mlsys26-flashinfer-contest

Code review and debugging assistant. Identifies bugs, performance issues, security vulnerabilities, and suggests optimizations.

Optimize MLSYS26 FlashInfer contest operators for NVIDIA B200/Blackwell with a reference-first workflow. Use when tasks require refreshing only the current project's local `reference/` git repos before optimization, searching high-quality official repos when local references are insufficient, deriving workload regimes from contest JSONL, running paired-baseline local gates, exploring 2-3 representative regimes before costly full sweeps, asking the user before expensive multi-GPU evaluation, extracting compact NCU metrics from Modal artifacts, archiving scratch outputs, applying CUDA 13.2 Blackwell tactics, and promoting only on repeatable mean-latency wins.

Create and push FlashInfer contest submission tags such as `submission-v3` in a git repository. Use when the user asks to "tag your submission", "create a submission tag", "push the latest submission tag", or prepare a FlashInfer contest repo so the evaluator will use the latest tagged commit. Always validate `config.toml` topology before committing or tagging, especially for multi-definition repos where extra or misplaced configs can make the evaluator pack an empty solution.