Error in loading model: CUDA_ERROR_UNSUPPORTED_PTX_VERSION

[mohammadian-reza]

Hi, I am running the code examples on Google Colab, but when I load the model with this code:
model = load_model("deepblast-v3.ckpt", "Rostlab/prot_t5_xl_uniref50", device='cuda')

I get this error:
[222] Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION
ptxas application ptx input, line 9; fatal: Unsupported version 8.5; current version is '8.4'

I tried to change the CUDA version, but it didn't work. I also tried to install VM exactly as mentioned in the installation part, but it didn't work

[mortonjt]

Thanks, does the cpu version work instead?

model = load_model("deepblast-v3.ckpt", "Rostlab/prot_t5_xl_uniref50", device='cpu')

You could also try reinstalling PyTorch to match the google colab coda version. You can see the output of nvcc --version to get the version. If it uses say cuda=11.8, you can run something like

# Example for CUDA 11.8 !pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 torchaudio==2.0.2+cu118 -f https://download.pytorch.org/whl/torch_stable.html

Give it a shot and let us know if it works and we can update the installation accordingly

[mohammadian-reza]

I tried it, but it shows this error. (I tried both gpu and cpu, none of them worked)

ValueError Traceback (most recent call last)
in <cell line: 0>()
2 y = 'QQNKELNFKLREKQNEIFELKKIAETLRSKLEKYVDITKKLEDQNLNLQIKISDLEKKLSDA'
3 # obtains alignment string specifying structural superposition
----> 4 pred_alignment = model.align(x, y)

11 frames
/usr/local/lib/python3.11/dist-packages/numba/core/typing/typeof.py in typeof(val, purpose)
35 msg = _termcolor.errmsg(
36 f"Cannot determine Numba type of {type(val)}")
---> 37 raise ValueError(msg)
38 return ty
39

ValueError: Cannot determine Numba type of <class 'torch.Tensor'>

[mohammadian-reza]

Is there any solution to this error? I installed the stated environment for running deepblast and tmvec, I don't know why I face this error.

[mortonjt]

Hi @mohammadian-reza thanks for raising this issue. We currently don't have the bandwidth to maintain google colab.

We are able to run TMvec locally, so it should just be a matter of getting the right cuda libraries installed. The most up to date fork is under : https://github.com/valentynbez/tmvec , see the install instructions there and let us know if they work for you.