What problem does it solve? Kernel changes in the FLA repository (Triton, Gluon, TileLang, CUDA backends) can silently regress throughput or latency, and merge requests often lack the profiling evidence reviewers need. This Skill standardizes how performance work is done: hardware baselines, benchmark workflows, Nsight Compute profiling, and MR-ready evidence requirements. ## Core Features & Use Cases - Hardware baseline rules: Defines which NVIDIA GPUs count as valid performance evidence (sm_90+ datacenter cards preferred, consumer cards reference-only). - Benchmark workflow: Provides quick sanity-check commands for training throughput and generation benchmarks, plus dense and variable-length workload coverage requirements. - NCU profiling guidance: Supplies minimal ncu --set full and --set source command templates, profile artifact layout under profile/<run_name>/, and integration with an optional external ncu-report-skill. - Use Case: Before opening an MR that modifies a Gated DeltaNet Triton kernel, run the before/after benchmark on an H100, capture NCU full and source profiles, and paste the key metrics summary into the MR description. ## Quick Start Ask the agent to profile the chunk_kda forward kernel with Nsight Compute and prepare before/after benchmark evidence for an MR touching fla/ops kernel code.
