NCCL version upgrade for PyTorch

[YingleiZhang]

🐛 Bug

Build PyTorch from source code would fail on some old Linux release.
This issue has been described here: NVIDIA/nccl#244, and Nvidia folks had it fixed (NVIDIA/nccl@7f2b337). However, that change has not been included in the PyTorch source code, so if people need to build PyTorch from source code on some old linux system, it would still fail for the same problem.

The current nccl version in the latest PyTorch release is still 2.4.8, i think it can be upgraded to 2.5.6 now.

I can see that there is another open ticket for the same issue: #29093

To Reproduce

Steps to reproduce the behavior:

1. Checkout PyTorch source code.
2. Build on a linux system with kernel that is below 3.9 (This is when SO_REUSEPORT was introduced).

Expected behavior

Build should succeed on Linux system that is below 3.9.

 - PyTorch Version (e.g., 1.0): 1.4 master
 - OS (e.g., Linux): Linux (Below 3.9)
 - How you installed PyTorch (`conda`, `pip`, source): source
 - Build command you used (if compiling from source): python setup.py develop
 - Python version: 3.6
 - CUDA/cuDNN version: irrelevant
 - GPU models and configuration: irrelevant
 - Any other relevant information: No

Additional context

No

[ngimel]

Thank you for reporting! We should update nccl submodule. cc @seemethere to route to the appropriate person.

[mattip]

This was fixed in gh-41608 which upgraded NCCL to 2.7.3

[zarzen]

are there ways to choose a different version of NCCL at runtime?
I have a modified NCCL, I would like to use with PyTorch distributed API, how could I do?