AMD gfx12 (RX 9070/9070 XT) offers a specialized MiniMax H3 Sage Attention node designed for memory-efficient processing in ComfyUI. This tool is compatible with KJNodes' existing patches, enabling users to seamlessly integrate it into their workflows without modifications.
- Provides a direct pass-through for sequences up to 30,000, ensuring no additional overhead on standard attention paths.
- Automatically implements head-chunking for sequences exceeding 30,000, significantly reducing peak memory usage by approximately 37%.
- Maintains compatibility with existing MiniMaxLowVRAMAttention configurations, allowing for flexible integration into various workflows.
Context
This tool serves as a dedicated node for AMD gfx12 architecture, specifically targeting the RX 9070 and RX 9070 XT graphics cards. Its primary purpose is to enhance memory efficiency while maintaining performance in the MiniMax H3 framework of ComfyUI.
Key Features & Benefits
The MiniMax H3 Sage Attention node optimizes memory usage during attention computations, particularly for longer sequences. By utilizing a head-chunking technique, it mitigates out-of-memory (OOM) errors, enabling users to run larger models on hardware with limited VRAM.
Advanced Functionalities
The node automatically engages head-chunking when processing sequences longer than 30,000 tokens, distributing the attention heads into manageable groups. This approach allows it to handle extensive data without crashing, a significant improvement over traditional methods that may lead to system failures.
Practical Benefits
Incorporating this tool into ComfyUI workflows enhances overall efficiency by allowing users to work with longer sequences without the fear of exceeding memory limits. It streamlines the processing of complex models, ultimately improving workflow control and output quality.
Credits/Acknowledgments
This project is a collaborative effort, drawing from the work of the SageAttention team at Tsinghua University and the KJNodes repository by Kijai. The tool is licensed under the MIT License, promoting open-source collaboration and development.
