PyTorch: Apple Silicon Gets a Metal Overhaul

A cluster of MPS backend fixes and migrations moved several operators from MPSGraph to Metal kernels, closing correctness bugs in clamp, RMSNorm, attention, and linspace along the way, while a separate stack landed the first user-facing pieces of torch.compiler.precompile.

Duration: PT2M57S

Episode overview

This episode is a short developer briefing from PyTorch.

It explains recent repository work in plain language.

  • Show: PyTorch
  • Published: 2026-07-12T13:00:36Z
  • Audio duration: PT2M57S

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 July 12th, twenty twenty six, and this is PyTorch, your daily developer briefing.

Today's biggest signal is a coordinated push to modernize the Apple Silicon backend. At least six PRs touch MPS, and they're not random bug fixes — they're one migration effort landing in careful order.

PR 189625 brings ternary operation specializations up to parity with the existing binary path, building shared infrastructure so future ops don't need one-off kernels. Right behind it, PR 189624 fixes two dispatch bugs in that same ternary path — including one where sixty-four bit indexing accidentally recursed into…

Alongside that stack, three more MPS fixes landed independently: PR 189631 corrects batch and head broadcasting in scaled dot-product attention, PR 189617 fixes fused RMSNorm so the weight multiply happens in float32, matching CPU and CUDA numerics instead of drifting in half precision, and PR 189630 preserves…

Second theme: precompile is taking shape. Bob Ren's stack — PRs 189620, 189621, 189629, and 189622 — adds torch dot compiler dot precompile with an eager backend, then an inductor backend, tests, and documentation, split deliberately to stay under PyTorch's…

On…

Nearby episodes from PyTorch

  1. Dynamo's Great Type-System Cleanup
  2. Dynamo's CPython Test Push and a Rocky Exception-Handling Rewrite
  3. Distributed Backend Overhaul and a Wave of Cleanup
  4. The Great Runner Migration, and a Distributed Rollback
  5. Distributed Cleanup and CI Infrastructure Overhaul
  6. Correctness Fixes Sweep the Compiler Stack
  7. Weekly Recap - Release Hardening and Correctness Cleanup
  8. Determinism and Correctness Sweep