A flash attention extension for stable diffusion webui with AMD ZLUDA (gfx11xx, Windows) environments.
一个为 AMD GPU Windows ZLUDA 环境提供Flash attention优化方案的stable diffusion webui扩展插件
Flash attention 编译自： Flash attention Compiled From: https://github.com/Repeerc/flash-attention-v2-RDNA3-minimal
1.从本仓库安装插件。 Install extensions from this repository.
2.选择 Flash Attention 算法，然后点击SET激活。 Select the Flash Attention algorithm and click SET to activate it.
GPU: AMD RX7900 XTX 24GB
CPU: Intel 13900T
SD1.5 Model:cuteyukimixAdorable_neochapter[6ee4f31532], VAE:vae-ft-mse-840000-ema-pruned
SDXL Model:animagineXLV3_v30[1449e5b0b9], VAE:None
Clip skip:2
DPM++, 2M, 1024x1024, 50 Steps, BatchSize = 1
| Pytorch SDP | Flash Attention | ||
|---|---|---|---|
| Speed | 2.75 it/s | 3.04 it/s | |
| VRAM (shown in Task manager) | 9.1 GB | 8.6 GB |
DPM++, 2M, 1024x1536, 50 Steps, BatchSize = 1
| Pytorch SDP | Flash Attention | ||
|---|---|---|---|
| Speed | 1.70 it/s | 1.93 it/s | |
| VRAM (shown in Task manager) | 10.95 GB | 9.1 GB |
Tile VAE: Encoder Tile Size = 1536, Decoder Tile Size = 64
DPM++, 2M, 512x512, 50 Steps, BatchSize = 4
| Pytorch SDP | Flash Attention | ||
|---|---|---|---|
| Speed | 4.53 it/s | 4.64 it/s | |
| VRAM (shown in Task manager) | 9.08 GB | 3.5 GB |
768x768, BatchSize = 2
| Pytorch SDP | Flash Attention | ||
|---|---|---|---|
| Speed | 2.92 it/s | 3.36 it/s | |
| VRAM (shown in Task manager) | 19.0 GB | 3.69 GB |
960x960, BatchSize = 1
| Pytorch SDP | Flash Attention | ||
|---|---|---|---|
| Speed | 2.91 it/s | 3.36 it/s | |
| VRAM (shown in Task manager) | 15.9 GB | 3.61 GB |