📚 Book Launch: Rearchitecting LLMs (MEAP) + How This Repo Fits In

[peremartra]

The Book is Live! 🎉

After months of development, "Rearchitecting LLMs" is now available in Manning's Early Access Program (MEAP).

📖 Get early access here - 50% off until Feb 12

What's Available Now

Chapters 1-2 are live:

• Ch1: Why architectural tailoring matters + the complete pipeline
• Ch2: Your first end-to-end project (depth pruning + knowledge distillation on Gemma-3-270m)

Coming soon (Ch3-12):

• Advanced depth pruning strategies
• Width pruning in GLU architectures
• Knowledge distillation techniques
• Specialized fine-tuning
• Attention optimization
• Fair Pruning (bias-aware optimization)
• Capstone project

How the Book and Repo Work Together

This Repository:

• ✅ Complete, executable notebooks
• ✅ All code runs on Colab free tier (T4 GPUs)
• ✅ Quick experimentation and learning by doing
• ✅ Issues, discussions, and community contributions

The Book:

• 📖 In-depth explanations of each technique
• 📊 Why certain approaches work (and when they don't)
• 📝 Research paper breakdowns
• 🎯 Decision frameworks for your specific use case
• 💡 Architectural intuition beyond the code

Think of it this way:

• Repo = Hands-on laboratory
• Book = Your detailed lab manual

What Makes This Different

This isn't another fine-tuning tutorial. You'll learn to:

1. Modify model architecture physically (not just weights)
2. Make informed trade-offs between size, speed, and capabilities
3. Understand the "why" behind each optimization decision
4. Bridge research papers to production code

Example from Ch2:

# Start: Gemma-3-270m (268M params, 18 layers)
pruned_model = prune_model_depth(model, layers_to_remove=2)
# Result: 257M params, 16 layers, ~80% capabilities

# Recover with knowledge distillation
final_model = distill_knowledge(teacher=model, student=pruned_model)
# Result: 95% capabilities recovered, 10% faster inference

All on a free T4 GPU. No TPU pods required.

Building in Public

As I write Ch3-12, I'll be sharing the process on X/Twitter:

• Experiment results before publication
• Failed attempts and unexpected findings
• Paper implementations and adaptations
• Code previews

Follow along: @PereMartra

Supporting Library: optiPfair

The techniques in the book are also available through the optiPfair library:

pip install optipfair

This is an open-source project—contributions welcome!

Get Involved

Questions about the book? Ask here or in the Manning forum (MEAP readers get access)

Found a bug in the notebooks? Open an issue!

Have optimization ideas? PRs welcome!

Want to discuss your specific use case? Drop a comment below 👇

Launch Offer

50% off MEAP until February 12

• Link: https://hubs.la/Q040tvsK0

---

What optimization challenges are you tackling? What would you like to see covered in upcoming chapters?

Let's discuss 👇