xFormers
A memory-efficient attention algorithm library developed by Meta AI. It avoids storing the full attention matrix by recomputing intermediate activation values, reducing peak GPU memory usage by 30–50%. The library is widely adopted for memory optimization in AI image generation workflows.
Core Facts
Timeline (last 90 days)
Enabling xFormers can typically reduce VRAM usage by approximately 30%–50% and improve generation speed by 15%–25%.
PyTorch 2.x可能需要CUDA 11.8或12.1，不同版本的xFormers对PyTorch版本有严格要求
整合包通过xFormers或--medvram参数等显存优化技术，利用注意力机制优化和分块计算降低峰值显存占用，使低配显卡也能运行
xFormers is a Transformer optimization library open-sourced by Meta; its Memory-Efficient Attention avoids storing the full attention matrix in VRAM through block-wise computation and can improve attention computation speed by 2–4×.
对于8GB显存用户，启用量化和注意力优化可节省约30-40%的显存占用
All Facts (6)
Informer reduces the self-attention computation complexity from O(n²) to O(n log n) through ProbSparse sparse attention.
70%UnverifiedxFormers is a Transformer optimization library open-sourced by Meta; its Memory-Efficient Attention avoids storing the full attention matrix in VRAM through block-wise computation and can improve attention computation speed by 2–4×.
60%UnverifiedEnabling xFormers can typically reduce VRAM usage by approximately 30%–50% and improve generation speed by 15%–25%.
50%UnverifiedPyTorch 2.x可能需要CUDA 11.8或12.1，不同版本的xFormers对PyTorch版本有严格要求
50%Unverified整合包通过xFormers或--medvram参数等显存优化技术，利用注意力机制优化和分块计算降低峰值显存占用，使低配显卡也能运行
50%Unverified对于8GB显存用户，启用量化和注意力优化可节省约30-40%的显存占用
50%