PyTorch with CUDA 11.8

To install PyTorch 2.2.0 with CUDA 11.8 support:

pip install torch==2.2.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

PyTorch with CUDA 12.4

For CUDA 12.4 compatibility, use:

pip install torch==2.4.0+cu124 --extra-index-url https://download.pytorch.org/whl/cu124

LMdeploy Minimum Requirements

LMdeploy requires at least PyTorch 2.2.2 with CUDA 11.8:

pip install torch==2.2.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

vllm 0.4.0 Compatibility

vllm 0.4.0 with CUDA 11.8 requires PyTorch 2.1.2:

pip install torch==2.1.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

Installation Process Details

During installation, you may observe dependencies being resolved:

nvidia-cuda-nvrtc-cu11==11.8.89
nvidia-cuda-runtime-cu11==11.8.89
nvidia-cuda-cupti-cu11==11.8.87
nvidia-cudnn-cu11==8.7.0.84

These CUDA-specific package are automatically downloaded from PyTorch's official repository.

Important Consideration

When using requirements files, note that standard pip install -r requirements.txt commands cannot automatical resolve CUDA-specific package variants. The CUDA version must be explicitly specified in each PyTorch installasion command.