Deep Dive into DeepSeek, Volume II

Follow a DeepSeek request from its interface to a correct, measured result on GPUs.

A model architecture alone does not explain how a user request becomes a response. Between the interface and the result, state must be represented, scheduled, moved, synchronized, stored, and checked. Deep Dive into DeepSeek, Volume II makes that execution path visible.

Inside Volume II

• Systems foundations, model serving, scheduling, and runtime state
• DeepSeek-V4 runtime behavior and speculative decoding
• FlashMLA, DeepGEMM, TileKernels, and TileLang kernel surfaces
• Tensor, pipeline, expert, sequence, and context parallelism
• DeepEP communication and fused expert-parallel execution
• Expert placement with EPLB and pipeline scheduling with DualPipe
• 3FS storage and distributed data processing with smallpond
• Precision, quantization, and numerical correctness
• Long-context execution, system oracles, replay, and failure analysis
• NVIDIA kernel measurements, benchmarking, and evaluation protocols

The book follows the ownership and movement of request state across serving runtimes, GPU kernels, distributed communication, and storage. It identifies which component owns each handoff, keeps performance reasoning attached to correctness, and distinguishes conclusions established by source inspection from conclusions established by measurement.

Written for machine-learning engineers, systems practitioners, researchers, and students, this volume builds on the model vocabulary established in Volume I while reintroducing the systems contracts it needs. No prior GPU programming or distributed-systems programming experience is required.

For readers who want to understand what happens between a model interface and a measured GPU result, this is a mechanism-first guide to the DeepSeek systems stack.