Problem

Initial 3-shard configuration resulted in long per-shard execution times (~5h 26m average).

Solution Implemented

Increased test sharding from 3 to 5 shards to improve per-shard execution time.

Results Achieved

Per-shard time: 5h 26m → 3h 13m
Reduction: 40.6% improvement
Total time: ~16 hours (unchanged - sequential execution)
Tests per shard: ~6,800 → ~4,100

Benefits

• [x] Faster feedback per shard
• [x] Better load balancing
• [x] Easier debugging (smaller test batches)
• [x] Aligns with PyTorch standards (5 shards like Ubuntu)
• [x] Ready for parallel execution if more runners added

References

• Analysis: 5_SHARD_ANALYSIS_REPORT.md
• Workflow: .github/workflows/rhel-build-test.yml
• Performance run: https://github.com/subinz1/pytorch/actions/runs/20745368086