How to check the version of NCCL

Question

I remotely access High-performance computing nodes. I am not sure about NVIDIA Collective Communications Library (NCCL) is installed in my directory or not. Is there any way to check whether the NCCL is installed or not?

Amorphous

did you try nvcc --version? – Stithy

you can do python -c "import torch;print(torch.cuda.nccl.version())" with pytorch. I wish I new the terminal command without pytorch. – Stithy

i tried print(torch.cuda.nccl.version()), it throws me this error: AttributeError: module 'torch._C' has no attribute '_nccl_version' – Measly

Answer 1

You can try

locate nccl| grep "libnccl.so" | tail -n1 | sed -r 's/^.*\.so\.//'

or if you use PyTorch:

python -c "import torch;print(torch.cuda.nccl.version())"

Check it this link Command Cheatsheet: Checking Versions of Installed Software / Libraries / Tools for Deep Learning on Ubuntu

For containers, where no locate is available sometimes, one might replace it with ldconfig -v:

ldconfig -v | grep "libnccl.so" | tail -n1 | sed -r 's/^.*\.so\.//'

Droll

When I enter `locate nccl| grep "libnccl.so" | tail -n1 | sed -r 's/^.*\.so\.//'1, it show nothing. – Amorphous

torch/lib/libtorch_cuda.so: undefined symbol: ncclCommRegister – Pale

Answer 2

You can usually do this in the command line:

nvcc --version

you might have to run:

sudo apt install nvidia-cuda-toolkit

too.

---

As the other answerer mentioned, you can do:

torch.cuda.nccl.version()

in pytorch. Copy paste this into your terminal:

python -c "import torch;print(torch.cuda.nccl.version())"

I am sure there is something like that in tensorflow.

Stithy

NVCC is a general CUDA C++ compiler. It doesn't report NCCL (communications library) version. The first part of the answer is wrong. – Gillmore

i tried print(torch.cuda.nccl.version()), it throws me this error: AttributeError: module 'torch._C' has no attribute '_nccl_version' – Measly