Develop and benchmark optimized CUDA kernels for HuggingFace diffusers and transformers.
npx skills add https://github.com/m0at/claudemd --skill cuda-kernels-m0at
Please help me install this Agent Skill. Skill: cuda-kernels Source: https://github.com/m0at/claudemd/tree/main/skills/cuda-kernels Command: npx skills add https://github.com/m0at/claudemd --skill cuda-kernels-m0at
CUDA kernel development and benchmarking guidance for NVIDIA GPUs used with HuggingFace diffusers and transformers. It helps practitioners design, implement, and benchmark custom CUDA kernels, compare against baseline implementations, and integrate with the HuggingFace Kernels Hub for loading pre-compiled kernels on H100, A100, and T4.
Run the end-to-end and RMSNorm benchmarks to compare optimized kernels against baseline implementations.
High-intent search queries and answers about installing and using this skill.
To benchmark custom CUDA kernels against baseline PyTorch implementations, run the provided end-to-end video generation and micro-benchmark scripts. These workflows compare optimized kernel performance against baseline implementations to measure GPU acceleration gains on NVIDIA H100, A100, and T4 hardware.
Integrate pre-compiled CUDA kernels from HuggingFace Kernels Hub into diffusers pipelines using the get_kernel function. This integration pattern allows loading and applying external kernels with minimal code changes to your existing transformers and diffusers workflows.
Available CUDA kernel patterns for optimizing diffusion models and LLMs include RMSNorm, RoPE, GEGLU, and AdaLN. These core patterns provide optimized implementations for common operations in transformer architectures and diffusion pipelines.
Yes, you can write new CUDA kernels for NVIDIA GPUs running HuggingFace diffusers. It provides guidance for developing custom CUDA operations targeting diffusion models and LLMs, with practical patterns and integration guides for GPU-accelerated AI workloads.
Yes, PyTorch and diffusers are required dependencies to benchmark CUDA kernel performance. The benchmarking workflows also depend on transformers, numpy, and accelerate to compare optimized kernels against baseline implementations in end-to-end video generation scenarios.
Custom CUDA kernels provide optimized execution for specific operations like RMSNorm and RoPE compared to standard PyTorch operations. Benchmarking workflows demonstrate measurable performance improvements on NVIDIA GPUs, particularly for compute-intensive diffusion model and LLM workloads.
drunkcodingDevelop, benchmark, and integrate CUDA kernels for HuggingFace Diffusers and Transformers.
danielfleischerBenchmark optimized CUDA kernels for diffusers and transformers on NVIDIA GPUs.
dotruruBuild and benchmark CUDA kernels for NVIDIA GPUs in HuggingFace workflows.
FutureMLS-LabDiagnose SGLang CUDA crashes by recording kernel API calls and tensor metadata.
