PyTorch: Undefined Contracts and the MPS Coverage Push
Today's activity centers on making implicit assumptions explicit—undefined gradient strides, silent NaN inconsistencies, and moved-from argument stacks all got documented or fixed—while a single contributor kept closing MPS backend gaps with eight new operator implementations.
Duration: PT2M38S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-08-29T13:00:36Z
- Audio duration: PT2M38S
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 August 29th, and today's PyTorch activity has a clear throughline: quietly wrong assumptions getting caught and documented.
Start with correctness edge cases. Jason Ansel's PR 195210 documents that backward gradients can have undefined strides — this came directly from a real bug where torch dot cat's backward silently fed a non-contiguous view into a kernel that assumed contiguous layout, and neither gradcheck nor opcheck caught it. In…
A related thread is exposing state that's already been consumed or freed. Georgia Phillips's PR 195213 fixes nativert error dumps that printed "None" for list arguments because unboxing had already moved them out of the stack — the fix rebuilds the stack from the execution frame after the fact. Daniel Ohayon's PR…
Meanwhile, Benjamin Demaille single-handedly kept shipping MPS coverage — fractional max pool in two and three dimensions, segment reduce, adaptive average pool 3D, linalg eig, and two margin loss variants, all previously CPU, CUDA, and XPU only. That's the most tests-per-PR density in the batch.
Worth flagging: a reverted MXFP6 packing fusion PR and a reverted "drop unused includes" cleanup both broke internal builds, a…
W…
Nearby episodes from PyTorch
- A New CuteDSL Reduction Engine and the Great Test Untangling
- ROCm Ten Point Oh and the Precompile Stack
- Hardening the Merge Pipeline and the Attention Stack
- FlyDSL Lands and the GEMM Compiler Stack Grows
- AOT Autograd Tightens Its Grip on Backward Tangents
- Weekly Recap - Compiler Correctness and the Long Tail of Cleanup
- Precompile Guards Get a Rewrite
- Precompile Hardening and the Great Cleanup