🧠 Bringing Modern AI Inference to Ancient Silicon

retrolm is a transformer-based LLM inference engine written from scratch in C and x86 Assembly, designed to run on Intel Pentium II hardware with FreeDOS.

🛠️ The Challenge

• Zero Dependencies: No PyTorch, NumPy, or BLAS—all linear algebra written from scratch
• No AI Assistance: Core implementation written without LLMs (documentation and tests were LLM-assisted for efficiency)
• 32-bit x86 Assembly: Hand-optimized routines for critical operations
• Classic References: K&R's C Programming Language and Hyde's Art of Assembly Language
• Cross-Compiled: DJGPP toolchain targeting FreeDOS

---

🖥️ Target Hardware

Component	Minimum	Recommended
CPU	Pentium II 233MHz	Pentium II 400MHz+
RAM	64 MB	128 MB
OS	FreeDOS 1.2+	FreeDOS 1.3

---

🔧 Quick Start

Prerequisites: Docker Desktop

# Clone and setup
git clone https://github.com/agaz1985/retrolm.git
cd retrolm
chmod +x *.sh
make build

# Development
make run          # Fast: Build & run on Linux (32-bit)
make dos          # Deploy: Build retrolm.exe for FreeDOS
make shell        # Debug: Interactive container shell
make clean        # Remove build artifacts

Deploy to Pentium II:

1. make dos → creates build/retrolm.exe
2. Transfer via USB/floppy/serial to FreeDOS machine
3. Run: C:\> retrolm.exe

Toolchain: NASM (asm) • GCC -m32 (Linux) • DJGPP (DOS) • Docker/Ubuntu 22.04

---

🤔 Why?

An exploration of what happens when modern AI meets 1998 hardware constraints—deep diving into low-level optimization, memory management, and vintage architecture.

---

📚 References

• The C Programming Language (2nd Ed.) - Kernighan & Ritchie
• The Art of Assembly Language - Randall Hyde

Built with determination, coffee, and deep appreciation for vintage computing. ☕