nvidia - How can I know if NCCL is installed? - Stack Overflow

admin管理员组

文章数量:1434908

Very simple question. I have access to a multi-node machine and I have to do some NCCL tests. In the readme it says

If CUDA is not installed in /usr/local/cuda, you may specify CUDA_HOME. Similarly, if NCCL is not installed in /usr, you may specify NCCL_HOME.

I can see that CUDA is installed but (and here it comes my question)

how can I know if NCCL is installed? and where?

---

Other info

I have done

find /usr -name "libnccl.so*" 2>/dev/null

and I found this file. However when I di

find /usr -name "nccl.h" 2>/dev/null

it was not found. Obviously I could not build even the simplest

#include <stdio.h>
#include <nccl.h>

int main() {
    printf("NCCL version: %d\n", NCCL_VERSION_CODE);
    return 0;
}

(Btw, I think the OS is CentOS)

Very simple question. I have access to a multi-node machine and I have to do some NCCL tests. In the readme it says

If CUDA is not installed in /usr/local/cuda, you may specify CUDA_HOME. Similarly, if NCCL is not installed in /usr, you may specify NCCL_HOME.

I can see that CUDA is installed but (and here it comes my question)

how can I know if NCCL is installed? and where?

---

Other info

I have done

find /usr -name "libnccl.so*" 2>/dev/null

and I found this file. However when I di

find /usr -name "nccl.h" 2>/dev/null

it was not found. Obviously I could not build even the simplest

#include <stdio.h>
#include <nccl.h>

int main() {
    printf("NCCL version: %d\n", NCCL_VERSION_CODE);
    return 0;
}

(Btw, I think the OS is CentOS)

• nvidia
• hpc
• nccl

Share Improve this question Follow asked Nov 17, 2024 at 13:17 KansaiRobotKansaiRobot 10.2k2222 gold badges112112 silver badges227227 bronze badges

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 0

It is likely you have the runtime:

sudo yum install -y libnccl

But not the development environment:

sudo yum install -y libnccl-devel

As an alternative, since you have the HPC tag, most HPC cluster tend to have their code under modules (env mod, or lmod) and those are usually outside /usr. You can look with

module avail nccl

If it is there you could load the module and should have access to the development environment.

For the actual finding, If it is in a module, the the previous command will tell, and you can check in the module file to see if any variable like nccl_home is set which might make it easier.

You can also use ldconfig, which might work (if it doesn't show, it could be a false negative as there are other reasons other than not being installed that could cause the negative), it prints all the shared libraries cached by the system.

ldconfig -p | grep libnccl

Finally, specific to this case, try to run nvidia-smi if it is installed (and in path), it should print an output indicating the version (and maybe location?) of nccl.

本文标签： nvidiaHow can I know if NCCL is installedStack Overflow