README.md

---
title: VLIW Kernel Optimizer
emoji: "⚡"
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: false
license: mit
---

# VLIW Kernel Optimization via Reinforcement Learning

Train a language model to generate optimized VLIW/SIMD kernels using test-time RL training.

## Goal
- **Baseline:** 147,734 cycles
- **Target:** 1,363 cycles (108x speedup)

## How it works
1. Model generates kernel code
2. Simulator evaluates cycle count
3. RL training improves the model based on rewards

## Usage
1. Select a model (Qwen2.5-Coder-7B recommended)
2. Configure training steps (50 recommended)
3. Click "Start Training"
4. Monitor progress - training continues even if you close the browser

## Hardware
Requires A10G GPU (HF Spaces Pro)