Understand this story
On November 13, 2026, the PyTorch blog announced that PyTorch's Helion DSL has added compile-time heuristics, which analyze the kernel IR and data flow to classify it as matmul, reduction, or pointwise, and directly provide configurations such as block size, num_warps, and num_stages without the need for autotuning. According to the blog, on H100, a linear attention kernel is 1.18x faster than the hand-written Triton FLA forward pass, 6 vLLM reduction kernels are 1.27x faster than their hand-written CUDA/C++, and the general reduction kernel suite reaches 1.09x of torch.compile max-autotune. The heuristics have been merged into Helion and are enabled by default on the Triton backend for H100 and B200.
AI-generated from public reports
Latest development · Report
PyTorch Helion Introduces Compile-Time Heuristics, Selecting Kernel Configurations Without AutotuningDevelopments
1 confirmed developments
- Source:PyTorch Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 13
PyTorch BlogFirst-party
PyTorch Helion Introduces Compile-Time Heuristics, Selecting Kernel Configurations Without Autotuning
PyTorch's Helion DSL adds compile-time heuristics that analyze the kernel IR and data flow to classify it as matmul, reduction, or pointwise, and directly provide configurations such as block size, num_warps, and num_stages without autotuning. On H100, the linear attention kernel is 1.18x faster than the hand-written Triton FLA forward pass, vLLM's 6 reduction kernels are 1.27x faster than its hand-written CUDA/C++, and the general reduction kernel suite reaches 1.09x of torch.compile max-autotune. The heuristics have been merged into Helion and are enabled by default on the Triton backend for H100 and B200.