PyTorch: FlyDSL Lands and the GEMM Compiler Stack Grows
A large FlyDSL feature stack shipped across export, Inductor, and AOTI, giving custom GEMM kernels a full compile-to-package pipeline, while parallel work tightened NVFP4 GEMM autotuning, cache correctness, and cross-platform build health.
Duration: PT3M12S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-08-25T13:00:33Z
- Audio duration: PT3M12S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good day, it's August 25th, and here's what mattered in PyTorch overnight.
The dominant story is FlyDSL, a new kernel path getting wired end to end. Six related pull requests from garroud build out the whole pipeline: an opaque operator boundary for custom ops in PR 194637, export-time wrapping in PR 194636, decomposition control in PR 194639, Inductor lowering of captured launchers in PR…
Second theme: GEMM performance and correctness are getting more rigorous. Mlazos's NVGEMM stack - PRs 194648, 194649, 194651, and 194652 - fuses scaled GEMM output scaling, improves NVFP4 inference autotuning with cold-weight prefetch, and hardens the CuTeDSL compile protocol under monitoring. A related commit…
Third theme: caching and cudagraph correctness. Desertfire's PRs 194632 and 194633 fix a case where AOTAutograd cache could reuse a stale cudagraph decision across forward and backward, and make cudagraph partitioning aware of subgraph internals instead of treating them as opaque blocks. Separately, ad8e's PR 194658…
Smaller but concrete fixes: an MPS pinned-memory race with in-flight GPU buffers, PR 194662; MPS reductions now failing loudly instead of returning silently wrong results, PR…
Wha…
Nearby episodes from PyTorch
- 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
- Precompile's Guard Overhaul
- Compiler Stacks, Symmetric Kernels, and Build Hygiene
- Precompile Grows Up, and the Great Typo Sweep
- Correctness Cleanup and a Suspicious Cluster