PyTorch: Autograd Overhead Cuts and a New GEMM Backend Migration
The autograd hot path and the NVGEMM inductor backend both got sustained performance surgery this cycle, while a new distributed communication backend for Apple Silicon and a fresh NCCL rewrite pushed further into review. Developers touching compiled autograd, matrix multiply tuning, or distributed code should expect meaningful behavior and performance shifts.
Duration: PT2M53S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-07-14T13:00:41Z
- Audio duration: PT2M53S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
It's July 14th, 2026, and this is your PyTorch briefing.
The clearest signal today: two separate teams are independently squeezing Python overhead out of hot paths, and a new distributed backend landed for Apple hardware.
First, autograd. Edward Yang's collaborator Richard Zou is running a multi-PR stack shrinking AOTAutograd and custom Function overhead. PR 189759 stops detaching every saved view at runtime, fixing a long-standing correctness gap tied to issue 94990. PR 189790 moves save-for-backward bookkeeping into C++. And a…
Second theme: the NVGEMM inductor backend. Michael Lazos landed a large stack, PRs 189771 through 189841, migrating off an internal cutlass API to the official cutlass operators interface, while layering in performance work — swap-A-B tiling for small-M decode shapes, addmm epilogue fusion, avoiding a 14-second…
Third: distributed communication continues to expand. Tristan Rice's NCCL2 backend stack added one-sided RMA window support and fault-tolerant reconfiguration with backend-agnostic tests shared with Gloo. Separately, PR 189856 introduces TCCL, a new Thunderbolt-based collective communication library for Apple…
Infrastructure notes: CI keeps migrating…
Nearby episodes from PyTorch
- Hardening Boundary Checks and MPS Correctness
- Weekly Recap - Metal Kernels, Compile Reliability, and Dynamo's Long Tail
- Apple Silicon Gets a Metal Overhaul
- Dynamo's Great Type-System Cleanup
- Dynamo's CPython Test Push and a Rocky Exception-Handling Rewrite
- Distributed Backend Overhaul and a Wave of Cleanup
- The Great Runner Migration, and a Distributed Rollback
- Distributed Cleanup and CI Infrastructure Overhaul