sglang-diffusion-ako4all-kernel

To optimize SGLang diffusion kernels safely, use a disciplined workflow that builds custom microbenchmarks, profiles with ncu, and validates improvements against real denoising workloads before porting changes back to production.

The best way to profile Triton and CUDA JIT kernels is using ncu profiling within a clean AKO4ALL checkout to establish baselines, diagnose register or occupancy limitations, and iterate on one optimization idea at a time.

To validate kernel optimizations for real denoising workloads, port the best candidate back to SGLang, run targeted tests, compare denoise performance dumps, and execute end-to-end validation to preserve correctness evidence.

You should optimize a diffusion kernel hotspot in SGLang only after profiling has established a genuine hotspot, ensuring that the optimization effort targets actual performance bottlenecks rather than unvalidated microbench gains.

Yes, you need a clean AKO4ALL repository checkout to benchmark and optimize CUDA kernels, because cloning, synchronizing, and verifying repository hygiene before optimization work prevents contamination of baseline performance metrics.

Yes, the kernel optimization workflow supports correctness checking for models like Qwen, FLUX, Wan, Hunyuan, and MOVA by requiring custom correctness-checked microbenchmarks and preserving correctness evidence during end-to-end validation.