release v1.1.0

🚀 Release v1.1.0 - CosyVoice3 Support & Performance Enhancements

We're excited to announce a major update to LightTTS with significant improvements in model support, performance, and infrastructure!

✨ What's New

🎤 CosyVoice3 Model Support

• Added full support for CosyVoice3: LightTTS now supports the latest CosyVoice3 architecture alongside CosyVoice2
• Tested with Fun-CosyVoice3-0.5B-2512 model in float16 precision
• Seamless model switching and compatibility with existing APIs

⚡ Performance Improvements

• Faster inference speed: Optimized inference pipeline delivering better throughput and lower latency
• Accelerated startup time: Significantly reduced model loading and service initialization time
• Enhanced streaming performance across all GPU configurations

Benchmark Highlights (RTX 4090D, float16)

• Non-streaming mode: Up to 4.10 QPS with 8 workers (RTF: 0.36)
• Streaming mode: TTFT 50% 0.33s with 1 worker (RTF: 0.21)

🔧 Infrastructure Updates

• Updated CUDA & PyTorch: Upgraded to PyTorch 2.9.1 with CUDA 12.8 for better performance and compatibility
• Improved dependency management and stability
• Enhanced Docker image with optimized runtime environment

📊 Performance Benchmarks

Full benchmarks available in the README for both NVIDIA RTX 4090D and RTX 5090 configurations, covering streaming and non-streaming modes.

Test configuration:

• Model: Fun-CosyVoice3-0.5B-2512
• Datatype: float16
• Test scripts: test/test_zs_speed.py (non-stream), test/test_zs_stream.py (stream)

🐛 Bug Fixes & Improvements

• Various stability improvements
• Enhanced error handling and logging
• Better memory management

📦 Installation

Pull the latest Docker image:
docker pull lighttts/light-tts:latest Or build from source