search

alembics / disco-diffusion

Other
7.47k stars 1.12k forks source link

is it possible to use multi GPUs to speed up the inference? #115

Open youchangxin opened 2 years ago

MoYuFly commented 2 years ago

Same question, request reply

aletts commented 2 years ago

Currently, Disco Diffusion doesn't support this. In theory there are ways to do this.. but it hasn't been implemented. Its history of being primarily used via Google Colab (and by relatively few users running locally, often with a single consumer Nvidia GPU) made implementation of it low priority.

MoYuFly commented 2 years ago

“Allow 3D animation to run on GPU's other than GPU0”，Can I use multiple GPUs, or can I specify other GPUs？

aletts commented 2 years ago

It's easy to make it run on a different GPU, an it's something that I do sometimes. You can search for the "cuda:0" string in disco.py and change it to your desired GPU index. Then you can e.g. run several instances of Disco Diffusion on the same system to use all the system's GPUs in parallel (working on separate jobs).

MoYuFly commented 2 years ago

It's inconvenient to access foreign Internet. I hope to realize this function as soon as possible. Thank you very much！

youchangxin commented 2 years ago

It's easy to make it run on a different GPU, an it's something that I do sometimes. You can search for the "cuda:0" string in disco.py and change it to your desired GPU index. Then you can e.g. run several instances of Disco Diffusion on the same system to use all the system's GPUs in parallel (working on separate jobs).

Thank you for your suggestion, but it may not fit my job well. I am struggling with multiprocess to speed up the diffusion step.

MoYuFly commented 2 years ago

I hope to send your multi process tutorial based on this project. Thank you very much！