diffusion-kernel

To build and test custom CUDA kernels for diffusion models, use battle-tested templates for operations like RMSNorm and RoPE, wrap them with JIT loaders, and validate correctness against PyTorch baselines using provided unit tests.

torch.compile integrates with custom Triton kernels by patching the diffusion model before compilation, ensuring PyTorch compatibility and allowing the JIT CUDA launchers to function correctly within the compiled execution graph.

The best way to profile Triton kernels for diffusion throughput is using Nsight Compute to analyze occupancy and vectorized memory access, enabling iterative performance optimization to maximize end-to-end denoise speed.

Yes, you can use CuTe DSL patterns for common diffusion operations. The Skill provides reusable kernel templates for operations like AdaLN, GEGLU, and SiLU gate that are ready to adapt for CuTe DSL integration.

Yes, you need Triton and PyTorch to develop JIT CUDA kernels using this workflow, as the templates enforce PyTorch compatibility through torch.compile wrap_triton and rely on Triton for JIT compilation.

Custom CUDA kernels may fail diffusion model integration if they lack vectorized memory access, warp reductions, or proper wrapping with Python JIT launchers before applying torch.compile to the denoising stage.