Summary

Pyre Code is a self-hosted ML coding practice platform enabling users to implement modern AI system internals from scratch. Targeting ML engineers preparing for interviews and developers seeking hands-on learning, it offers 68 problems covering diverse AI concepts. Users receive instant feedback via a local grading service, fostering practical implementation skills without GPU requirements or complex setup.

How It Works

Users implement solutions in a browser-based Monaco editor. A local FastAPI grading service, powered by torch_judge, executes submissions against predefined tests, providing immediate pass/fail results. This test-driven approach reinforces practical understanding of algorithms and architectures, with progress tracked locally via SQLite.

Quick Start & Requirements

• Prerequisites: Python 3.11+, Node.js 18+.
• Installation: Recommended: git clone https://github.com/whwangovo/pyre-code.git && cd pyre-code && ./setup.sh. Alternatives include Conda, manual venv, or Docker (docker compose up --build).
• Running: Web app at http://localhost:3000, grading service at http://localhost:8000.
• Links: GitHub Repository
• Resource Footprint: Fully local, no GPU needed. Docker persists progress via volumes.

Highlighted Details

• Problem Set: 68 problems cover fundamentals, normalization, attention variants, position encodings, architectures (GPT-2, ViT, MoE), training, distributed systems, inference optimizations (KV cache, quantization), alignment (RLHF), diffusion models, and SSMs (Mamba).
• Learning Paths: Curated paths include "Transformer Internals," "Train a GPT from Scratch," and "Inference & Distributed Training."
• Test-Driven: Focuses on passing specific test cases for practical understanding.
• No GPU Required: Accessible without specialized hardware.

Maintenance & Community

Contributions are welcomed via PRs for new problems, bug fixes, or documentation. Explicit community channels or maintainer details are not provided.

Licensing & Compatibility

Distributed under the permissive MIT License, allowing commercial use and integration into closed-source projects. The judge engine is also MIT licensed.

Limitations & Caveats

Designed for local learning, not production deployment. Requires specific Python (3.11+) and Node.js (18+) versions. Setup involves managing local development environments.