一份统一 12 章的深度报告,覆盖软件架构第一性原理、执行热路径、KV-Cache、分布式、编译图、Ascend Overlay 与性能手册。
NEW · 专题深度
百万序列 Prefill：vLLM 如何解决 + 仍开放的难题
全模块分析
vllm_ascend_architecture_full.html
主报告 · 12 章
vllm_architecture_unified.html
专题 · NEW
HBM 索引架构设计
| § | 章 | 核心 |
|---|---|---|
| 0 | Overview & v0→v1 | v0 4 条缠结→v1 拆分根因 |
| 1 | Software-Architecture First Principles | "按变化速率切"公理 + 8 模块职责表 + 跨合约表 |
| 2 | Architecture Layers | EngineCore / Scheduler / Worker / ModelRunner / Attention / Platform / Compilation |
| 3 | Execution Hot Path | schedule → 5 阶段 forward + slot_mapping 页表 walk |
| 4 | KV-Cache (4-layer) | Manager / Coordinator / BlockPool / Connector (MLA forcing) |
| 5 | Parallel & Distributed | TP/PP/DP/EP + CP + EPLB + PD disagg |
| 6 | Compilation & Graph | FULL/PIE/eager + ACL Graph diff + 70 fusion + 10+ 量化 |
| 7 | Ascend Overlay | 5 后段 / ACL 图 / KV-Pool 3-backend / EPLB / patch 层 |
| 8 | Perf Handbook | OS → allocator → HCCL(AIV) → fusion + 特性互斥表 |
| 9 | Generic HW-Adaptation Playbook | 8 步 + MUST/SHOULD/NICE + fork/plugin/patch 决策 |
| 10 | 90-Day Pulse | ~1060 commit 主题分布 + 5 条进阶物理 |
| 11 | References & Drill Anchors | file:line 全索引 + 外链 |
| 目录 | 内容 |
|---|---|
_archived/arch_analysis_round1/ | round-1: spine + L1–L4 agent + FP + community_dive |
_archived/analysis_dsa_round2/ | round-2: vllm_ascend_dsa_analysis + kvcomp 稀疏谱系 |
_archived/ascend_arch_round3/ | round-3: platform/worker/attention/distributed 6 章 |
_raw/ | copy 备份 |