Problem installing scGPT

[imerelli]

Hi, I'm trying to install scGPT on a cluster with cudatoolkit=11.6

I'm able to install the right pytorch and flash-attn, with:

mamba create -n scgpt python==3.10 r-base=4.4 
pip install torch==1.13.0+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
pip install "flash-attn<1.0.5"

But then with:

pip install scgpt
...
Collecting torch>=1.13.0 (from scgpt)
  Using cached torch-2.8.0-cp310-cp310-manylinux_2_28_x86_64.whl.metadata (30 kB)
...

It automatically update torch to 2.8, which breaks the environment.

How is it possible to avoid this?

[loversaber]

I 've encountered the same issue

[rivzzzzz]

same issue here, and that leads to torchtext error:

     17 # the following import has to happen first in order to load the torchtext C++ library
---> 18 from torchtext import _extension  # noqa: F401

[yilitai]

I 've encountered the same issue