Hy Mt

English | 简体中文 | 繁體中文 | 日本語

HY-MT Translation Service

🚀 All-in-One Docker deployment for Tencent HunyuanMT 1.5 translation model with Web UI, REST API, and MCP Server support.

✨ Features

• 🌐 38 Languages Support - Chinese, English, Japanese, Korean, French, German, Spanish, and 31 more
• 🎨 Modern Web UI - Dark/Light theme toggle, drag & drop file upload, real-time progress display
• ⚡ Streaming Translation - Server-Sent Events (SSE) for real-time output, perfect for long texts
• 🔧 Full Parameter Control - Temperature, Top-P, Top-K, repetition penalty adjustable
• 📚 Terminology Intervention - Custom term mapping for domain-specific translations
• 🤖 MCP Server - Model Context Protocol support for AI assistants (Claude, etc.)
• 🐳 One-Click Deployment - All-in-One Docker image with all models pre-downloaded
• 🔄 Smart GPU Management - Auto GPU selection, idle timeout, memory release
• 🔀 Multi-Model Support - Switch between 4 models (1.8B/7B, base/FP8) via UI or API

🎯 Model Selection Guide

Model	VRAM	Speed	Quality	Recommendation
HY-MT 7B	16GB	⭐⭐⭐⭐	⭐⭐⭐⭐⭐	🏆 Best Choice - Highest quality, fast speed
HY-MT 1.8B	6GB	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	Good for limited VRAM
HY-MT 1.8B FP8	4GB	⭐⭐⭐	⭐⭐⭐⭐	For VRAM < 6GB
HY-MT 7B FP8	10GB	⭐⭐	⭐⭐⭐⭐⭐	7B quality with less VRAM

💡 Tip: If you have 16GB+ VRAM, use HY-MT 7B for best results. FP8 models save memory but are slower due to runtime decompression.

📸 Screenshot

🚀 Quick Start

Docker Run (Recommended)

# One command to start (uses 7B model by default)
docker run -d --gpus all \
  -p 8021:8021 \
  -v ./models:/app/models \
  --name hy-mt \
  neosun/hy-mt:latest

# Access Web UI
open http://localhost:8021

The Docker image (~43GB) includes all 4 models pre-downloaded. No external downloads needed!

Docker Compose

Create docker-compose.yml:

services:
  hy-mt:
    image: neosun/hy-mt:latest
    container_name: hy-mt
    ports:
      - "8021:8021"
    environment:
      - MODEL_NAME=tencent/HY-MT1.5-7B  # Recommended for 16GB+ VRAM
      - GPU_IDLE_TIMEOUT=300
      - HF_ENDPOINT=https://huggingface.co  # Use https://hf-mirror.com for China
    volumes:
      - ./models:/app/models
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

docker compose up -d

📋 Requirements

Requirement	Minimum	Recommended
GPU	NVIDIA GPU with 6GB+ VRAM	16GB+ VRAM (for 7B model)
CUDA	11.8+	12.4+
Docker	20.10+	24.0+
nvidia-docker	Required	-

Verify GPU Support

# Check NVIDIA driver
nvidia-smi

# Check Docker GPU support
docker run --rm --gpus all nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi

📊 Performance Benchmark

Tested on NVIDIA L40S GPU, translating English to Chinese:

Model	Short (61 chars)	Medium (530 chars)	Long (1.8K chars)	Extra Long (4.2K chars)
HY-MT 7B	0.4s	4.4s	17.7s	43.0s
HY-MT 1.8B	0.4s	3.6s	14.0s	32.3s
HY-MT 1.8B FP8	1.1s	10.8s	38.1s	92.9s
HY-MT 7B FP8	2.9s	28.5s	115.6s	274.1s

⚠️ Why are FP8 models slower?

This is counter-intuitive but technically correct:

Comparison	Speed Change	Reason
1.8B FP8 vs 1.8B	2.7x slower	Runtime decompression overhead
7B FP8 vs 7B	6.4x slower	More parameters = more decompression

FP8 quantization is designed to save VRAM, not to speed up inference. The model is stored in 8-bit format but needs to be decompressed to 16-bit for GPU computation at runtime. This decompression happens for every token generation.

When to use FP8:

• ✅ When VRAM is limited (< 16GB for 7B, < 6GB for 1.8B)
• ❌ Not for speed optimization
• ❌ Not for batch processing (speed loss accumulates)

See Benchmark Report for detailed analysis.

🔑 Key Optimization: Chunk Size

Critical finding: Smaller chunk size = Better translation quality

Chunk Size	Quality	Notes
500 chars	❌ Poor	Mixed languages in output
300 chars	⚠️ Fair	Some untranslated residue
150 chars	✅ Excellent	Complete, accurate translation

The service uses MAX_CHUNK_LENGTH=150 by default for optimal quality.

Why? HY-MT model tends to "slack off" on long inputs, only translating part of the content. Shorter chunks force the model to fully translate each segment.

See Optimization Guide for details.

⚙️ Configuration

Environment Variables

Variable	Default	Description
PORT	8021	Service port
MODEL_NAME	tencent/HY-MT1.5-7B	HuggingFace model name
MODEL_PATH	./models	Local model cache path
GPU_IDLE_TIMEOUT	300	Auto-release GPU after idle (seconds)
NVIDIA_VISIBLE_DEVICES	auto	GPU ID (empty = auto select)
HF_ENDPOINT	https://huggingface.co	HuggingFace mirror URL

Using .env File

# Copy example config
cp .env.example .env

# Edit as needed
vim .env

📖 API Usage

Basic Translation

curl -X POST "http://localhost:8021/api/translate" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello, how are you?",
    "target_lang": "zh"
  }'

Response:

{
  "status": "success",
  "result": "你好，你好吗？",
  "elapsed_ms": 358,
  "model": "tencent/HY-MT1.5-7B",
  "chunks": 1
}

Streaming Translation (SSE)

curl -N "http://localhost:8021/api/translate" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Long article to translate...",
    "target_lang": "en",
    "stream": true
  }'

With Terminology Intervention

curl -X POST "http://localhost:8021/api/translate" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Apple released iPhone 16",
    "target_lang": "zh",
    "terms": {"Apple": "苹果公司", "iPhone": "苹果手机"}
  }'

Output: 苹果公司发布了苹果手机16

File Upload Translation

curl "http://localhost:8021/api/translate/file" \
  -F "[email protected]" \
  -F "target_lang=zh" \
  -F "stream=true"

Switch Model

curl -X POST "http://localhost:8021/api/models/switch" \
  -H "Content-Type: application/json" \
  -d '{"model": "tencent/HY-MT1.5-1.8B"}'

📚 API Endpoints

Endpoint	Method	Description
/	GET	Web UI
/api/translate	POST	Translate text (supports streaming)
/api/translate/file	POST	Upload and translate file
/api/translate/batch	POST	Batch translation
/api/translate/stream	POST	Streaming translation (SSE)
/api/languages	GET	List supported languages
/api/models	GET	List available models
/api/models/switch	POST	Switch translation model
/api/gpu/status	GET	GPU status and memory info
/api/gpu/offload	POST	Release GPU memory
/api/config	GET	Service configuration
/health	GET	Health check
/docs	GET	Swagger API documentation

🌍 Supported Languages

Language	Code	Language	Code	Language	Code
Chinese	zh	English	en	Japanese	ja
Korean	ko	French	fr	German	de
Spanish	es	Portuguese	pt	Russian	ru
Arabic	ar	Thai	th	Vietnamese	vi
Italian	it	Dutch	nl	Polish	pl
Turkish	tr	Indonesian	id	Malay	ms
Hindi	hi	Traditional Chinese	zh-Hant	Cantonese	yue

And 17 more languages. See /api/languages for full list.

🛠️ Tech Stack

• Model: Tencent HY-MT1.5 (1.8B & 7B)
• Backend: FastAPI + Uvicorn
• Frontend: Vanilla JS with Dark/Light Mode
• Container: NVIDIA CUDA 12.4 base image
• Streaming: Server-Sent Events (SSE)
• MCP: Model Context Protocol for AI integration

📁 Project Structure

hy-mt/
├── app_fastapi.py      # Main FastAPI application
├── mcp_server.py       # MCP Server for AI assistants
├── benchmark.py        # Performance benchmark script
├── templates/
│   └── index.html      # Web UI (Dark/Light theme)
├── docs/
│   ├── BENCHMARK_REPORT.md    # Performance test report
│   ├── OPTIMIZATION_GUIDE.md  # Long text optimization guide
│   └── QUICK_REFERENCE.md     # API quick reference
├── Dockerfile          # All-in-One Docker build
├── docker-compose.yml  # Docker Compose config
├── start.sh           # Quick start script
├── test_api.sh        # API test script
└── .env.example       # Environment config template

🔧 Advanced Usage

Manual Start (Development)

# Clone repository
git clone https://github.com/neosun100/hy-mt.git
cd hy-mt

# Install dependencies
pip install torch transformers accelerate fastapi uvicorn

# Run
python -m uvicorn app_fastapi:app --host 0.0.0.0 --port 8021

MCP Server Integration

For AI assistants like Claude Desktop, add to MCP config:

{
  "mcpServers": {
    "hy-mt": {
      "command": "python",
      "args": ["/path/to/hy-mt/mcp_server.py"],
      "env": {
        "HY_MT_API": "http://localhost:8021"
      }
    }
  }
}

Available MCP tools:

• translate - Translate text
• list_languages - Get supported languages
• list_models - Get available models
• switch_model - Switch translation model

See MCP_GUIDE.md for details.

🐛 Troubleshooting

Issue	Solution
Model download slow	Set HF_ENDPOINT=https://hf-mirror.com (China mirror)
GPU out of memory	Use quantized model: tencent/HY-MT1.5-1.8B-FP8
Container won't start	Check nvidia-smi and nvidia-docker installation
Translation incomplete	Already optimized with chunk size 150
Container shows unhealthy	Wait 1-2 minutes for model loading

📝 Changelog

v2.0.1 (2026-01-03)

• 🏆 Default model changed to HY-MT 7B (best quality & speed)
• 🩺 Added Docker HEALTHCHECK for container health monitoring
• 📦 Container status now shows (healthy) when ready

v2.0.0 (2026-01-03) - True All-in-One

• 🎯 All 4 models pre-downloaded in Docker image - No external downloads needed!
• 📦 Image size: ~43GB (includes all models)
• 🏆 Recommended: HY-MT 7B for best quality and speed
• 📊 Added performance benchmark report
• 🔧 Added benchmark.py for reproducible testing

v1.2.0 (2026-01-03)

• 🔀 Multi-model support (4 models: 1.8B, 1.8B-FP8, 7B, 7B-FP8)
• 🔄 Model switching via UI and API
• 📝 MCP Server: added list_models and switch_model tools
• 🐛 Fixed model name display in translation response

v1.0.0 (2026-01-03)

• 🎉 Initial release
• ✨ All-in-One Docker image
• ⚡ Streaming translation with SSE
• 🎨 Dark/Light theme Web UI
• 🔧 Long text optimization (chunk size 150)
• 🤖 MCP Server support

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

📄 License

This project is based on Tencent HunyuanMT. See License.txt for details.

🙏 Acknowledgments

• Tencent Hunyuan - Original HY-MT model
• HuggingFace - Model hosting

---

⭐ Star History

Star History Chart

📱 Follow Us