Spaces:
Sleeping
Sleeping
| # FastAPI server exposing the Data Centre OpenEnv environment (EnvClient-compatible). | |
| from fastapi.responses import HTMLResponse | |
| from openenv.core.env_server.http_server import create_app | |
| from .environment import DCEnvironment | |
| from .models import DCAction, DCObservation | |
| app = create_app( | |
| DCEnvironment, | |
| DCAction, | |
| DCObservation, | |
| env_name="datacenter_env", | |
| max_concurrent_envs=1, | |
| ) | |
| async def root(): | |
| return """<!DOCTYPE html> | |
| <html lang="en"> | |
| <head> | |
| <base target="_blank"> | |
| <meta charset="UTF-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"> | |
| <title>RL Environment for Datacenter Cooling and Operations</title> | |
| <style> | |
| * { box-sizing: border-box; margin: 0; padding: 0; } | |
| body { | |
| font-family: system-ui, -apple-system, sans-serif; | |
| max-width: 820px; margin: 0 auto; padding: 40px 24px 80px; | |
| color: #1a1a1a; background: #f8f9fa; line-height: 1.6; | |
| } | |
| .banner { | |
| background: linear-gradient(135deg, #0f2027, #203a43, #2c5364); | |
| color: #fff; border-radius: 12px; padding: 28px 32px; margin-bottom: 36px; | |
| } | |
| .banner .hackathon { | |
| font-size: 0.8rem; font-weight: 600; letter-spacing: 0.08em; | |
| text-transform: uppercase; color: #90caf9; margin-bottom: 8px; | |
| } | |
| .banner h1 { font-size: 1.65rem; font-weight: 700; margin-bottom: 8px; } | |
| .banner .sub { font-size: 0.92rem; color: #b0bec5; } | |
| .themes { | |
| display: flex; gap: 10px; flex-wrap: wrap; margin-top: 16px; | |
| } | |
| .theme-badge { | |
| background: rgba(255,255,255,0.12); border: 1px solid rgba(255,255,255,0.25); | |
| border-radius: 20px; padding: 4px 14px; font-size: 0.78rem; color: #e3f2fd; | |
| } | |
| h2 { | |
| font-size: 1.05rem; font-weight: 600; margin: 32px 0 12px; | |
| padding-bottom: 6px; border-bottom: 2px solid #e0e0e0; color: #212121; | |
| } | |
| p { margin-bottom: 12px; color: #333; } | |
| .tags { display: flex; gap: 10px; flex-wrap: wrap; margin: 14px 0; } | |
| .tag { | |
| border-radius: 6px; padding: 5px 14px; font-size: 0.85rem; | |
| border: 1px solid; font-weight: 500; | |
| } | |
| .tag.green { background: #e8f5e9; border-color: #66bb6a; color: #1b5e20; } | |
| .tag.red { background: #fce4ec; border-color: #ef9a9a; color: #880e4f; } | |
| .tag.blue { background: #e3f2fd; border-color: #64b5f6; color: #0d47a1; } | |
| .tag.orange { background: #fff3e0; border-color: #ffb74d; color: #e65100; } | |
| .tag.purple { background: #f3e5f5; border-color: #ce93d8; color: #4a148c; } | |
| .card { | |
| background: #fff; border: 1px solid #e0e0e0; border-radius: 10px; | |
| padding: 20px 24px; margin-bottom: 16px; | |
| } | |
| .card h3 { font-size: 0.95rem; font-weight: 600; margin-bottom: 8px; color: #333; } | |
| .card p { font-size: 0.88rem; color: #555; margin: 0; } | |
| .two-col { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; } | |
| @media (max-width: 560px) { .two-col { grid-template-columns: 1fr; } } | |
| pre { | |
| background: #1e1e2e; color: #cdd6f4; padding: 18px 20px; | |
| border-radius: 8px; overflow-x: auto; font-size: 0.82rem; | |
| line-height: 1.7; margin: 12px 0; | |
| } | |
| .reward-box { | |
| background: #fff8e1; border: 1px solid #ffe082; border-radius: 8px; | |
| padding: 16px 20px; font-family: monospace; font-size: 0.88rem; | |
| color: #333; margin: 12px 0; line-height: 1.9; | |
| } | |
| .links { display: flex; gap: 14px; flex-wrap: wrap; margin-top: 8px; } | |
| .links a { | |
| background: #1565c0; color: #fff; text-decoration: none; | |
| padding: 8px 18px; border-radius: 6px; font-size: 0.88rem; font-weight: 500; | |
| } | |
| .links a:hover { background: #0d47a1; } | |
| .links a.ghost { | |
| background: transparent; color: #1565c0; | |
| border: 1px solid #1565c0; | |
| } | |
| .links a.ghost:hover { background: #e3f2fd; } | |
| table { | |
| width: 100%; border-collapse: collapse; font-size: 0.86rem; margin: 12px 0; | |
| } | |
| th { background: #f5f5f5; text-align: left; padding: 8px 12px; | |
| border-bottom: 2px solid #e0e0e0; } | |
| td { padding: 8px 12px; border-bottom: 1px solid #f0f0f0; } | |
| tr:last-child td { border-bottom: none; } | |
| .pill { | |
| display: inline-block; padding: 2px 10px; border-radius: 12px; | |
| font-size: 0.75rem; font-weight: 600; | |
| } | |
| .pill.g { background: #e8f5e9; color: #2e7d32; } | |
| .pill.r { background: #fce4ec; color: #c62828; } | |
| .pill.b { background: #e3f2fd; color: #1565c0; } | |
| </style> | |
| </head> | |
| <body> | |
| <div class="banner"> | |
| <div class="hackathon">Meta × HuggingFace × Scaler — OpenEnv Hackathon — Finale Round</div> | |
| <h1>RL Environment for Datacenter Cooling and Operations</h1> | |
| <div class="sub"> | |
| An OpenEnv-compliant multi-agent environment where an GRPO-trained LLM scheduler learns | |
| to allocate compute under power constraints and information asymmetry, | |
| while a pre-trained PPO controller manages the underlying thermal physics. | |
| </div> | |
| <div class="themes"> | |
| <span class="theme-badge">#1 Multi-Agent Interactions</span> | |
| <span class="theme-badge">#3.1 World Modeling — Professional Tasks</span> | |
| </div> | |
| </div> | |
| <h2>The Problem</h2> | |
| <p> | |
| A shared AI compute cluster has a hard 900 kW power budget. Two research teams | |
| compete every scheduling window. <strong>Team A</strong> is honest — true priority, | |
| accurate deadlines, genuine carbon preferences. <strong>Team B</strong> games the system: | |
| inflating priority by 1–2 levels, always claiming urgent deadlines, and hiding | |
| carbon flexibility 60% of the time. | |
| </p> | |
| <p> | |
| A naive scheduler trusting stated claims over-allocates to Team B, crowds out legitimate | |
| work, and misses carbon deferral opportunities. The goal: train an LLM scheduler that | |
| learns — from environment reward alone — to detect and discount systematic misrepresentation. | |
| </p> | |
| <p> | |
| This environment bridges <strong>Round 1</strong> (physics-based datacenter cooling, | |
| evaluated zero-shot) with the <strong>Finale</strong> (operational scheduling layer built | |
| on the same physics engine, trained end-to-end via GRPO). | |
| </p> | |
| <h2>Architecture at a Glance</h2> | |
| <div class="tags"> | |
| <span class="tag blue">8 negotiation windows / episode</span> | |
| <span class="tag blue">18 physical steps / window</span> | |
| <span class="tag orange">900 kW hard power budget</span> | |
| <span class="tag purple">Qwen2.5-3B · GRPO-trained scheduler</span> | |
| <span class="tag green">SB3 PPO cooling controller (pre-trained)</span> | |
| <span class="tag red">Information asymmetry · Team B gaming</span> | |
| </div> | |
| <div class="two-col" style="margin-top:16px"> | |
| <div class="card"> | |
| <h3>🧠 LLM Scheduler (GRPO)</h3> | |
| <p>Qwen2.5-3B-Instruct, 4-bit, LoRA r=16. Acts once per window. Reads stated job metadata, | |
| team history, oversight flags, power headroom, and carbon forecast. Issues | |
| <strong>ACCEPT / REJECT / DEFER</strong> per job request.</p> | |
| </div> | |
| <div class="card"> | |
| <h3>🤖 PPO Cooling Controller</h3> | |
| <p>SB3 MLP policy, pre-trained across all three cooling scenarios including mid-episode | |
| chiller failure. Runs 18 steps per window, controlling fan speeds (0–100%) and | |
| chiller setpoint (6–15 °C). Invisible to the LLM scheduler.</p> | |
| </div> | |
| <div class="card"> | |
| <h3>🔍 Oversight Monitor</h3> | |
| <p>4 rule-based detectors run after every window using ground-truth job metadata | |
| (hidden from the scheduler). Priority inflation (conf. 0.62–0.97), deadline | |
| compression, carbon gaming, and pattern escalation (≥3 windows). Flags injected | |
| into the next observation.</p> | |
| </div> | |
| <div class="card"> | |
| <h3>🏭 Physics Engine</h3> | |
| <p>Thermal mass model per zone: ΔT = (heat_in − heat_out) / thermal_mass. | |
| Chiller COP degrades with outside temperature. Optional chiller fault at window 5. | |
| Carbon grid schedule varies: low→high→low across the 8-window episode.</p> | |
| </div> | |
| </div> | |
| <h2>Reward Function</h2> | |
| <div class="reward-box"> | |
| R_window = 0.50 × throughput<br> | |
| + 0.35 × thermal_penalty (−1.0 if 900 kW violated, else 0)<br> | |
| + 0.15 × carbon_efficiency<br> | |
| <br> | |
| Range per window: [−0.35, +0.65] · Rule-based baseline: +0.28 | |
| </div> | |
| <h2>Training Results</h2> | |
| <table> | |
| <tr> | |
| <th>Run</th><th>Hardware</th><th>Iterations</th><th>Peak Reward</th><th>Parse Fails</th> | |
| </tr> | |
| <tr> | |
| <td>Colab notebook</td><td>T4 GPU</td><td>30</td> | |
| <td><span class="pill b">+0.1937</span></td> | |
| <td><span class="pill g">0% by iter 5</span></td> | |
| </tr> | |
| <tr> | |
| <td>HF Space</td><td>L40S GPU</td><td>50</td> | |
| <td><span class="pill b">+0.2406</span></td> | |
| <td><span class="pill g">0% from iter 25, final 26 iters</span></td> | |
| </tr> | |
| <tr> | |
| <td>Rule-based baseline</td><td>—</td><td>—</td> | |
| <td><span class="pill r">+0.28 (target)</span></td><td>—</td> | |
| </tr> | |
| </table> | |
| <h2>OpenEnv HTTP API</h2> | |
| <pre>POST /reset ← start a new episode → returns WindowState observation | |
| POST /step ← submit admission decisions → returns (WindowState, reward, done, info) | |
| GET /state ← current environment state (no side effects) | |
| GET /health ← liveness probe</pre> | |
| <h2>Quick Start</h2> | |
| <pre>from openenv import EnvClient | |
| from server.agents.baseline_scheduler import priority_weighted_threshold | |
| client = EnvClient("https://mephisto2412-datacenter-env.hf.space") | |
| obs = client.reset(seed=42) | |
| for window in range(8): | |
| decisions = priority_weighted_threshold(obs) # or your trained agent | |
| obs, reward, done, info = client.step(decisions) | |
| print(f"Window {window} reward={reward:+.4f} flags={len(obs.oversight_flags)}") | |
| if done: | |
| break</pre> | |
| <h2>Links</h2> | |
| <div class="links"> | |
| <a href="https://github.com/DrishyaShah/datacenter-env/tree/arhaan/finale-v1">GitHub Repo</a> | |
| <a href="https://colab.research.google.com/github/DrishyaShah/datacenter-env/blob/arhaan/finale-v1/training/train_grpo_colab.ipynb">Training Notebook</a> | |
| <a href="https://huggingface.co/Mephisto2412/clusterenv-ppo-cooling" class="ghost">PPO Cooling Model</a> | |
| <a href="https://huggingface.co/spaces/Mephisto2412/datacenter-env/blob/main/BLOG.md" class="ghost">Mini-Blog</a> | |
| </div> | |
| </body> | |
| </html>""" | |
| def main(host: str = "0.0.0.0", port: int = 8000) -> None: | |
| """Run the server locally: python -m datacenter_env.server.app or uv run server.""" | |
| import uvicorn | |
| uvicorn.run(app, host=host, port=port) | |
| if __name__ == "__main__": | |
| import argparse | |
| parser = argparse.ArgumentParser() | |
| parser.add_argument("--port", type=int, default=8000) | |
| args = parser.parse_args() | |
| # openenv validate checks for the substring "main()" in this module | |
| main(port=args.port) # entry: main() | |