Helion enables GPU kernel settings that bypass autotuning

Helion now selects GPU kernel settings at compile time without benchmarking alternative configurations, PyTorch announced. The heuristics are merged and enabled by default for H100 and B200 GPUs using the Triton backend. The compiler examines operations and dataflow to choose settings such as block sizes and parallel execution groups. PyTorch reports performance of 1.27x vLLM's handwritten CUDA operators across 153 kernel-and-shape combinations covering six reduction kernels on H100, compared with 0.58x for Helion's previous default. The change offers developers an alternative to spending GPU time tuning each kernel and shape. Kernels whose dimensions are not all known at compile time still fall back to the original default.