viable/strict/1790898650: [CUDA][LinAlg] Nopiv LU - better heuristics (#193004)
- PyTorch: 1775 events in the last 90 days
- PyTorch: 1753th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790897185
What happened
Without these we inflict up to 7x regression on inputs with low batch count, and up to 2x for high-batched inputs. With about 540 points (gathered for A100, H100, GB200), this heuristic gets about 98% right. This is a prep work for LU improvements without the pivoting where MAGMA still outperforms cuBLAS/cuSOLVER. Pull Request resolved: #193004 Approved by: https://github.com/eqy
Summary assembled by rule from the sources below