Using cached nvidia_cusolver_cu11-11.4.0.1-2-py3-none-manylinux1_x86_64.whl (102.6 MB) Collecting nvidia-cusparse-cu11==11.7.4.91 (from torch>=2.0.0->vllm) Using cached nvidia_cusparse_cu11-11.7.4.91-py3-none-manylinux1_x86_64.whl (173.2 MB) Collecting nvidia-nccl-cu11==2.14.3 ...
I guess we are using the system NCCL installation to be able to pip install nvidia-nccl-cu12 during the runtime. If we would use the third_party/nccl module I assume we would link NCCL into the PyTorch binaries. If so, we should make sure to update the install_cuda.sh NCCL version ...