Can I use torch2.3.0 + cuda 11.8 on Runpod?
I want upgrade my touch version to 2.3.0, can it works on Runpod ?
16 Replies
Hmm not sure, try it out. Why not use the current available template version of torch?
Torch 2.3.0 works with CUDA 11.8 but I suggest using CUDA 12.1 instead
Because xformers 0.0.26 (RECOMMENDED, linux & win) Install latest stable with pip: Requires PyTorch 2.3.0. I am testing now.
Traceback (most recent call last):
2024-06-05T09:04:11.388043478Z File "/runpod-volume/55fd91b5/prod/instantid/src/handler.py", line 14, in <module>
2024-06-05T09:04:11.388417474Z import diffusers
2024-06-05T09:04:11.388430521Z ModuleNotFoundError: No module named 'diffusers'.
I confirm that diffusers 0.27.0 is already installed. Why does this issue always occur in serverless?
It doesn't always occur, its specific to your installation. Maybe you installed it into a venv on a network volume or something.
I see
Check it with pip, and check if you get venv on when launching your app
ok , I will check.
Dependencies are listed in the requirements.txt file，When running runpodctl project dev, don't these dependencies get automatically downloaded and installed?
xformer is kinda meh choice right now as PyTorch have native split attention
Thanks， I will check to delete xfomer or not.
You can open .txt file and put # in front of xformer
example
#xformers
Ok, I will check code if or not need xfomers.
I mean you can always install it and just remove version of xformers so it would install latest
xformers latest is 0.0.26,it need torch 2.3.0, but now I catch problem when i use torch2.3.0.I am checking now.
Should I install requirements.txt in my prod env? or ,when I use 'runpodctl project deploy' , there dependencies will be auto installed on prod env ?
To be true havent got chance to use deploy yet
btw fireice does the
projects on runpod directly calls the handler.py without any other scripts (like .sh files)?no other scripts,. runpod.serverless.start({"handler": handler}) in handler.py
ohh