Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Machine Learning, Deep Learning & AI Reference Library

This page is a verified index of primary sources for machine learning & ai: official documentation, developer and API portals, source repositories, SDKs, downloadable or offline documentation, a two-track learning path, and free-access research literature.

It is a navigation layer, not a tutorial. Where the rest of this book explains a concept, this page tells you which document to open to get the authoritative answer, and in what order to read things. Every link was HTTP-verified on the date shown below; sources that block automated checkers but work in a browser are flagged rather than silently dropped.

Frameworks, model hubs, GPU kernels and ML compilers, distributed training, inference serving, MLOps tooling, and a two-track path from building backpropagation by hand to reading vLLM and writing Triton kernels.

89 entries across 7 categories, plus 57 education & reference-implementation resources (25 basic / 32 advanced), plus 22 video resources and 8 conference sources.

Every link HTTP-verified on 2026-10-07.

Google devsite properties (tensorflow.org, ai.google.dev) redirect automated clients to an authentication page; they load normally in a browser.

Contents

1. Core frameworks

PyTorch

  • Docs: pytorch.org/docs/stable/index.html
  • Developer / API: pytorch.org/tutorials
  • Source: github.com/pytorch/pytorch
  • SDKs & repos: torchvision, torchaudio, torchtune, TorchServe; C++ and Python APIs
  • Downloadable / offline: Docs built per release and downloadable; source tree cloneable
  • Note: The research default. torch.compile and the dispatcher are where the interesting engineering is — see also the PL/compilers index.

TensorFlow

  • Docs: tensorflow.org/api_docs
  • Developer / API: tensorflow.org
  • Source: github.com/tensorflow/tensorflow
  • SDKs & repos: Keras, TFX, TF Lite / LiteRT, TensorFlow.js, TF Serving
  • Downloadable / offline: API docs per version
  • Note: Still dominant in some production and mobile settings. Google’s devsite redirects automated clients to an auth page; loads fine in a browser.

JAX

  • Docs: docs.jax.dev/en/latest
  • Source: github.com/jax-ml/jax
  • SDKs & repos: Flax, Optax, Haiku, Equinox; XLA backend
  • Downloadable / offline: Sphinx docs
  • Note: Composable transforms (grad, jit, vmap, pmap) over NumPy semantics. The cleanest mental model of autodiff in any framework.

Keras

scikit-learn

XGBoost

LightGBM

PyTorch Lightning

fastai

  • Docs: docs.fast.ai
  • Source: github.com/fastai/fastai
  • SDKs & repos: Layered API over PyTorch
  • Downloadable / offline: Docs site
  • Note: The library and the course are designed together; use both or neither.

2. Models, hubs & transformer libraries

Hugging Face

  • Docs: huggingface.co/docs
  • Developer / API: huggingface.co/docs/api-inference/index
  • Source: github.com/huggingface
  • SDKs & repos: transformers, datasets, tokenizers, accelerate, peft, trl, diffusers
  • Downloadable / offline: All docs per library; models and datasets downloadable
  • Note: The de facto model distribution layer for the field. The transformers source is also the best cross-architecture reference for how models are actually implemented.

Transformers

Diffusers

OpenAI API

Anthropic / Claude API

Google Gemini API

Mistral AI

Llama

Ollama

3. GPU compute, kernels & compilers

CUDA

cuDNN

Triton (OpenAI)

TensorRT

OpenXLA / XLA

  • Docs: openxla.org/xla
  • Source: github.com/openxla/xla
  • SDKs & repos: Compiler backend for JAX and TensorFlow; StableHLO
  • Downloadable / offline: Docs site
  • Note: Read alongside MLIR — XLA is one of the best-documented production ML compilers.

MLIR

Apache TVM

ONNX

  • Docs: onnx.ai/onnx
  • Developer / API: onnx.ai/onnx/intro
  • Source: github.com/onnx/onnx
  • SDKs & repos: Model interchange format; operator spec
  • Downloadable / offline: Spec in-repo
  • Note: Read the operator spec when debugging export problems — most issues are unsupported ops.

ONNX Runtime

4. Training at scale & inference serving

DeepSpeed

Megatron-LM

Ray Train

vLLM

  • Docs: docs.vllm.ai/en/latest
  • Source: github.com/vllm-project/vllm
  • SDKs & repos: PagedAttention, continuous batching, OpenAI-compatible server
  • Downloadable / offline: Docs site
  • Note: The current default for serving open-weight LLMs. The PagedAttention design is worth reading as a systems idea.

llama.cpp

NVIDIA Triton Inference Server

TorchServe

BentoML

KServe

5. Experiment tracking, data & orchestration

MLflow

Weights & Biases

DVC

Kubeflow

LangChain

LlamaIndex

6. Fine-tuning, evaluation, retrieval & RL

PEFT

  • Docs: huggingface.co/docs/peft
  • Source: github.com/huggingface/peft
  • SDKs & repos: LoRA, QLoRA, prefix tuning, adapters
  • Downloadable / offline: Docs site
  • Note: Parameter-efficient fine-tuning is how most people actually adapt models. Read the LoRA docs before the paper.

TRL

Unsloth

  • Docs: docs.unsloth.ai
  • Source: github.com/unslothai/unsloth
  • SDKs & repos: Memory- and speed-optimized fine-tuning with hand-written Triton kernels
  • Downloadable / offline: Docs + free Colab notebooks
  • Note: The kernels are worth reading if you want to see Triton used for real work.

Axolotl

FlashAttention

SGLang

  • Docs: docs.sglang.ai
  • Source: github.com/sgl-project/sglang
  • SDKs & repos: High-throughput LLM serving with RadixAttention prefix caching
  • Downloadable / offline: Docs site
  • Note: The main alternative to vLLM. Comparing their scheduling designs is instructive.

TensorRT-LLM

lm-evaluation-harness

FAISS

Sentence Transformers

Gymnasium

CleanRL

  • Docs: docs.cleanrl.dev
  • Source: github.com/vwxyzjn/cleanrl
  • SDKs & repos: Single-file implementations of RL algorithms with logged benchmark runs
  • Downloadable / offline: Docs site
  • Note: Every algorithm in one readable file, with reproduced results. The best way to actually understand an RL algorithm.

Stable-Baselines3

Flax

Optax

  • Docs: optax.readthedocs.io
  • Source: github.com/google-deepmind/optax
  • SDKs & repos: Composable gradient transformations and optimizers for JAX
  • Downloadable / offline: Sphinx docs
  • Note: Reading Optax is the clearest way to understand what an optimizer actually does.

Mojo

7. Research papers & open-access literature

arXiv

  • Docs: arxiv.org
  • Developer / API: info.arxiv.org/help/api/index.html
  • SDKs & repos: Preprints across all of CS; most systems, ML and PL work appears here before publication
  • Downloadable / offline: Every paper is a free PDF. Bulk access documented at info.arxiv.org/help/bulk_data/index.html; a full-text API at export.arxiv.org
  • Note: Not peer-reviewed. Treat an arXiv-only paper as a claim, not a result — but it is where you will read almost everything first.

ar5iv

  • Docs: ar5iv.labs.arxiv.org
  • SDKs & repos: Renders any arXiv paper as responsive HTML instead of PDF
  • Downloadable / offline: Free; swap arxiv.org/abs/ID for ar5iv.labs.arxiv.org/html/ID
  • Note: Makes papers readable on a phone and searchable in-page. Underused.

alphaXiv

  • Docs: alphaxiv.org
  • SDKs & repos: arXiv papers with a public comment and discussion layer
  • Downloadable / offline: Free
  • Note: Useful when a paper is contested — the discussion often contains the critique you were looking for.

Semantic Scholar

  • Docs: semanticscholar.org
  • Developer / API: api.semanticscholar.org/graph/v1
  • SDKs & repos: 200M+ papers with citation graph, influential-citation scoring and TLDR summaries
  • Downloadable / offline: Free Graph API (semanticscholar.org/product/api), bulk datasets available on request
  • Note: The best free citation graph. ‘Highly influential citations’ is a genuinely useful filter for finding what actually mattered.

OpenAlex

  • Docs: openalex.org
  • Developer / API: api.openalex.org/works
  • SDKs & repos: Fully open catalogue of works, authors, venues and institutions; successor to Microsoft Academic Graph
  • Downloadable / offline: Entirely free API with no key required; complete database snapshots downloadable
  • Note: The only large-scale bibliographic database that is open all the way down, including bulk snapshots.

DBLP

  • Docs: dblp.org
  • Developer / API: dblp.org/faq/13501473.html
  • SDKs & repos: Authoritative CS bibliography — complete author and venue listings
  • Downloadable / offline: Free; full XML dump downloadable
  • Note: The fastest way to find everything a given researcher has published, and to see a conference’s full programme by year.

OpenReview

  • Docs: openreview.net
  • Developer / API: docs.openreview.net
  • Source: github.com/openreview
  • SDKs & repos: ICLR, NeurIPS, COLM and dozens of other venues — papers plus the full review threads
  • Downloadable / offline: Free; REST API
  • Note: Reading the reviews and author rebuttals teaches you how the field evaluates work. Nothing else exposes this.

CORE

  • Docs: core.ac.uk
  • Developer / API: core.ac.uk/services/api
  • SDKs & repos: Aggregates 300M+ open-access papers from repositories worldwide
  • Downloadable / offline: Free API and bulk datasets
  • Note: Good for finding the green open-access copy when a publisher’s version is paywalled.

Unpaywall

  • Docs: unpaywall.org
  • Developer / API: api.unpaywall.org
  • SDKs & repos: Finds legal free copies of paywalled papers via DOI
  • Downloadable / offline: Free API; browser extension
  • Note: Legal, author-deposited copies only. Install the extension and most paywalls simply stop appearing.

Papers We Love

  • Docs: paperswelove.org
  • Source: github.com/papers-we-love/papers-we-love
  • SDKs & repos: A curated, categorised repository of classic CS papers with local meetup talks
  • Downloadable / offline: Repo cloneable; many PDFs mirrored in-repo
  • Note: The best starting point if you do not yet know which papers matter in a subfield.

The Morning Paper (archive)

  • Docs: blog.acolyer.org
  • SDKs & repos: Adrian Colyer’s daily paper summaries, 2014-2021
  • Downloadable / offline: Free archive, still online
  • Note: No longer updated, but the back catalogue of ~1000 summarised papers is one of the great free CS resources.

USENIX Proceedings

  • Docs: usenix.org/publications/proceedings
  • SDKs & repos: OSDI, SOSP (co-published), NSDI, ATC, FAST, Security — the core systems venues
  • Downloadable / offline: Every paper free, immediately, with no membership. Often with recorded talks
  • Note: USENIX made everything open access years before the rest of the field. If a systems paper exists, check here first.

ACM Digital Library

  • Docs: dl.acm.org
  • SDKs & repos: SIGMOD, ASPLOS, PLDI, POPL, SoCC and the ACM journals
  • Downloadable / offline: Partly open: ACM Open and author-paid OA papers are free; others are paywalled
  • Note: 403s to automated clients; loads in a browser. For paywalled items, check arXiv, the author’s homepage or Unpaywall first — the free copy usually exists.

IEEE Xplore

  • Docs: ieeexplore.ieee.org
  • SDKs & repos: ISCA, MICRO, HPCA and the IEEE journals
  • Downloadable / offline: Mostly paywalled; abstracts free
  • Note: Almost always worth searching the author’s page or arXiv instead. Architecture authors in particular post preprints widely.

DROPS / LIPIcs (Dagstuhl)

  • Docs: drops.dagstuhl.de
  • SDKs & repos: ECOOP, ITP, CONCUR, SAT and many theory venues
  • Downloadable / offline: 100% open access, free PDFs, Creative Commons licensed
  • Note: A fully open publisher. Every paper, always free, no exceptions.

arXiv cs.LG / cs.CL / cs.CV

Papers with Code

  • Docs: paperswithcode.com
  • SDKs & repos: Papers linked to implementations, with benchmark leaderboards by task
  • Downloadable / offline: Free; datasets and results available in bulk
  • Note: The leaderboards are useful for orientation but gameable — check whether the comparison is fair before believing a ranking.

Hugging Face Papers

  • Docs: huggingface.co/papers
  • SDKs & repos: Daily curated arXiv selection with community discussion and linked model weights
  • Downloadable / offline: Free
  • Note: The most practical daily filter on the arXiv firehose, because papers are linked to runnable artifacts.

NeurIPS Proceedings

  • Docs: papers.nips.cc
  • SDKs & repos: Every NeurIPS paper since 1987
  • Downloadable / offline: All free PDFs, plus supplementary material
  • Note: Fully open archive going back nearly four decades.

PMLR

  • Docs: proceedings.mlr.press
  • SDKs & repos: Proceedings of Machine Learning Research — ICML, AISTATS, CoLLAs, UAI and more
  • Downloadable / offline: 100% open access
  • Note: If it was published at ICML, it is free here. No exceptions.

JMLR

  • Docs: jmlr.org
  • SDKs & repos: Journal of Machine Learning Research
  • Downloadable / offline: Open access since 2000
  • Note: Longer, more rigorous papers than the conference venues. Founded specifically as an open-access protest.

ACL Anthology

  • Docs: aclanthology.org
  • Source: github.com/acl-org/acl-anthology
  • SDKs & repos: Every NLP paper from ACL, EMNLP, NAACL, COLING and more — 100,000+ papers
  • Downloadable / offline: All free; full BibTeX and metadata dumps in the repo
  • Note: The model open-access archive. Complete, searchable, and permanently free.

CVF Open Access

  • Docs: openaccess.thecvf.com/menu
  • SDKs & repos: CVPR, ICCV, WACV papers in full
  • Downloadable / offline: All free
  • Note: The entire computer vision conference output, openly hosted.

OpenReview

  • Docs: openreview.net
  • Developer / API: docs.openreview.net
  • SDKs & repos: ICLR, NeurIPS and others with full review threads
  • Downloadable / offline: Free; API
  • Note: Read the reviews. For contested papers this is more informative than the paper.

MLCommons / MLPerf

  • Docs: mlcommons.org/benchmarks
  • SDKs & repos: Standardised training and inference benchmarks with audited submissions
  • Downloadable / offline: Free results and rules
  • Note: The only ML performance numbers with a credible verification process behind them.

Google Research publications

  • Docs: research.google/pubs
  • SDKs & repos: Transformers, BERT, TPU architecture, MapReduce, Spanner
  • Downloadable / offline: Free PDFs
  • Note: Search here before arXiv for anything Google-originated.

Microsoft Research publications

Apple Machine Learning Research

Amazon Science

Lil’Log

  • Docs: lilianweng.github.io
  • SDKs & repos: Long-form technical surveys of research areas
  • Downloadable / offline: Free
  • Note: Usually the best single explainer on a topic, with full citations. Start here before the primary papers.

Distill

  • Docs: distill.pub
  • Source: github.com/distillpub
  • SDKs & repos: Interactive visual explanations of ML concepts
  • Downloadable / offline: Free, archived
  • Note: No longer publishing, but the back catalogue remains the high-water mark for explaining ML visually.

Education & reference implementations

Two tracks: Basic builds the foundations, Advanced is about reading and extending real implementations. Everything listed is free and publicly accessible.

Basic

25 resources across 5 topics.

Build it from scratch first

  • Neural Networks: Zero to Hero (Karpathy) — Build backpropagation, then a language model, then GPT — on video, in order, from nothing. The best free introduction to deep learning that exists.
  • micrograd — An autograd engine in roughly 100 lines. Read every line; autodiff stops being magic.
  • nanoGPT — A GPT you can train yourself and actually read. ~300 lines of model code.
  • nn-zero-to-hero repo — The notebooks accompanying the video series.

Structured courses

  • fast.ai Practical Deep Learning — Top-down: get working models first, derive theory later. Free, and effective for people who learn by doing.
  • Stanford CS231n — Computer vision and CNNs. Notes and assignments public; the notes remain among the clearest written.
  • Stanford CS224n — NLP with deep learning. Slides and assignments public.
  • Stanford CS229 — Classical machine learning done properly. The lecture notes are a standalone reference.
  • MIT Introduction to Deep Learning — A compact bootcamp with all lectures and labs free.
  • Hugging Face courses — Free courses on NLP, diffusion, RL, audio and agents, each with runnable notebooks.

Books, free and complete

The libraries to learn first

  • PyTorch tutorials — Start with the 60-minute blitz, then the training loop from scratch.
  • PyTorch docs — Learn to read these rather than searching for examples.
  • scikit-learn — Do not skip classical ML. Most real problems are still tabular.
  • XGBoost — On structured data this is still frequently the right answer.
  • Hugging Face Transformers — The practical path to using pretrained models.
  • Ollama — Run open-weight models on your own machine in one command.

Finding and reading papers

  • Papers We Love — Start here when you do not yet know which papers matter. Curated by subfield, with recorded talks.
  • The Morning Paper archive — Around a thousand papers summarised in plain language. No longer updated; still one of the best free CS resources.
  • Semantic Scholar — Free citation graph. The ‘highly influential citations’ filter is the fastest way to find what a paper actually changed.
  • Unpaywall — Install the extension. Most paywalls stop appearing, legally, because the author deposited a copy.
  • ar5iv — Read any arXiv paper as HTML instead of a two-column PDF. Swap arxiv.org/abs for ar5iv.labs.arxiv.org/html.

Advanced

32 resources across 6 topics.

Graduate coursework

Read production model code

  • Transformers source — Every major architecture in one consistent structure. Unmatched as a comparative reference.
  • nanoGPT — Return to it after reading production code; the minimal version clarifies the essentials.
  • Diffusers — Schedulers and pipelines for diffusion models, readable.
  • llama.cpp — Quantization, GGUF, and CPU inference — a systems education in C++.
  • vLLM — PagedAttention and continuous batching; a genuinely novel systems contribution.

GPU kernels & performance

  • Triton tutorials — Write fused GPU kernels in Python. Work through the matmul and attention tutorials.
  • CUDA Programming Guide — The memory hierarchy and occupancy chapters determine most real performance.
  • CUDA samples — Reference kernels to compare your own against.
  • cuDNN docs — What the primitives underneath your framework actually do.
  • TensorRT — Inference-side graph optimization and precision calibration.

ML compilers & intermediate representations

  • OpenXLA / XLA — Operator fusion and layout assignment, documented properly.
  • MLIR — The dialect infrastructure under most modern ML compilers. Also read the PL/compilers index.
  • Apache TVM — Autotuning and codegen for heterogeneous targets.
  • ONNX operator spec — Read the spec when export breaks — it usually will.
  • ONNX Runtime — Execution providers and graph optimizations.

Distributed training

  • DeepSpeed — ZeRO stages 1-3 and offloading, with papers alongside the code.
  • Megatron-LM — Tensor and pipeline parallelism, reference implementation.
  • PyTorch distributed docs — FSDP, DDP and the collective primitives underneath.
  • Ray Train — Orchestration across the cluster; see the distributed-systems index too.
  • JAX — pmap, shard_map and explicit sharding give the clearest model of parallelism in any framework.

Keeping current & research

  • arXiv cs.LG — Where the field publishes first. Unfiltered, so pair with curation.
  • Papers with Code — Papers linked to implementations and benchmark tables.
  • Lil’Log (Lilian Weng) — Long-form technical surveys; usually the best single explainer on a new topic.
  • Distill.pub — Archived but still the high-water mark for visual explanation of ML concepts.
  • NeurIPS — Proceedings free.
  • ICML — Proceedings free.
  • ICLR — Reviews are public on OpenReview — reading them teaches you how the field evaluates work.

Video courses, channels & talks

22 resources across 2 groups. Every channel and playlist below was fetched and title-verified on 2026-10-09. Handles drift and several plausible-looking handles resolve to the wrong channel, so a 200 response is not proof of identity — the links here were each checked against the channel title.

Channels & conference recordings

  • Andrej Karpathy — The best free depth available: backprop, transformers and GPT built from scratch on camera.
  • Two Minute Papers — Fast paper coverage; headlines overstate, so read the abstract before you cite.
  • Yannic Kilcher — Long-form paper reviews — literature depth you will not get from a blog.
  • 3Blue1Brown — Visual linear algebra and calculus — the maths prerequisite for all of ML.
  • StatQuest — The clearest intuition available for the statistics that ML assumes.
  • DeepLearning.AI — Short courses run by Andrew Ng & co.; vendor-leaning, verify against the papers.
  • Hugging Face — The open LLM ecosystem: transformers, datasets, agents, tool calling.
  • PyTorch — Official PyTorch tutorials and PyTorch Conference talks.
  • TensorFlow — Official TensorFlow channel — dev summit talks and tutorials.
  • OpenAI — Model releases and DevDay talks; primary for API and product behaviour.
  • Google DeepMind — Research overviews and model announcements from DeepMind.
  • sentdex — Applied Python ML/NLP walkthroughs; solid for getting something running.
  • ICML — Official ICML channel; plenaries and tutorials, coverage varies by year.
  • ICLR — Channel titled simply “Iclr” — check the video against the conference site before citing it.
  • The CVF — The Computer Vision Foundation: CVPR/ICCV material, fully open access.
  • Latent Space — Long-form practitioner interviews; the fastest way to find out what people are actually shipping.
  • TWIML AI Podcast — Interview archive going back years; uneven, occasionally excellent.
  • Google TechTalks — Foundational ML talks by the authors of the papers you are reading.

Lectures & playlists

Note: The only topic here where video is genuinely better than text for prerequisites; treat paper-summary channels as a discovery feed, never as a citation.

Conference videos, notes & archives

8 resources across 2 groups. Conference recordings are the primary-source tier of video: the speaker is usually an author of the paper, and where a talk exists the proceedings entry is often open at the same link. Every URL here returned 200 on 2026-10-09 unless the note says otherwise.

ML publishes more open proceedings than any other field here — and more video of people explaining other people’s papers, which is not the same thing.

Conference channels & video archives

  • NeurIPS proceedings — Every paper, with links to the authors’ sites; video coverage varies by year.
  • CVF Open Access — CVPR/ICCV/WACV papers with supplementary video — fully open, the model the other venues should copy.
  • FOSDEM video archive — AI/ML and data devrooms; small but free.
  • MLSys — The ML-systems conference: open proceedings, and the venue for the engineering half of ML (serving, compilation, distributed training).
  • CVPR 2026 — Conference site for the vision flagship; papers and supplementary video via the CVF open-access portal above.
  • KDD — The applied data-science conference; proceedings partly open, tutorials often published.
  • EMNLP 2026 — The NLP venue; papers in the ACL Anthology, recordings inconsistent.

Notes, proceedings & paper-adjacent archives

  • Hugging Face Papers — Daily trending papers with discussion attached; the fastest discovery feed, and a mediocre archive.

If you only do three things

  1. Watch Karpathy’s Zero to Hero in order and type the code yourself. micrograd, then makemore, then GPT.
  2. Read nanoGPT completely (~300 lines), then read modeling_llama.py in Transformers. That gap is production ML.
  3. Do the Triton tutorials. Once you can write a fused kernel, framework performance stops being a black box.

Honest notes

  • Do not skip classical ML. Most real problems are tabular, and XGBoost still beats deep models on them regularly.
  • Two different things are called Triton: OpenAI’s GPU kernel language and NVIDIA’s inference server. They are unrelated.
  • JAX gives the clearest mental model of autodiff and parallelism, even if you ship PyTorch.
  • The transformers source is better architecture documentation than most papers — one consistent structure across hundreds of models.
  • The Goodfellow book is still good on fundamentals and dated on architectures. Use it for theory, not for current practice.
  • Check framework version in any documentation you read. This ecosystem deprecates faster than any other in this series.