D-Ogi/ComfyUI-Attention-Optimizer

Implements automatic benchmarking of multiple attention backends (PyTorch SDPA, Flash Attention, SageAttention, xFormers) within ComfyUI, detecting optimal implementations per GPU and model configuration, then applies the fastest via model patching. Caches benchmark results by model hash to eliminate repeated overhead, with support for diverse architectures (SDXL, Flux, video models) and head dimensions. Provides granular backend selection with per-GPU recommendations and detailed performance reporting.