how-to-optim-algorithm-in-cuda is a study and reference collection for CUDA kernel optimization and GPU systems work.
The repository addresses the challenge of learning practical GPU optimization by gathering hands-on examples, implementation notes, and architectural insights across multiple layers of the GPU stack. It covers concrete kernel implementations for common operations like reduction and softmax, explores high-level optimization frameworks such as CUTLASS and Triton, documents low-level instruction set details through PTX ISA notes, and synthesizes optimization patterns specific to LLM inference and training workloads.
This material suits developers working on GPU-accelerated systems who want to understand optimization from first principles rather than relying solely on library abstractions. It is particularly valuable for those building or tuning inference serving systems, implementing custom kernels, or diving into PyTorch internals on the GPU side. The collection spans from handwritten CUDA kernels through DSL-based approaches like Triton and CuTe, allowing someone to choose an appropriate abstraction level for their problem. The repository explicitly includes LLM-specific optimization notes, making it relevant for anyone working on large language model deployment or training infrastructure.
The project maintains active curation focused on CUDA kernels, LLM inference optimization, and AI infrastructure work. Older material is being consolidated or replaced with English-language entry points, indicating ongoing effort to keep content current and accessible. The repository functions as a public engineering notebook rather than a polished library, with content organized by topic across directories covering kernel implementations, lecture notes, framework-specific material, and architectural papers.
