LLM Inference

An opinionated and incomplete survey of LLM inference and serving runtimes from a systems and infrastructure lens.

• LLMs and Transformers Introduction, embeddings, transformers and attention mechanisms
• Inference and the KV Cache Inference execution and the KV cache
• Sharding a Model Pipeline, tensor, and expert parallelism
• Batching, Scheduling, and Paging Continuous batching, Orca, and PagedAttention
• I/O-Aware Kernels FlashAttention and FlashInfer
• Speculative Decoding Speculative decoding, EAGLE, and Medusa Trees
• Prefill-Decode Scheduling and Disaggregation Chunk prefill and prefill-decode disaggregation
• KV Cache Management and Offload Prefix caching and KV offload
• Appendix: Overview of Training Fine-tuning, RLHF, RLAIF, quantization, and alignment techniques
• Appendix: GPU Hardware Architecture, CUDA and ROCm, kernels and Triton, memory hierarchy
• Appendix: Inference Runtimes LLM Serving Stacks, TensorRT, Triton, vLLM, and SGLang