mini-avo

MiniAVO is an evolutionary search over Triton GPU kernels. Each generation an LLM (OpenRouter or Vultr Serverless Inference) rewrites the current best kernel, and candidates are filtered by a Python syntax check and an ahead-of-time Triton compile for the target architecture — sm_80 through sm_121, no GPU required — then scored from the compiled PTX and cubin. Where a GPU is available, every Nth generation is also checked against a torch reference and timed, and measured latency outranks the static score when choosing the next parent. Compile errors, wrong answers and timings all become instructions in the next generation's prompt, kept for the run or persisted across runs. Four GPU MODE-style problems ship with it — GEMM, reduction, batched Cholesky, AlphaFold3 triangle multiplication — and every run writes the full lineage, including each failed candidate and why, plus a submission script for the best kernel.