---
title: Compiler Tetris — Toy-IR RL Demo
emoji: 🧩
colorFrom: gray
colorTo: indigo
sdk: docker
sdk_version: 5.12.0
app_file: app.py
pinned: false
license: mit
---

# Compiler Tetris (Toy-IR) — Hugging Face Space

Interactive **CPU** demo for the Meta / OpenEnv-style **compiler phase-ordering** project:

- **Deliverable 2 — state translation:** JSON Toy-IR → compact pseudo-assembly for the LLM.
- **Deliverable 2 — format enforcement:** resilient extraction of a JSON **pass array** from noisy model text.
- **CompilerOptimizationEnv** (mock Role 1 engine + mock passes): step through passes and inspect rewards.

Full **GRPO / Unsloth** training is not run here (heavy GPU + long installs). Use your notebooks on Colab or a GPU Space for training.

## Source notebooks (parent repo)

- `compiler_optimization_grpo.ipynb`
- `role2_deliverable3_training_loop (2) (1).ipynb`
- `compiler_tetris (1).ipynb`
- `metahack1 (1).ipynb`

## Deploy this folder as a new Space

1. Create a new Space on Hugging Face (SDK: **Gradio**).
2. Upload the contents of this `hf_space/` directory to the Space repository root (`app.py`, `requirements.txt`, `README.md`).
3. Optional: add `Dockerfile` only if you switch the Space to **Docker** (see below).

## Optional: Docker SDK instead of Gradio SDK

If you want the Space to build from `Dockerfile`, change the YAML header to:

```yaml
sdk: docker
```

and remove Gradio-specific keys (`sdk_version`, `app_file`). The container runs `python app.py`, which listens on the `PORT` environment variable provided by Spaces.

## Local run

```bash
cd hf_space
pip install -r requirements.txt
python app.py
```

Then open `http://127.0.0.1:7860`.