.cache/torch_extensions/py38_cu117/utils/utils.so about visorgpt HOT 5 CLOSED

Hi, I will check the problem soon. It seems to be the incompatible problem of torch and deepspeed. Maybe you can try to train the model using the environment below:

torch=1.7.1
deepspeed==0.7

from visorgpt.

When I run pip3 install -r requirements.txt, I met the following error

ERROR: Could not install packages due to an OSError: [Errno 2] 没有那个文件或目录: '/opt/conda/conda-bld/flit-core_1644941570762/work/source/flit_core'
https://github.com/Sierkinhane/VisorGPT/blob/6f1252f9859f62707fe31faf4913b9e4efad05b3/requirements.txt#L36

It seems that the requirements.txt needs rearrangements.

from visorgpt.

We don't have A100. We use 3090 for training and inference. We can run the gradio inference but fail to run training.

from visorgpt.

This problem has been solved by compiling deepspeed from source instead of using pip install deepspeed
microsoft/DeepSpeed#2187.

• Uninstall DeepSpeed pip uninstall deepspeed
• clone repo git clone https://github.com/microsoft/DeepSpeed.git && cd DeepSpeed
• install DS_BUILD_UTILS=1 pip install .

Thanks a lot.

from visorgpt.

Great!

from visorgpt.