A model architecture alone does not explain how a request becomes a correct, measured result on GPUs. Between prompt and response, state must be represented, scheduled, moved, synchronized, stored, and checked.

Deep Dive into DeepSeek, Volume II: Systems, Kernels, and Integration makes that execution path visible. It follows model requirements through serving and speculative decoding; FlashMLA, DeepGEMM, TileKernels, and TileLang; parallel execution; DeepEP, Mega MoE, EPLB, and DualPipe; 3FS and smallpond; and finally precision, quantization, long-context systems, correctness, replay, and NVIDIA runtime measurements.

A recurring request trace shows how rendered tokens, attention pages, routed expert rows, kernel operands, communication buffers, pipeline dependencies, stored chunks, and replay records cross component boundaries. Each performance claim remains attached to its workload, layout, data type, hardware, and correctness gate. The result is both a conceptual guide to DeepSeek infrastructure and a practical framework for reading repositories, interpreting benchmarks, and validating integrations.

Written for ML systems engineers, infrastructure developers, CUDA programmers, performance engineers, researchers, and readers who want to understand how sparse language models run. Basic deep-learning familiarity is assumed; no prior GPU or distributed-systems programming experience is required. Volume II reintroduces the contracts it needs, so it can be read independently, while Volume I provides the fuller model and training foundation.