PyTorch/torchdata Version Compatibility Issue with CUDA 11.8

[ssiddhantsharma]

Complex dependency conflicts between PyTorch, torchdata, and DGL when using CUDA 11.8:

• PyTorch 2.2.0+cu118 required for CUDA 11.8
• torchdata 0.6.1 requires PyTorch 2.0.1 exactly
• torchdata 0.7.0+ missing torchdata.datapipes module expected by DGL
• Installing compatible torchdata downgrades PyTorch to 2.1.0
• Creates conflicts with torchaudio 2.2.0+cu118 and torchvision 0.17.0+cu118

Additional Context

• Original conda script specified CUDA 12.1 packages, causing mixed CUDA version conflicts
• Similar issue reported in [RFdiffusion #306](ModuleNotFoundError: No module named 'torchdata.datapipes' RosettaCommons/RFdiffusion#306)

Solution

Downgrade to compatible versions:

pip install torch==2.2.1 torchdata==0.7.1
pip install "numpy<2"

Alternative for newer torchdata:

pip install torchdata==0.9.0  # if torchdata 0.10.0+ causes issues

[schloegl]

I can add, that I tried this on python 3.11 with venv, with pytorch cuda 12.1 using

    pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 --index-url https://download.pytorch.org/whl/cu121
    pip install -r CA_RFdiffusion.requirements.txt

The requirements file is attached.
CA_RFdiffusion.requirements.txt