• Sources: primary, discussion
  • Summary: A vLLM post by NVIDIA's Benjamin Chislett describes an agent-assisted CuTe DSL decode path for Kimi K3 on GB300. It cuts per-layer decode from 23 kernels and 130 microseconds to 13 kernels and 69 microseconds, with 1.31x to 1.92x higher decode throughput depending on batch size and tensor parallelism. The authors say agent-written kernels outpace line-by-line review. They propose a fast-track RFC that keeps changes inside model code, bounds hardware scope, requires correctness and long-context evals, and lets contributors run standard evals on their own GPUs. Their audit found a compiler reordering of activation reads ahead of a synchronization wait that simple functional checks missed.
  • Why it matters: The post is a concrete case of replacing line-by-line human review of agent-generated code with isolation, scoped support, and reproducible evidence, and it lists a failure mode that review must still catch.