附录 B · 论文与资料清单

Attention 与 Kernel - FlashAttention (2022) / FlashAttention-2 (2023) / FlashAttention-3 (2024) - PagedAttention · vLLM (SOSP 2023) - Triton: An Intermediate Language and Compiler (2019)

调度与 Serving - Orca: Continuous Batching (OSDI 2022) - Sarathi-Serve: Chunked Prefill (OSDI 2024) - SGLang / RadixAttention (2024) - DistServe / Splitwise: PD 分离 (2024)

量化 - GPTQ (2023) / AWQ (MLSys 2024) / SmoothQuant (ICML 2023) - FP8 Formats for Deep Learning (NVIDIA, 2022)

模型结构 - GQA (EMNLP 2023) - DeepSeek-V2 (MLA) / DeepSeek-V3 (MTP + MoE) - DeepSeek-V3.2（DSA，2025） / DeepSeek-V4（CSA/HCA，2026） / DeepSeek-V4.1-Flash（CED/CSA2，2026） - Qwen3.8-Next（GDN/QSA/GR，2026） - Mixtral of Experts (2024)

投机解码 - Fast Speculative Decoding (Leviathan et al., 2023) - Medusa (2023) / EAGLE (2024)

并行 - Megatron-LM: TP (2019) / GPipe、PipeDream: PP