First stable release of Auto GPU Kernel that extends beyond the competition submission and flashinfer.
- Support arbitrary tasks and git repositories: describe the objective, how to measure, and how to validate in natural language. The agent generates
harness/validate.py+harness/benchmark.pyto be used for optimization. - Add tool
kbench: validation, benchmarking, A/B testing, and experiment history. - Add tool
kopt: scaffolds a harness and drives the optimization agent (via OMP) with budget, iteration, and time limits. - Add tool
kopt watch: a live web view for following agent messages, optimization chart, experiment files.