search

Stability-AI / stablediffusion

High-Resolution Image Synthesis with Latent Diffusion Models
MIT License
38.83k stars 5.01k forks source link

Add mps-cpu support #163

Open Tps-F opened 1 year ago

Tps-F commented 1 year ago

Summary

・Adapted stablediffusion to work on M1 mac(mps). ・ Created an automatic device detection system for #147

Feature

・Support mps for txt2img.py, img2img.py, gradio and streamlit ・Add mac config(v2-inference-v-mac.yaml)

Of course, it can also run on cuda

oscarbg commented 1 year ago

Hi, nice work.. Can share perf on M1 CPU and GPU like https://github.com/Stability-AI/stablediffusion/pull/147#issuecomment-1402153477 for recent Intel IPEX pull request?

Tps-F commented 1 year ago
This is the result in my environment for txt2img.py! CPU Memory
Apple M1 16GB
Batch size Model name device argument Avg gen time How fast compared to cpu
1 SD2-v (768px) CPU --precision full 273s N/A
1 SD2-v (768px) MPS 100s 273%
4 SD2-v (768px) CPU --precision full 1010s N/A
4 SD2-v (768px) MPS 577s 175%
1 SD2-base (512px) CPU --precision full 271s N/A
1 SD2-base (512px) MPS 98s 276%
4 SD2-base (512px) CPU --precision full 1014s N/A
4 SD2-base (512px) MPS 600s 169%
oscarbg commented 1 year ago

Thanks for testing! Hope can be merged soon!

Tps-F commented 1 year ago

If there is a problem, I'll revert 90d4c71

jflatow commented 1 year ago

Can confirm this works more or less on M2 w/ 16GB (although I'm not sure how you generated 768x768, that crashes for me in allocating a buf)