes Center Stage
- Hardware-Agnostic Cleanup and MPS Correctness Push
- Weekly Recap - Multi-Rank Compilation and Backend Correctness
- One Rank to Compile Them All