PyTorch: Test Infrastructure Gets a Deep Clean
A wave of changes tackled slow and flaky test infrastructure across distributed, ROCm, and CI systems, while a fourteen-part precompile refactor moved torch.compiler.precompile from a callable to a module-based API.
Duration: PT2M34S
Episode overview
This episode is a short developer briefing from PyTorch.
It explains recent repository work in plain language.
- Show: PyTorch
- Published: 2026-09-17T13:00:27Z
- Audio duration: PT2M34S
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 September 17th, 2026, and this is PyTorch.
Today's biggest story isn't one feature, it's a coordinated push to make the test suite faster and more honest.
Start with distributed test store, PR 197324. That file cost fifteen minutes of wall clock per shard for tests that only need twenty-four seconds. The cause: process isolation was hiding state leaks, like process groups that never got torn down and retry logic that silently swallowed the wrong errors. Jeff Daily's…
That pattern repeats. PR 197340 strips ROCm test skips that no longer reflect real failures, some dating back to 2021, and corrects two that were actually masking convergence failures on newer architectures. A companion commit, 197313, stops ROCm's default test shards from silently duplicating fifty minutes of…
Second theme: the precompile API is being rebuilt in public. Bobrenjc93 landed a fourteen-part ghstack replacing the old callable `torch.compiler.precompile` with a proper module offering capture, load, and explicit tracer classes. The pieces are deliberate: guard tracking so dropped guards don't silently widen…
Two other items worth flagging. FSDP2 got a crash fix for unsharding under inference…
Nearby episodes from PyTorch
- Sharding, Fake Tensors, and the Great Test Cleanup
- Correctness Fixes and the AOT Compile Hardening Push
- Compile-On-One-Rank Gets Serious About Device Portability
- Weekly Recap - Compiler Correctness and MPS Backend Consolidation
- Apple Silicon Hardening and Dynamo's Global State Cleanup
- Hardening the Compile-and-Ship Pipeline
- Distributed Training Takes Control of Its Own Timing
- Guard Serialization Gets a Precision Pass