Problem to generate the data

[raphaelmourad]

Hi,
first, thank you for this great work!

I have a problem when I want to generate scgpt embedding with the corresponding notebook.

When I run:
adata_copy_embeddings = scg.tasks.embed_data(
adata_copy,
model_dir,
gene_col='index',
batch_size=1,
return_new_adata=True,
device='cpu',
#use_fast_transformer=False,
)

I get this:
"The kernel appears to have died. It will restart automatically."
scGPT - INFO - match 5235/5666 genes in vocabulary of size 60697.
/home/mourad/.local/lib/python3.10/site-packages/torch/nn/modules/transformer.py:282: UserWarning: enable_nested_tensor is True, but self.use_nested_tensor is False because encoder_layer was not TransformerEncoderLayer
warnings.warn(f"enable_nested_tensor is True, but self.use_nested_tensor is False because {why_not_sparsity_fast_path}")
Embedding cells: 0%| | 0/111445 [00:00<?, ?it/s]

Please could you help me? I have already preprocessed the data and save it.

[yanwu2014]

Thanks! And at first glance I think you've run out of CPU memory, can you describe the machine you're using and the dataset you're trying to embed?

[raphaelmourad]

Hi Yu, I have 64Gb of RAM on my CPU. I checked the system monitor of Ubuntu and I didn't see that all the memory was taken.

[yanwu2014]

Hmm so I'm not able to reproduce this error, can you also create an issue on the scGPT repo since it seems like an scGPT issue? https://github.com/bowang-lab/scGPT/issues

In the meantime I've found the scGPT install to be a bit tricky, this recipe worked for me if you want to give it a shot

conda create -n scgpt-env python=3.10 pip
conda activate scgpt-env

conda install conda-forge::cudatoolkit=11.7 conda-forge::cudatoolkit-dev=11.7 conda-forge::cudnn=8.9.7.29
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install packaging cellxgene-census GitPython transformers datasets tensorboard wandb rich scib ipython torchtext scvi-tools
pip install flash-attn'<'1.0.5 --no-build-isolation
pip install scgpt --no-deps