updating paperspace pytorch without starting from the scratch

[CoqueTornado]

Hello, I have installed a lot of Loras and extensions thus for that I would like to know how to upgrade the torch version. I know, I know, this looks like a dumb question. But I have already tried with this sentence pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

and it upgrades everything but... when starting the magical code of thelastben it looks like everything gets back to the torch 1.13 or something like that I have installed since march if do you know what I mean.

I think it is related with the Sytem Packages of the code, so I should start another notebook from the scratch, wiping out this one I built with all the extensions, loras, textual inversions and so on.

If anybody knows a solution please do not hesitate to let me know.

2 question (this is less important). Can we manage to use an external storage drive in paperspace for the models, the loras or the outputs? I mean not a temporaly one. Maybe the output could be... just please let me know if you managed this issue too.

Thank you in advance!

[CoqueTornado]

ok now it shows "free(): invalid pointer" in the Start Stable-Diffusion box... I think I have messed up everything... I will get back to torch 1.12 but... maybe is the docker, so when I instaled this version it works only with torch 1.12 and I have to get a new empty notebook and install the newest one probably. Please light us the path...

thank you!

[CoqueTornado]

1.13.1+cu116 is what it does when installing from the scratch... the docker goes like that, it looks like there is not option to upgrade on paperspace thelastben?

[TheLastBen]

Paperspace is working on having containers that hold the latest torch and CUDA versions, I can't promise the date, but I know they are working on it.

[CoqueTornado]

thus for this, the hot potatoe is in the roof of Paperspace I bet. So there is no option to raise the Pytorch to 2.0?

I have tried like a crazy 5 or 6 options but it crashes the "play" asking for dependences or strange errors like the free invalid pointer I told you before.... and then suddenly it goes back to the PyTorch 1.13.1+cu116

So as long as it is done over a docker of Paperspace we can not do anything to solve this.

I just wanted the SDP backend or the Reactor...

maybe it is time to move on towards cagliostro or buying a new laptop hahahah....

[pdannell]

As a note, I think I noticed that you can't run SDV1.5 Animated Diff V3 either until Pytorch is updated.

Figured I would throw this here before people post new issues on it, related to the following error:

• AttributeError: module 'torch.nn.functional' has no attribute 'scaled_dot_product_attention'

[SRagy]

Here's a workaround.

1. Run the notebook as usual.
2. Open a terminal from the icon on the left menu bar.
3. Paste this and press enter: pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 xformers==0.0.23post1
4. Restart the kernel and rerun the notebook. The updated packages will now persist.

[pdannell]

Here's a workaround. 1. Run the notebook as usual. 2. Open a terminal from the icon on the left menu bar. 3. Paste this and press enter: pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 xformers==0.0.23post1 4. Restart the kernel and rerun the notebook. The updated packages will now persist.

This worked for getting mine up to date without breaking everything.