Stable Diffusion (DALLE-2 clone) on AMD GPU
Mean Gene Hacks
Aug 23, 2022 23:53
Create an account to hide videos, mark them watched, or save them to Don't Miss.
Create an account to hide videos, mark them watched, or save them to Don't Miss.
Create an account to hide videos, mark them watched, or save them to Don't Miss.
Chapters
Get Stable Diffusion running on your AMD GPU without needing CUDA.
Note: Tested on Radeon RX68XX and 69XX series GPU's with Ubuntu 20.04/22.04 and ArchLinux. User reports indicate this method is working on several other GPU's including Radeon VII, Vega 56/64, 6700(XT), 6600(XT) [with env workaround], and even 5700(XT) series (see below for additional step).
Should theoretically work on RX470/480/570/580 GPU's with at least 8GB of ram too. You can check whether your AMD GPU supports ROCm here: https://community.amd.com/t5/knowledge-base/amd-rocm-hardware-and-software-support-document/ta-p/489937
ENV Workaround for 66XX(XT) and some 67XX(XT) GPU's:
https://github.com/RadeonOpenCompute/ROCm/issues/1756#issuecomment-1160386571
Update: This method also works on the 56XX/57XX(XT) series, but requires an additional step:
https://www.reddit.com/r/StableDiffusion/comments/ww436j/comment/imp7bx2/
RX 470/480/570/580/590 Users: 8GB RAM required. The bundled PyTorch has dropped support for these older cards, you will have to follow my video instructions, but use a different pytorch binary that includes support linked here: https://www.reddit.com/r/StableDiffusion/comments/ww436j/comment/ipkfm35/
STEP 1: Prepare ROCm Kernel Drivers
a) ArchLinux - AUR - See Video
b) Ubuntu - See video - https://github.com/RadeonOpenCompute/ROCm-docker/blob/master/quick-start.md
STEP 2: Download Docker Image & Setup Container
https://hub.docker.com/r/rocm/pytorch
STEP 3: Clone GIT repo
https://github.com/CompVis/stable-diffusion
STEP 4: Download Public Model Weights (v1.4)
https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
Step 5: Setup Conda Environment
https://pytorch.org/
#stablediffusion #stable-diffusion #amd
Education