What problem does it solve? SGLang ships MoE kernel configs keyed by Triton version and device name, and on newer AMD GPUs like the MI355X no matching configs exist at all, so every fused-MoE layer silently runs default configs. This Skill guides you through capturing real shapes, tuning kernels, deploying configs correctly, and verifying from server logs that the tuned config actually engaged rather than falling back. ## Core Features & Use Cases - Version-aware config deployment: Place tuned JSON configs in the correct triton_<version> directory and deploy via SGLANG_MOE_CONFIG_DIR, avoiding the crash and silent-replacement traps of that override. - Log-based engagement verification: Distinguish an exact config hit from a version fallback or a default-config miss by reading SGLang's startup log lines. - Graph-capture-aware benchmarking: Account for HIP graph capture on the decode path, where kernel changes only take effect across server restarts and eager-mode timing is misleading. - Use Case: On an MI355X node serving a MoE model with SGLang, run the shipped tuning_fused_moe_triton.py tuner against captured production shapes, write results into triton_3_6_0, and confirm an exact-hit log line before A/B benchmarking throughput. ## Quick Start Ask the agent to tune the fused-MoE kernels for your SGLang server on an AMD Instinct GPU and verify from the server log that the tuned config loaded as an exact hit.
