autotuning

PyTorch Helion now integrates NVIDIA CompileIQ's Advanced Control Files into its autotuner, enabling joint search over kernel configurations and low-level ptxas compiler controls to extract last-mile performance from already autotuned GPU kernels.

learnc 是一个工具，可在不修改 C 代码的情况下，针对用户编写的 C 评分函数自动调优代码中的命名常量，并在保持代码不变的情况下写回新值；在 JPEG 上实现 9.1% BD-rate 压缩提升，在 WebRTC VAD 上 F1 从 0.810 提升至 0.859。

PyTorch's Helion kernel DSL is integrated into vLLM's linear backend, where a single autotuned GEMM implementation covers Standard GEMM, Split-K, and Swap-AB variants, outperforming CUTLASS and DeepGEMM backends on NVIDIA Hopper GPUs with over 10% throughput gains on some workloads.

The Hugging Face Kernels project now supports Helion, a high-level DSL for writing performant ML kernels, enabling users to build, autotune, and ship kernels seamlessly.

A technique to accelerate matrix multiplication when M and N are small but K is large, as encountered in MoE routers and small-batch decoding, by decomposing K and running partial GEMMs in parallel. The approach beats PyTorch Inductor on most shapes using a custom Triton kernel.

This blog explores using LLM-guided autotuning to accelerate kernel configuration search in PyTorch's Helion DSL, replacing the slower Likelihood-Free Bayesian Optimization approach.