v1.2.3 - SageAttention Support

[Enemyx-net]

⚡ SageAttention Integration

• New Attention Type: Added "sage" option to attention_type parameter
• Quantized Attention: Uses INT8/FP8 precision for faster inference
• Performance Boost: Speedup in generation with minimal quality impact

🚀 Performance Improvements

Speed Comparison

• SageAttention: Up to 2-3x faster than standard attention
• Memory Efficient: Reduced memory bandwidth usage
• Quality Maintained: Negligible quality difference in output

✨ How It Works

SageAttention uses quantized operations (INT8/FP8) instead of full precision:

• ✅ Faster Matrix Operations: Quantized math is significantly faster
• ✅ Lower Memory Bandwidth: Smaller data types reduce memory transfer
• ✅ Intelligent Quantization: Preserves important information while reducing precision
• ✅ Hardware Acceleration: Leverages modern GPU tensor cores

⚙️ Requirements

For SageAttention:

• GPU: NVIDIA GPU with CUDA support
• Library: Install sageattention
• CUDA: Compatible CUDA version installed

🎛️ Usage

Select "sage" from the attention_type dropdown:
auto: Let transformers choose (default)
eager: Standard implementation
sdpa: Scaled Dot Product Attention
flash_attention_2: Flash Attention 2
sage: SageAttention (NEW) - quantized for speed

💾 Installation

Install via ComfyUI Manager or manually:
git clone https://github.com/Enemyx-net/VibeVoice-ComfyUI

📋 Notes

SageAttention is only available for NVIDIA GPUs with CUDA
Falls back to standard attention if requirements aren't met
Ideal for production environments prioritizing speed
Compatible with all model variants including 4-bit quantized models

---

This discussion was created from the release v1.2.3 - SageAttention Support.