# 算力集市

**Repository Path**: tianyeHome/compute_bazaar

## Basic Information

- **Project Name**: 算力集市
- **Description**: 做Ai时代的房东，让每一张显卡都不闲着
- **Primary Language**: Python
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-10-17
- **Last Updated**: 2026-10-24

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

<p align="center">
  <img src="https://img.shields.io/badge/Compute-Bazaar-58A6FF?style=for-the-badge&labelColor=0D1117" alt="Compute Bazaar">
  <br>
  <strong>Make Every GPU Count</strong>
</p>

<p align="center">
  <img src="https://img.shields.io/badge/Flutter-3.27+-02569B?logo=flutter&logoColor=white" alt="Flutter">
  <img src="https://img.shields.io/badge/Python-3.10+-3776AB?logo=python&logoColor=white" alt="Python">
  <img src="https://img.shields.io/badge/Dart-3.6+-0175C2?logo=dart&logoColor=white" alt="Dart">
  <img src="https://img.shields.io/badge/License-Apache--2.0-green" alt="License">
  <img src="https://img.shields.io/badge/Platform-Windows-0078D4?logo=windows&logoColor=white" alt="Platform">
</p>

<p align="center">
  <a href="https://computebazaar.netlify.app"><strong>🌐 Visit Website</strong></a>
  &nbsp;|&nbsp;
  <a href="https://computebazaar.netlify.app/privacy-policy"><strong>🛡️ Privacy Policy</strong></a>
  &nbsp;|&nbsp;
  <a href="https://www.bilibili.com/video/BV13enu6eEMG/"><strong>🎬 Demo Video</strong></a>
</p>

---

## Compute Bazaar

> The bulletin board of the AI era — securely and conveniently distribute idle GPU computing power to those who need it.

**Compute Bazaar** is a LAN computing power sharing platform based on Flutter and FastAPI. It is dedicated to securely and conveniently distributing idle GPU computing power within a LAN, serving as "the bulletin board of the AI era."

For detailed product design, see: [Product Design Whitepaper](docs/product_design.md)<br/>
**Official Website**: [https://computebazaar.netlify.app](https://computebazaar.netlify.app)

---

## Core Features

| Feature | Description |
|:-------:|-------------|
| 🚀 **Instant LAN Discovery** | UDP broadcast-based auto-discovery of idle compute nodes on the same subnet, zero-config plug-and-play |
| 🖥️ **Real Hardware Monitoring** | Real-time GPU utilization, VRAM, and temperature via pynvml, 2s refresh + fl_chart ECG-style scrolling line charts |
| 🛡️ **Docker Sandbox Isolation** | All tasks execute in Docker containers, whitelist validation + 60s timeout auto-kill, Windows calls Docker via WSL |
| 🔍 **Real Environment Detection** | Settings page real-time detection of NVIDIA driver / GPU / CUDA / Docker engine status, green/red visual indicators |
| 🔐 **Landlord Authorization** | Must agree to the "Compute Sharing Rental Agreement" when switching to landlord mode, real HTTP handshake to detect Agent, green security banner on success |
| 🎨 **Geek-style Dual-mode UI** | One-click switch between "Tenant" and "Landlord" identities, minimal dark mode, GitHub-style color scheme |

---

## Tech Architecture

| Layer | Technology |
|-------|------------|
| **Frontend** | Flutter + GetX + MVVM (Melos Monorepo) |
| **Backend Agent** | Python + FastAPI + pynvml |
| **Task Isolation** | Docker container (Windows via WSL, `--runtime=nvidia` GPU passthrough) |
| **Communication** | UDP broadcast (discovery) + HTTP REST API (task scheduling) |
| **Database** | Drift (SQLite) |
| **GPU Collection** | pynvml (NVIDIA official Python bindings) |
| **Charts** | fl_chart (ECG-style scrolling line charts) |

---

## Project Structure

```
compute_bazaar/
├── agent_backend/                  # Python FastAPI backend
│   ├── main.py                     # Entry: mount routes + UDP broadcast lifecycle
│   ├── requirements.txt
│   ├── routers/                    # GPU / Task / Env / Heartbeat / Discovery
│   └── services/                   # gpu_service / task_manager / task_executor / udp_broadcaster
│
├── apps/
│   └── compute_bazaar_local/       # Flutter Windows desktop app
│       ├── lib/pages/tenant/       # Tenant mode: Node Discovery / Task Submit / Usage View
│       ├── lib/pages/landlord/     # Landlord mode: GPU Monitor / Task Mgmt / Revenue Stats
│       └── pubspec.yaml
│
└── packages/                       # Dart shared packages (Melos Monorepo)
    ├── cb_models/                  # Data models (ComputeNode, GpuInfo, Task, AppConstants)
    ├── cb_core/                    # Controllers + Services (AppController, AgentClientService)
    ├── cb_ui/                      # Common UI (MainShell, SloganWatermark, Theme)
    ├── cb_discovery/               # UDP node discovery (RawDatagramSocket)
    ├── cb_data/                    # Drift database layer
    ├── cb_ad/                      # Ad placeholder
    └── cb_share/                   # Sharing
```

---

## Feature Showcase

### Tenant Mode (Use Computing Power)

<p align="center">
  <img src="docs/images/节点发现页.png" alt="Node Discovery" width="720">
</p>

| Page | Function |
|------|----------|
| **Node Discovery** | UDP listener auto-discovers LAN nodes, fetches real GPU info, displays IP / hostname / GPU / VRAM |
| **Task Submit** | Auto-selects first online node and first template, selected cards show bold blue border + green checkmark badge, dynamic bottom hints, highlighted submit button when ready |
| **Usage View** | Poll task list every 3s, clickable stat items for filtering + SegmentedButton status filter + search (ID/name/command) + paginated loading, show desensitized task name / status / duration / stdout / stderr, tap card to expand |

<p align="center">
  <img src="docs/images/任务提交执行结果页.png" alt="Task Result" width="720">
</p>

### Landlord Mode (Provide Computing Power)

<p align="center">
  <img src="docs/images/task_result_docker.png" alt="Docker Task Result" width="720">
</p>
<p align="center"><i>Docker sandbox isolation execution, real nvidia-smi output returned</i></p>
<p align="center">
  <img src="docs/images/landlord_gpu_monitor.png" alt="GPU Monitor" width="720">
</p>
<p align="center"><i>Real-time monitoring of GPU utilization, VRAM and temperature</i></p>

<p align="center">
  <img src="docs/images/landlord_settings_green.png" alt="Environment Detection All Green" width="720">
</p>
<p align="center"><i>One-click authorization, environment detection (NVIDIA driver / GPU / CUDA / Docker) all green ✅</i></p>

| Page | Function |
|------|----------|
| **GPU Monitor** | Request /api/gpu/metrics every 2s, real-time render utilization / VRAM / temperature + fl_chart ECG-style scrolling charts (30-point sliding window) |
| **Task Management** | Connect to /api/task/list, clickable stat items for filtering + SegmentedButton status filter + search + paginated loading, show desensitized task name / status / duration, support task termination |
| **Revenue Stats** | Calculate time slices / service duration / completed tasks based on real data, show recent service records |
| **Agent Settings** | Real-time detect Agent process / NVIDIA driver / GPU / CUDA / Docker engine status; blue "Start & Authorize Agent" button when offline, red "Stop & Deactivate Service" button when running |

### General Features

- **Online / Offline toggle**: Manually switch online status; submit button auto-disables when offline
- **Agent not started warning**: Red warning banner at top when Agent unreachable in landlord mode
- **Authorization status banner**: After authorization succeeds, a green banner "🛡️ Authorized for rental, sandbox isolation enabled" appears at the top
- **Rental authorization agreement**: First switch to landlord mode requires agreeing to the "Compute Sharing Rental Agreement" (no sleep / platform as info intermediary only / liability separation / sandbox security), followed by real HTTP handshake to detect Agent
- **Docker isolation badge**: Task submit page clearly marks 🛡️ Tasks run in Docker isolated sandbox
- **Log system**: In-app log viewer with search / level filter / auto-scroll / copy / export JSON
- **Agent process log capture**: Auto-launched Python backend stdout/stderr written to App log page in real-time, tagged [Agent], ERROR log + auto-restart on abnormal exit; uvicorn INFO/WARNING level stderr correctly identified as normal logs, not errors
- **Heartbeat anti-misjudgment**: 30s polling + /api/health priority probe + 3 consecutive failures (90s) before disconnect + proactive heartbeat confirm, avoids single network jitter misjudgment; heartbeat failure counting paused during Agent startup, no yellow/red warning before first ready
- **Agent standalone packaging**: PyInstaller build.spec packages as compute-bazaar-agent.exe, frontend prioritizes exe launch, falls back to python main.py; mode_switch_dialog home dialog "Agree & Start" auto-launches Agent

---

## Backend API

### GPU Monitoring

| Method | Path | Description |
|--------|------|-------------|
| GET | `/api/gpu/info` | Return GPU list (name, VRAM) |
| GET | `/api/gpu/metrics` | Return GPU utilization, VRAM usage, temperature |
| GET | `/api/gpu/stream` | SSE real-time GPU metrics stream |

### Task Management

| Method | Path | Description |
|--------|------|-------------|
| POST | `/api/task/submit` | Submit controlled task (Docker container isolation, 60s timeout) |
| GET | `/api/task/list` | List all tasks |
| GET | `/api/task/{id}/status` | Query single task status and output |
| POST | `/api/task/{id}/stop` | Force stop task |

**Allowed task types (whitelist)**: `system_check` (System Check Docker), `ai_inference` (AI Inference Sim)

**Docker execution commands**:
```bash
# Windows (WSL)
wsl docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all nvidia/cuda:11.0-base nvidia-smi

# Linux/Mac
docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all nvidia/cuda:11.0-base nvidia-smi
```

> 🛡️ All tasks execute in Docker isolated containers, never directly invoke host commands. 60-second timeout auto-kill. Task names are desensitized (e.g., "System Check (Docker)" instead of raw docker command).

### Other

| Method | Path | Description |
|--------|------|-------------|
| GET | `/api/env/check` | Environment detection (NVIDIA driver, GPU, CUDA, Docker engine) |
| GET | `/api/health` | Agent health check |
| GET | `/api/heartbeat/status` | Heartbeat status |
| POST | `/api/heartbeat/confirm` | Confirm heartbeat (prevent unconfirmed_count accumulation) |

---

## UDP Node Discovery Protocol

| Parameter | Value |
|-----------|-------|
| Broadcast port | 39878 |
| Broadcast interval | 3 seconds |
| Broadcast address | 255.255.255.255 |
| Timeout removal | Remove node after 45s without heartbeat |

---

## 📥 Download & Experience
- **Windows Client**: The app is currently under review on the Microsoft Store. Please visit [the official website](https://computebazaar.netlify.app) for installation instructions.
- **Build from Source**: See the "Quick Start" section below.

## Quick Start

### Requirements

| Dependency | Version | Purpose |
|------------|---------|---------|
| **Flutter** | 3.27+ / Dart 3.6+ | Frontend desktop app |
| **Python** | 3.10+ | Agent backend service |
| **Docker Desktop** | Latest stable | Task sandbox isolation (Windows via WSL2) |
| **NVIDIA Container Toolkit** | Latest stable | Docker GPU passthrough (`--runtime=nvidia`) |
| **NVIDIA GPU + driver + CUDA** | Any CUDA 11+ capable GPU | Landlord mode GPU collection & task execution |
| **Windows** | 10/11 | Current single-machine platform |

> 💡 **Tip**: If you only use tenant mode (submitting tasks to remote nodes), you don't need Docker or an NVIDIA GPU on your machine.

### 🚀 Detailed Startup & Verification Guide (with Docker/WSL Isolation)

To ensure LAN discovery and Docker sandbox isolation work properly, follow these steps in order. It's recommended to open **three terminal windows** to observe logs.

#### 1. Start Backend Agent Service (PowerShell Window A)

```powershell
# Activate Python virtual environment (modify name to match your setup)
conda activate ai_env

# Enter backend directory
cd agent_backend

# Install dependencies (first time or when dependencies change)
pip install -r requirements.txt

# Start FastAPI service
python main.py
```

On successful startup, you should see output like:

```
INFO:     Started server process [xxxxx]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:39877 (Press CTRL+C to quit)
[UDP] Broadcasting node info to 255.255.255.255:39878 ...
```

**Verify the backend is running**:

```powershell
# Health check
curl http://127.0.0.1:39877/api/health
# Expected: {"status":"ok","agent":"running"}

# Environment check (confirm GPU / Docker are recognized)
curl http://127.0.0.1:39877/api/env/check
# Expected: response with driver_version, gpu, cuda, docker fields

# GPU info
curl http://127.0.0.1:39877/api/gpu/info
# Expected: GPU name, total VRAM, etc.
```

#### 2. Verify Docker + WSL2 GPU Isolation (PowerShell Window B)

> ⚠️ This step is only needed on first deployment or when Docker/GPU configuration seems broken.

```powershell
# Confirm WSL2 is installed and Docker Desktop is running
wsl --list --verbose
# Expected: Docker Desktop's WSL2 distribution listed

# Verify Docker is available inside WSL2
wsl docker --version

# Verify NVIDIA Container Toolkit works inside WSL2
wsl docker run --rm --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=all nvidia/cuda:11.0-base nvidia-smi
# Expected: GPU info matching host nvidia-smi output
```

If `nvidia-smi` fails to detect GPU inside the container, check:
- Docker Desktop → Settings → Resources → WSL Integration is enabled
- NVIDIA Container Toolkit is properly installed (`nvidia-ctk --version`)
- Host `nvidia-smi` works correctly

#### 3. Start Frontend Flutter App (PowerShell Window C)

```powershell
# Enter frontend directory
cd apps/compute_bazaar_local

# Install dependencies (first time or when dependencies change)
flutter pub get

# Start Windows desktop app
flutter run -d windows
```

On successful startup, the Flutter window opens automatically, defaulting to **Tenant Mode**.

#### 4. End-to-End Verification

**Tenant Mode Verification**:
1. Open the app, default to tenant mode
2. Wait for LAN nodes to appear on "Node Discovery" page (local node auto-discovered when backend is running)
3. On "Task Submit" page, select target node → select task template (System Check Docker / AI Inference Sim) → submit
4. View task execution status and output on "Usage View" page
5. Switch to PowerShell Window A to observe backend logs, confirm Docker container startup and task execution records

**Landlord Mode Verification**:
1. Confirm Agent backend is running in Window A
2. Switch to "Landlord Mode" via top-right toggle
3. First-time switch requires agreeing to the "Compute Sharing Rental Agreement"; after clicking "Agree & Start", the frontend sends a real HTTP handshake to `http://127.0.0.1:39877/api/health`
4. Handshake success → authorization complete, green banner "🛡️ Authorized for rental, sandbox isolation enabled" appears at top
5. Handshake failure → dialog prompts "Local Agent service not detected, please ensure python main.py is running in terminal"
6. View real-time GPU status on "GPU Monitor" page (2s refresh, ECG-style charts)
7. View and manage running tasks on "Task Management" page
8. On "Agent Settings" page, confirm environment detection status (NVIDIA driver / GPU / CUDA / Docker all green ✅)
9. Click red "Stop & Deactivate Service" button to clear authorization

**Docker Isolation Verification**:
1. Submit a "System Check (Docker)" task in landlord mode
2. Switch to PowerShell Window A, observe `docker run` command in backend logs
3. In Window B, run `wsl docker ps` to confirm the container is running
4. After task completes (~10-30s), run `wsl docker ps` again to confirm container auto-cleanup (`--rm` mode)

#### 5. Troubleshooting

| Issue | Resolution |
|-------|------------|
| Backend startup `ModuleNotFoundError` | Confirm `pip install -r requirements.txt` was run, check Python virtual environment is activated |
| `/api/health` returns connection refused | Confirm backend is running and port 39877 is not occupied: `netstat -ano \| findstr 39877` |
| GPU info returns empty | Confirm host `nvidia-smi` works, pynvml dependency installed: `pip show pynvml` |
| Docker container fails to start | Confirm Docker Desktop is running, WSL2 integration enabled, NVIDIA Container Toolkit installed |
| `nvidia-smi` fails inside container | Check `--runtime=nvidia` is effective, try `wsl docker run --rm --runtime=nvidia ...` manually |
| No nodes on discovery page | Confirm UDP broadcast started (logs should show `[UDP] Broadcasting`), check firewall allows UDP 39878 |
| Flutter compilation error | Confirm Flutter 3.27+: `flutter --version`, run `flutter clean && flutter pub get` |
| Task persistence lost | Confirm `agent_backend/data/tasks.db` exists and is writable, backend auto-restores on restart |

---

## Contributing

1. Fork this repository
2. Create feature branch: `git checkout -b feature/your-feature`
3. Commit changes: `git commit -m 'Add some feature'`
4. Push branch: `git push origin feature/your-feature`
5. Submit Pull Request

### Development Standards

- Dart code follows [Effective Dart](https://dart.dev/guides/language/effective-dart) guidelines
- Python code follows PEP 8
- Commit messages follow Conventional Commits format

---

## License

This project is licensed under [Apache License 2.0](LICENSE).

---

<p align="center">
  Compute Bazaar — Make Every GPU Count
</p>