File size: 2,383 Bytes
0feb6c8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 | ---
title: PyTorch Debug Env
emoji: 🔥
colorFrom: orange
colorTo: red
sdk: docker
app_port: 7860
short_description: Multi-step RL environment for diagnosing broken PyTorch training jobs
tags:
- openenv
- pytorch
- reinforcement-learning
- debugging
- ml-training
- agent
pinned: true
---
# PyTorch Debug Env 🔥
A complete [OpenEnv](https://meta-pytorch.org/OpenEnv/) environment for the **Meta PyTorch Hackathon** where an AI agent investigates and diagnoses broken PyTorch training jobs.
## Quick Start
```python
from openenv import AutoEnv, AutoAction
env = AutoEnv.from_env("ArchCoder/pytorch-debug-env")
Action = AutoAction.from_env("ArchCoder/pytorch-debug-env")
with env.sync() as client:
result = client.reset(task_id="easy")
action = Action(
current_hypothesis={
"bug_type": "missing_zero_grad",
"affected_file": "train.py",
"confidence": 0.7
},
commit_diagnosis=False
)
step_result = client.step(action)
```
## API Endpoints
| Endpoint | Method | Description |
|----------|--------|-------------|
| `/` | GET | Environment info |
| `/health` | GET | Health check |
| `/reset?task_id=easy` | POST | Start new episode |
| `/step` | POST | Submit hypothesis + action |
| `/state` | GET | Current episode state |
## Tasks
| Task | Difficulty | Description |
|------|-----------|-------------|
| `easy` | ⭐ | Single-file bug — missing `zero_grad`, wrong loss |
| `medium` | ⭐⭐ | Multi-file root cause — data leakage, scheduler mismatch |
| `hard` | ⭐⭐⭐ | Silent failure — memory leak, AMP overflow, red herrings |
## Reward Structure
- **Hypothesis delta** (60%) — reward for improving your bug hypothesis each step
- **Investigation** (20%) — reward for inspecting the right files
- **Final diagnosis** (20%) — accuracy of committed diagnosis vs ground truth
Scores range from `0.0` to `1.0`. Partial credit for correct bug category on hard tasks.
## Environment State
Each episode provides a synthetic PyTorch repo with:
- Source files (`train.py`, `model/`, `data/`, `config/`)
- Loss curves and GPU memory profiles
- Training logs with realistic noise and red herrings
The agent reveals files progressively across up to 5–6 steps, refining its hypothesis before committing a final diagnosis.
## Author
**Priyansh Saxena** — IIIT Gwalior
|