sglang-diffusion-ako4all-kernel

To optimize SGLang diffusion kernels, you reproduce the hotspot in a clean AKO4ALL repository, establish correctness and performance baselines, profile representative shapes with ncu, and iterate one-change-at-a-time improvements before porting the best candidate back.

ncu profiling in this workflow measures representative shapes to diagnose hardware-level causes of slow diffusion kernels. It establishes a performance baseline and records targeted optimization iterations to ensure improvements are validated before integration.

Yes, production validation ports the tuned kernel back to SGLang and runs targeted tests, kernel benchmarks, and denoise performance dumps to confirm the optimization transfers to real model denoising for Qwen, FLUX, Wan, Hunyuan, and MOVA.

Yes, AKO4ALL repository hygiene is required before tuning begins. The workflow clones or refreshes a sibling repository and verifies its branch, remotes, cleanliness, and synchronization with upstream to prevent inconsistent tuning states.

The best way to prevent unvalidated microbenchmark improvements is building a custom harness that establishes correctness baselines first, then recording one-change-at-a-time optimization iterations with ncu profiling and end-to-end denoise validation.

Performance wins fail to transfer when kernels are validated only through microbenchmarks. This workflow requires model-level denoise validation, performance dumps, and generated output comparisons to ensure the optimization survives real diffusion workloads.