search

potamides / DeTikZify

Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ
Apache License 2.0
453 stars 11 forks source link

Sample fail on W11 #2

Open Malik-Hacini opened 3 months ago

Malik-Hacini commented 3 months ago

When running the tool from webui, it crashes after loading the shard.

Traceback :

File "C:\Users\juioi\Desktop\Detikzify.venv\Lib\site-packages\torch\nn\functional.py", line 2546, in layer_norm return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

potamides commented 3 months ago

Thanks for reporting this issue! Are you trying to run inference on CPU by any chance?

Malik-Hacini commented 3 months ago

I have never ran any model on my machine before, and haven't touched any settings.

potamides commented 3 months ago

Do you have a GPU on your machine? If yes, what is the output of import torch; print(torch.cuda.is_available())? The webui always tries to load models in half precision to save some memory but this is not supported for CPU inference. If you remove the following line it should work also on a cpu:

https://github.com/potamides/DeTikZify/blob/545f8494a073bf63df6cab5d3428d3707264fdc7/detikzify/webui/webui.py#L102

But I would expect this to be prohibitively slow. If it is usable, however, I am more than happy to implement CPU support.

Malik-Hacini commented 3 months ago

I do have an RTX 2050 on my machine (laptop). The output of cuda.is_available() is False. What should I do to get it to true ?

Malik-Hacini commented 3 months ago

Alright I fixed my pytorch w/ CUDA installation. I have another issue, should I keep going here or create antoher one ?

potamides commented 3 months ago

Feel free to create a new one. Be aware, however, that the VRAM of your RTX 2050 will likely not suffice to run the 7b models, for them you would need about 20gb.