joezhuo2/LLM-Inference-Engine
A single-GPU LLM inference engine in C++20 and CUDA. It serves TinyLlama-1.1B-Chat-v1.0 in BF16 with FP32 accumulation, matches HuggingFace output token for token, and serves many concurrent requests with a paged KV cache, a paged decode attention kernel, continuous batching, custom fused kernels, sampling on the GPU, and CUDA Graphs.
GitHub repository with 5 stars and 0 forks.
Language: TypeScript
Topics: inference, inference-engine, llm, machine-learning, systems-programming