Helion's New CuTe DSL and Pallas Backends and How We Built Them Via Agents
Talkat KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 · 2026-09-07..09-09·
summary
Helion adds two new backends: a CuteDSL backend targeting NVIDIA Blackwell and a Pallas TPU lowering that emits VM-prefetched pipelines. Autotuner chooses between emit_pipeline and onroll lowering (onroll wins for ~8k sequence length, emit_pipeline for ~32k), and hybrid LM+hill‑climb autotuners outperform either alone. Agents run multi‑day agentic hill‑climbing loops that modify compiler search knobs and produce kernels that match or beat Tokamax and other baselines across many shapes, while source kernels stay ~100× shorter and the system generalizes across dozens of kernels.
AI-written summary · about
chapters (13)
- 0:00Helion and its backends
- 1:06Autotuning and pipeline strategies
- 10:02Performance vs Tokenmax
- 11:08Jason Enol and agentic workflows
- 11:51AI agents reshaping development
- 12:57Agentic hill climbing loops
- 14:21Hill climbing kernels vs compilers
- 17:12Agentic compiler optimization
- 19:40Helion's autotuner advantages
- 21:05Hybrid autotuner backends
- 22:09Attention kernel on NVIDIA GPUs
- 23:13Future of performance convergence
- 24:59Q&A
related talks
24mHelion 1.0: A High-Level DSL for Performance Portable Kernels
28mHelion: A High-level DSL for Kernel Authoring
11mQiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a...
08mA Halide Backend for TorchInductor
18mYour Coding Agent Should Do AI System Engineering
09mLightning Talk: Vllm-triton-backend: How To Get State-of-the-art Performance on...
23mTorchTPU: Expanding TPU Programmabil...
10mAuto-tuning vllm