What problem does it solve? Kernel changes in the FLA repository can silently regress throughput or latency, and merge requests often lack the profiling evidence reviewers need. This Skill standardizes how NVIDIA GPU performance work is done, from hardware baselines to NCU profiling and MR-ready benchmark reports. ## Core Features & Use Cases - Hardware baseline rules: Defines which GPUs count as valid performance evidence (sm_90+ datacenter cards, preferred sm_100/sm_103) and which are reference-only. - MR performance evidence checklist: Requires before/after benchmarks, NCU full and source profiles, dense and variable-length workload coverage, and an explicit conclusion on regressions. - Profiling workflow and artifact layout: Provides ready-to-use ncu and benchmark commands plus a standard profile/<run_name>/ directory structure that keeps .ncu-rep files out of git. - Use Case: After optimizing a Triton kernel in fla/ops, run the prescribed benchmark and NCU commands, then paste the summarized metrics into the MR description so reviewers can verify the improvement. ## Quick Start Ask the agent to profile the chunk_kda forward kernel with NCU and prepare before/after benchmark evidence for an MR on an H100.
