Hi there,
Thank you so much for the repo. Since I met a nccl problem on my windows machine.
But kinda confused by you repo how to use these things?
Thanks again.

It would be great if we could build a pip package for this.
Not sure where to start though, as I don't know where things should go under "normal": circumstances.

Any suggestions?

I build your cuda 11.7 project and got nccl64_134.11.7.dll .
Should I put this in cuda's bin folder?
I am currently using pytorch.

Hello, I'm having an issue where the .exe files are not showing up in the directory after I build the project. I'm trying to build the project for CUDA version 10.1. I get no errors from VS, and it tells me the build is successful. It creates nccl64_134.10.1.dll and copies cudart64_101.dll correctly, it just isn't creating the .exe files for some reason. I thought they might be somewhere else, but they're nowhere to be found. Any ideas?

I can't understand how to build this nccl on windows, can you give more detailed steps for install?

I would be very grateful ;)

please if you could provide with steps that make me able using the library with tensorf flow

I had to comment out CheckDelta in the test files to make them work ... what are the implications of this? I am extremely new to C++ and building via Visual Studio (like, a few days of experience), so I do not have strong debugging skills! But please let me know your thoughts. Additionally, when I use task manager to track GPU activity, I only see my original GPU doing everything in terms of processing. The new, 2nd GPU is showing copying activity, and VRAM utilization increases, so I know something is working. I wonder if it is not doing any processing because of the commented-out code! I highly appreciate any help, I bought a 2nd 3090, a beefy PSU, an NVLink, and redid all my wiring for my Windows system without knowing that cross GPU communication is not straightforward on Windows! Your repo gives me hope though.

If there is anything I can provide, please let me know. There really isn't much of an error code, it just says:

(base) [User]\GitHub\NCCL\windows\test\single\x64\Debug>all_reduce_test 300000000
Using devices
Rank 0 uses device 0 [0x24] NVIDIA GeForce RTX 3090
Rank 1 uses device 1 [0x2d] NVIDIA GeForce RTX 3090

                                             out-of-place                    in-place
  bytes             N    type      op     time  algbw  busbw      res     time  algbw  busbw      res
 300000000     300000000    char     sumCuda failure [User]\GitHub\NCCL\test\include\test_utilities.h:313 'invalid argument'

Which I traced back to the CheckDelta calls in the test codes.

nccl.10.1.vcxproj -> D:\NCCL-master\windows\x64\Release\nccl64_134.10.1.dll copy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin\cudart*.dll" "D:\NCCL-master\windows\x64\Release\" 系统找不到指定的路径。

in fact the file 'cudart64_101.dll' exists in 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin'

do u know how to fit it? thanks!