# The Atlas Book

> Documentation for Atlas — an open source, pure Rust and CUDA LLM inference
> engine. One ~75 MB binary, no Python, no PyTorch.

This is the reference documentation. It covers installing and running the
engine, the architecture of the kernel dispatch pipeline, per-crate references,
and operational guides. The rustdoc API reference is generated from the source
and published alongside it.

## Part I — Getting Started

- [Philosophy](https://docs.atlasinference.io/getting-started/philosophy.html)
- [Installation](https://docs.atlasinference.io/getting-started/installation.html)
- [Quickstart](https://docs.atlasinference.io/getting-started/quickstart.html)
- [Supported Models](https://docs.atlasinference.io/getting-started/models.html)
- [Troubleshooting](https://docs.atlasinference.io/getting-started/troubleshooting.html)

## Part II — Architecture

- [Philosophy: AI Kernel HyperCompiling](https://docs.atlasinference.io/architecture/philosophy.html)
- [Workspace Layout](https://docs.atlasinference.io/architecture/workspace.html)
- [Kernel Dispatch Pipeline](https://docs.atlasinference.io/architecture/dispatch.html)
- [SBIO: Business Logic vs I/O](https://docs.atlasinference.io/architecture/sbio.html)

## Part III — The Crates

- [atlas-core](https://docs.atlasinference.io/crates/atlas-core.html)
- [atlas-kernels](https://docs.atlasinference.io/crates/atlas-kernels.html)
- [spark-runtime](https://docs.atlasinference.io/crates/spark-runtime.html)
- [spark-comm](https://docs.atlasinference.io/crates/spark-comm.html)
- [spark-model](https://docs.atlasinference.io/crates/spark-model.html)
- [spark-server](https://docs.atlasinference.io/crates/spark-server.html)
- [atlas-spark-bench](https://docs.atlasinference.io/crates/atlas-spark-bench.html)

## Part IV — Engineering Deep Dives

- [CUDA Kernel Engineering](https://docs.atlasinference.io/deep-dives/kernels.html)
- [NVFP4 Quantization](https://docs.atlasinference.io/deep-dives/nvfp4.html)
- [FP8 Native Serving](https://docs.atlasinference.io/deep-dives/fp8.html)
- [Attention & Paged KV Cache](https://docs.atlasinference.io/deep-dives/attention.html)
- [MoE Routing & Experts](https://docs.atlasinference.io/deep-dives/moe.html)
- [SSM / Mamba / GDN Layers](https://docs.atlasinference.io/deep-dives/ssm.html)
- [Speculative Decoding (MTP)](https://docs.atlasinference.io/deep-dives/mtp.html)
- [Constrained Decoding (XGrammar)](https://docs.atlasinference.io/deep-dives/xgrammar.html)

## Part V — Operating Atlas

- [OpenAI-Compatible Server](https://docs.atlasinference.io/operations/server.html)
- [Tool Calling & Streaming](https://docs.atlasinference.io/operations/tools.html)
- [Multi-GPU & EP=2](https://docs.atlasinference.io/operations/multi-gpu.html)
- [Benchmarking](https://docs.atlasinference.io/operations/benchmarks.html)

## Part VI — The Project

- [Contributing](https://docs.atlasinference.io/project/contributing.html)
- [The Merge Lattice](https://docs.atlasinference.io/project/merge-lattice.html)
- [Security Policy](https://docs.atlasinference.io/project/security.html)
- [Release Notes](https://docs.atlasinference.io/project/releases.html)

## Appendix

- [Paper Summary](https://docs.atlasinference.io/appendix/paper.html)
- [A Category-Theoretic Perspective](https://docs.atlasinference.io/appendix/category-theory.html)
- [Glossary](https://docs.atlasinference.io/appendix/glossary.html)
- [Further Reading](https://docs.atlasinference.io/appendix/reading.html)

## Machine-readable

- [Full API reference](https://docs.atlasinference.io/api/): rustdoc for every public crate
- [Print view](https://docs.atlasinference.io/print.html): the entire book as one HTML document

## Optional

- [Atlas Inference](https://atlasinference.io): the project site — also at https://atlasinference.io/llms.txt
- [Engineering blog](https://blog.atlasinference.io): measured notes — also at https://blog.atlasinference.io/llms.txt
- [Source](https://github.com/Avarok-Cybersecurity/atlas): pure Rust and CUDA, AGPL-3.0-only
- [Recipes](https://github.com/Avarok-Cybersecurity/atlas-recipes): the model SSOT
- [Discord](https://discord.gg/RQcGakU2jW)