ML Infrastructure
The HuggingFace Training Stack
A visual map of the full HuggingFace ecosystem — Transformers, Datasets, PEFT, Accelerate, and TRL — showing how each library fits together in a real fine-tuning pipeline.
Read guide →Visual, interactive deep-dives into ML concepts — built to make complex ideas stick. New guides published regularly.
A visual map of the full HuggingFace ecosystem — Transformers, Datasets, PEFT, Accelerate, and TRL — showing how each library fits together in a real fine-tuning pipeline.
Read guide →Why language models need GPUs, where memory goes during training and inference, and exactly what happens when it runs out — with an interactive OOM simulator.
Read guide →From full fine-tuning to LoRA and QLoRA — a comprehensive visual guide covering when to fine-tune, how to choose a strategy, and the key hyperparameters that matter most.
Read guide →A beginner's guide to pretraining a small language model from scratch — covering tokenisation, architecture, data, memory, and training config through the lens of a real 125M-parameter model.
Read guide →Part 2: the base model was fluent but couldn't answer questions. This guide covers building a 10,000-pair synthetic QA dataset, full fine-tuning on a single GPU, loss masking, overfitting, and the exact math behind SFT.
Read guide →From the original 2017 "Attention is All You Need" paper to MQA, GQA, MLA, and FlashAttention — a visual field guide tracing how attention mechanisms evolved and why each innovation mattered.
Read guide →Why transformers need positional encoding, how RoPE encodes position as rotation in the complex plane, and why it became the dominant positional scheme from LLaMA to Qwen — built from first principles.
Read guide →Anthropic's "Building Effective Agents" taxonomy, visually: the augmented LLM, four workflow patterns, and the ReAct agent loop — each with a diagram and runnable Python.
Read guide →Before understanding any ML algorithm — from linear regression to large language models — you need to understand what ML is actually trying to compute. Builds that foundation from first principles.
Read guide →