llm_note
A collection of Chinese-language notes on LLM inference, transformer model structure and framework internals, covering quantization, attention optimizations, CUDA and Triton kernel programming.
llm_note is a set of study notes on large language models, organized into sections on transformer fundamentals, LLM quantization, inference optimization, and high-performance GPU programming. Topics include position encoding, MLA, SmoothQuant and AWQ quantization, FlashAttention versions one through three, KV-cache and CUDA graph optimizations, and tensor parallelism.
The high-performance computing section works through Triton and CUDA programming from GPU architecture basics to kernel development, and links out to external CUDA learning resources. The repository also documents a companion paid course that implements an inference framework from scratch using Triton and PyTorch, aimed at interview preparation for inference and high-performance-computing roles.