PyTorch: Compiler Stacks, Symmetric Kernels, and Build Hygiene
A cluster of ghstack-based PRs pushed compiled Muon optimizer speedups and standalone graph export forward, while a separate wave of changes tightened CUDA build targeting and test device-agnosticism across FSDP and inductor.
Duration: PT2M51S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-08-20T13:00:38Z
- Audio duration: PT2M51S
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 20th, and today's PyTorch activity centers on making the compiler stack do more, while trimming what actually gets built and tested.
The biggest connected effort is Mike Lazos's five-PR stack — 194177 through 194181 — adding a faster symmetric matrix product kernel for SM100 and SM110, teaching Inductor to recognize X times X transpose, and wiring compiled Muon optimizer updates through foreach map so batches of Newton-Schulz calculations run…
The second theme is build and dispatch hygiene, mostly from cyyever. Three separate PRs — 194195, 194213, and 194218 — stop CUDA kernels for memory-efficient attention, vectorized elementwise ops, and rowwise FP8 from being instantiated for architectures the build can never target. That shrinks binary size and build…
Third, a batch of eight FSDP test PRs from danieldale2026 — covering ignored modules, hybrid shard, overlap, multiple wrapping, multiple forward, optimizer state, and gradient accumulation — all decouple tests from hardcoded accelerator selection in favor of the device-test harness. Paired with can-gaa-hou's ongoing…
Worth flagging individually: PR 194015 fixes MPS matmul silently writing garbage into sliced…
Wh…
Nearby episodes from PyTorch
- 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
- Precompile Grows Up, and the Great Typo Sweep
- Correctness Cleanup and a Suspicious Cluster
- Numerical Correctness Sweep
- Weekly Recap - Release Stabilization and Dynamo's Object Model Overhaul