Knowledge note

Knowledge Base Introduction

Overview of this repository's documentation structure, category organization, and how to navigate or contribute to the knowledge base.

Updated September 16, 2026Confidence: High

Knowledge Base Introduction

This repository is an AI-maintained personal knowledge base published on GitHub Pages. Source material lives in raw/. Processed knowledge pages live in docs/. A searchable index and a chronological change log live in logs/.

Directory Structure

  • raw/ — Canonical source files. File content is not modified; filenames and category paths are maintained for retrieval.
  • derived/pdf-markdown/ — Generated Markdown extracted from PDFs, organized under the same category names as raw/ and docs/.
  • docs/ — AI-maintained markdown knowledge pages, divided into topic categories, each with subcategory folders for related projects or themes.
    • benchmarks/ — Benchmark designs, tasks, metrics, and empirical findings; subfolders agent-eval/ and serving-perf/.
    • frameworks/ — LLM serving and language-model programming systems; per-project subfolders such as vllm/, vllm-ascend/, triton/, triton-ascend/, sglang/, dspark/, harbor/, and deepseek/.
    • algorithms/ — Inference algorithms and kernel-level methods; subfolders foundations/, flashattention/, attention-variants/, and linear-attention/.
    • training/ — Model training, fine-tuning, transfer learning, and generalization; subfolders parallelism/, foundation-models/, deepseek/, kimi/, efficient-attention/, and fine-tuning/.
    • hardware/ — Numerics, hardware features, and accelerator-specific notes; subfolder quantization/.
  • sources.json — Machine-readable manifest linking raw sources, generated Markdown, and docs pages.
  • kb-categories.json — Canonical category registry and legacy category aliases.
  • logs/
    • index.md — Categorised index of all docs pages.
    • log.md — Chronological record of every ingest and edit.

Categories

Category When to use
benchmarks Papers or notes about benchmark construction, domains, metrics, task generation, or evaluation results.
frameworks Runtime systems, serving engines, programming frameworks, and application orchestration systems.
algorithms Inference algorithms, kernel implementations, scheduling methods, and mathematical procedures.
training Training dynamics, fine-tuning methods, transfer learning, and generalization analysis.
hardware Hardware formats, accelerator features, precision recipes, and numerics-focused notes.

Create a new category only when an existing one would make the page hard to find.

How to Use

Finding information

  1. Start with logs/index.md to browse by topic, person, concept, or source.
  2. Follow links to the relevant docs page.
  3. Check the page sources in the front matter to understand what evidence supports the content.

Ingesting a new source

  1. Choose one canonical category from kb-categories.json.
  2. Name the source with the deterministic source naming policy, for example raw/algorithms/attention/flashattention/flashattention-io-aware-exact-attention--arxiv-2205.14135v2.pdf.
  3. Add or update the corresponding sources.json entry.
  4. For PDFs, generate Markdown under the matching derived/pdf-markdown/{category}/{topic_path}/ directory.
  5. Follow the repo rules in AGENTS.md to read the source, create or update docs pages, and update logs/index.md and logs/log.md.

Maintaining quality

Run the docs lint check at any time:

./scripts/lint-docs.sh

For source naming and manifest checks:

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

Or invoke the lint-docs-cleanup Codex skill to have the agent fix safe issues and report deletion candidates.