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