zhongzhouTan-coder.github.io

An AI-maintained personal knowledge base — all content written by GitHub Copilot, published on GitHub Pages. The goal: test whether an agent-driven workflow can replace manual note-taking and do long-term knowledge precipitation automatically.

Browse the wiki →

What's inside

Five topic categories covering the LLM inference and training stack, built from primary sources (papers, codebases, technical blogs):

Category	Topics
Benchmarks	Agent evaluation (DeepSWE, τ-bench, Pier, AutoJudger), serving performance (EvalScope Perf, AISBench)
Frameworks	vLLM, vLLM Ascend, SGLang, Triton, Triton Ascend, DSpark, Harbor, DeepSeek
Algorithms	FlashAttention (v1–v4), attention variants (MQA, GQA, MLA, Collaborative Attention), linear attention, DeepSeek-V3.2 sparse attention
Training	Foundation models (GPT-1/2/3, LLaMA), parallelism (Megatron-LM, GPipe, sequence parallelism), DeepSeek-V4, Kimi K3/Linear, efficient attention (MSA, SWAT, Gated DeltaNet), fine-tuning (intrinsic dimensionality, Socratic-SWE)
Hardware	Quantization (FlatQuant, NVFP4), accelerator numerics

Every page links back to its original sources and related pages. See the wiki index for the full catalogue and the knowledge base introduction for navigation conventions.

How it works

flowchart LR
  R[raw/] -->|mineru / web capture| D[derived/]
  D -->|agent synthesizes| O[docs/]
  O -->|jekyll build| P[GitHub Pages]
  S[sources.json] -.->|tracks| R
  S -.->|tracks| D
  S -.->|tracks| O

Loading

1. Capture — PDFs and web pages land in raw/ as immutable sources and are registered in sources.json.
2. Extract — PDFs convert to Markdown via MinerU; web pages render through a headless browser. Output lands in derived/pdf-markdown/ or derived/web-markdown/.
3. Synthesize — An agent reads the extracted Markdown, writes a deep-dive docs page under docs/, cross-links related pages, creates term glossary entries, and updates the wiki index and change log.
4. Publish — Jekyll builds the site from docs/ and deploys to GitHub Pages.

The full workflow rules live in AGENTS.md.

Quick start

Browse the knowledge base:

# Start with the categorized index
open docs/logs/index.md

# Or read the knowledge base introduction
open docs/README.md

Search the wiki:

npm run wiki:search -- "paged attention" --limit 5
npm run wiki:search -- "pipeline bubble" --category training --json

Inspect the link graph:

npm run wiki:graph -- --top 10
npm run wiki:graph -- --json
npm run wiki:graph-data

Run integrity checks:

python3 scripts/checks/repository_integrity.py --json
python3 scripts/checks/source_names.py --json

Serve locally:

bash scripts/serve-local.sh

Repository code references use relative external-repos/ targets in their Markdown source, so editor previews open the local dependency checkout. Jekyll pages resolve the same references to provider-correct GitHub or GitCode links pinned to the inspected commit.

Refresh repository evidence:

# Fetch latest upstream and compare only the subsystem being studied. Relevant
# changes are deferred until the current evidence revision is 14 days old.
./scripts/run-in-workspace.sh python scripts/repositories/worktree.py sync \
  vllm-a0c092ee72c0 \
  --path vllm/v1/core \
  --sparse vllm/v1/core

# Retire or restore an inactive pinned worktree without losing its evidence.
./scripts/run-in-workspace.sh python scripts/repositories/worktree.py retire \
  vllm-a0c092ee72c0
./scripts/run-in-workspace.sh python scripts/repositories/worktree.py materialize \
  vllm-a0c092ee72c0

The helper keeps one partial bare object cache per upstream repository. It creates a detached revision worktree only when the requested paths changed and the default 14-day revision interval has elapsed. A deferred sync reports the next eligible date without creating a snapshot. Use --min-revision-interval-days N to tune the cadence, or --force-new-revision for an intentional urgent refresh. This keeps immutable evidence reproducible without producing a snapshot for every upstream change.

Prepare external repositories in a fresh workspace:

# Offline report: which registered pinned checkouts are locally readable?
./scripts/bootstrap-external-repos.sh --status

# Materialize only the repositories needed for the current task.
./scripts/bootstrap-external-repos.sh vllm-a0c092ee72c0 vllm-ascend-32a59d4e349c

# Or materialize every registered revision.
./scripts/bootstrap-external-repos.sh

The tracked docs/_data/code_repositories.json file is the portable lock table: it records each checkout path, provider URL, and exact commit. The ignored external-repos/ directory remains a disposable workspace cache. Repository hydration is explicit because cloning every codebase can consume substantial network bandwidth and disk space.

Lint docs:

bash scripts/lint-docs.sh

Web source capture

Install dependencies and capture a web page:

npm ci
npm run ingest:web -- \
  --url "https://example.com/article" \
  --category frameworks \
  --slug example-article

The command saves immutable HTML and metadata under raw/, writes readable Markdown under derived/web-markdown/, preserves inline SVG diagrams as local sidecar assets, and registers the captured revision in sources.json. See the web source instructions for renderer choices and the captured → ingested workflow.

Project structure

.
├── raw/                   Canonical source files (never modified)
├── derived/
│   ├── pdf-markdown/      Markdown extracted from PDFs
│   ├── web-markdown/      Markdown extracted from web pages
│   └── repo-analysis/     Code-reading notes from checked-out repos
├── docs/                  Published knowledge pages (Jekyll site root)
│   ├── algorithms/        Inference algorithms and kernels
│   ├── benchmarks/        Benchmark designs and evaluation
│   ├── frameworks/        Serving systems and runtimes
│   ├── hardware/          Numerics, quantization, accelerators
│   ├── training/          Pretraining, fine-tuning, parallelism
│   ├── terms/             Cross-paper technical glossary
│   └── logs/              Wiki index and chronological change log
├── scripts/               Ingest, lint, search, and graph utilities
├── external-repos/        Third-party repo checkouts (git-ignored)
├── sources.json           Manifest linking raw → derived → docs
├── kb-categories.json     Canonical category registry
└── AGENTS.md              Agent workflow rules