PyTorch reshards test jobs to target under 60 min per shard
Four longest-running test configs get more shards to cut time from 180 min to ~53 min
Deep Dive
A draft pull request (#190715) proposes increasing shard counts on the four longest-running PyTorch trunk test configs so each shard's median wall-clock lands under an hour, improving time-to-signal and node utilization. The cuda13.0-gcc11 default job goes from 5 to 14 shards (147m→~55m), distributed from 3 to 11 (180m→~53m), win-vs2022-cpu from 4 to 12
Key Points
- CUDA 13.0 default test shards increased from 5 to 14, cutting median runtime from 147m to ~55m
- Windows VS2022 CPU config went from 4 to 12 shards, dropping from 141m to ~53m median
- 30 total new shards added; ROCm and macOS pools excluded due to fixed/scarce resource constraints
Why It Matters
Faster CI feedback for PyTorch contributors means quicker iteration and better resource utilization across GPU and Windows pools.