h100-diffusers-kernels

Optimize CUDA kernels for diffusion models on H100 GPUs by applying tailored kernel templates for RMSNorm, RoPE, and AdaLN, then injecting them into your PyTorch and Diffusers pipelines to maximize performance.

Yes, you can use optimized CUDA kernels with LTX-Video in Diffusers by loading the pipeline on CUDA and running the kernel injection flow to patch and verify the integrated RMSNorm and RoPE kernels.

Accelerating diffusion models on H100 requires specific CUDA kernel patterns for RMSNorm, RoPE (1D/3D), GEGLU, SwiGLU, and AdaLN, focusing on block sizes, shared memory usage, and memory bandwidth considerations.

Yes, patching Diffusers kernels requires a CUDA toolchain, PyTorch and Diffusers bindings, and a reproducible injection workflow to successfully apply and verify the optimized H100 kernels.

Profile CUDA kernel performance for Stable Diffusion and DiT workflows using provided profiling commands to analyze block sizes, shared memory usage, and memory bandwidth bottlenecks on the H100 architecture.

An RMSNorm kernel underperforms in Diffusers when block sizes and shared memory usage are not tailored to the H100 architecture, requiring optimization guidance and a reproducible injection workflow to verify performance.