AI Roadmaps
← All roadmaps
AdvancedRoadmap 10 · 15 steps · 6h 40m of video

Run, Serve & Scale AI

Get models into production and make them fast: Docker, MLOps, monitoring, quantization, vLLM, GPUs, and FSDP.

Before this: Deep Learning. LLMs from Scratch for the inference steps.

  1. 2 min
    Step 1

    Docker in 100 seconds

    The tool every deployment starts with.

    Docker in 100 Seconds

    Fireship

  2. 26 min
    Step 2

    Docker for an ML project

    Package a model so it runs anywhere.

    Docker Simply Explained with a Machine Learning Project for Beginners

    Python Simplified

  3. 19 min
    Step 3

    Serve a model with FastAPI

    Turn a trained model into an HTTP API.

    Deploy ML models with FastAPI, Docker, and Heroku | Tutorial

    AssemblyAI

    GokuMohandas/Made-With-ML★ 50.1k
  4. 26 min
    Step 4

    MLOps explained

    Versioning, pipelines, deployment, and monitoring.

    MLOps Explained - What It Is, Why You Need It and How It Works

    TechWorld with Nana

    DataTalksClub/mlops-zoomcamp★ 15.6k
  5. 51 min
    Step 5

    Track experiments with MLflow

    Never lose the settings behind a good result.

    MLFlow Tutorial | ML Ops Tutorial

    codebasics

    mlflow/mlflow★ 28.9k
  6. 30 min
    Step 6

    Monitor models for drift

    Compare live inputs to training data and flag drift with Evidently reports.

    Evidently AI Tutorial-Open Source ML Models Monitoring and Observability

    Krish Naik

  7. 26 min
    Step 7

    Quantization

    8-bit, 4-bit, and GGUF: smaller models, small quality loss.

    Quantizing LLMs - How & Why (8-Bit, 4-Bit, GGUF & More)

    Adam Lucek

  8. 9 min
    Step 8

    llama.cpp

    The engine that runs LLMs on laptops and phones.

    What Is Llama.cpp? The LLM Inference Engine for Local AI

    IBM Technology

    ggml-org/llama.cpp★ 135.8k
  9. 9 min
    Step 9

    The KV cache

    Why inference memory grows with context length.

    The KV Cache: Memory Usage in Transformers

    Efficient NLP

  10. 32 min
    Step 10

    vLLM and PagedAttention

    Serve many users at once without wasting GPU memory.

    Fast LLM Serving with vLLM and PagedAttention

    Anyscale

    vllm-project/vllm★ 97.3k
  11. 25 min
    Step 11

    Speculative decoding

    A small draft model proposes tokens; the big model verifies several in one pass.

    How LLMs Get Faster Without Changing Their Outputs | Speculative Decoding

    Jia-Bin Huang

  12. 10 min
    Step 12

    Inference engines

    How serving systems batch, cache, and schedule.

    Inference Engines explained in 10min..

    Caleb Writes Code

  13. 28 min
    Step 13

    How GPUs work

    The hardware every model runs on, beautifully animated.

    How do Graphics Cards Work? Exploring GPU Architecture

    Branch Education

  14. 1h 14m
    Step 14

    GPUs for LLMs

    Stanford CS336: memory hierarchy, arithmetic intensity, and why kernels matter.

    Stanford CS336 I Language Modeling from Scratch | Spring 2025 | Lecture 5: GPUs

    Stanford Online

    gpu-mode/lectures★ 7.1k
  15. 33 min
    Step 15

    Sharded training with FSDP

    Train models that don't fit on one GPU.

    How Fully Sharded Data Parallel (FSDP) works?

    Ahmed Taha

    stas00/ml-engineering★ 19.7k

Build it

Watching isn't learning. Pick one and put it on GitHub.

Next roadmap