PyTorch Helion can now autotune Nvidia compiler controls, for 2 to 7% gains

PyTorch says Helion's autotuner can now search Nvidia ptxas compiler controls (via CompileIQ) jointly with kernel configs. On one already-tuned NVFP4 GEMV kernel on GB200 it measured a 4.7% speedup, up to 6.9% with multi-objective tuning. The post expects 2 to 7% on optimized kernels. It requires CUDA 13.3 or later, and CompileIQ is proprietary.

HEAT

OUTLETS

1

FIRST SEEN

Oct 20, 2026, 06:04 ET

LAST SIGNAL

6d ago