KDA-v0.5 Kernels Beat Human Winners by Up to 69% in FlashInfer Contest
songhan_mit · x · 2026-08-31
The KDA (Kernel Design Agents) team released v0.5 updates, where their Cute-DSL kernels outperformed human winners in the FlashInfer Kernel Contest by 16% to 69%. Key improvements include 1.69x speedup in GDN prefill, 1.41x in DSA attention, and 1.17x in FP8 MoE. The progress is attributed to integrated Cute-DSL primitives, better "humanize" workflows, a self-evolved kernel-wiki, and enhanced profiling skills (IKET). The team also detailed the "flame chase" flow using GPT-5.6-sol and fable-5 agents for iterative optimization.
More from coding & agent
- Gmail smart compose generates functional Python training code for free — dejanseo · 2026-08-31
- GLM 5.3 Flash beats Kimi K3 on same coding task at a third of the cost, self-repairs in 10 minutes — HowDevelop · 2026-08-31
- Dev builds a Recursive Language Model engine from scratch to digest arXiv papers — malliktwts · 2026-08-31
- GPT-5.6 Sol (med) dominates interactive coding agent benchmarks — steipete · 2026-08-31
- Observing model rollout behaviors: probing infrastructure — xeophon · 2026-08-31
- AI Agent Misinterprets Meeting Notes, Creates API Key Unsupervised — Altruistic-Toe4930 · 2026-08-31