gloo

gloo copied to clipboard

No way to use a ReductionFunction in CUDAAllreduce?

2

I noticed that the constructor does not take a reduction function, and there's no way to set it. https://github.com/facebookincubator/gloo/blob/master/gloo/cuda_allreduce_halving_doubling.h#L75 Is this intended?

ebetica

Doesn't work at the moment...

pietern

This would verify that we copy all the headers we need to copy in the install step.

pietern

Otherwise REALLY weird errors pop up (see for example https://github.com/pytorch/pytorch/issues/2835)

pietern

reference benchmarking in InifiniBand/RoCE

4

Hi! Firstly, thanks for the nice work. It's good to see the brief benchmark figures in README.md. It would be great if anybody can show the benchmarking result of `--transport...

changtimwu

Do this instead of assuming hostname(2) is resolvable. This is typically not the case on people's custom Ubuntu installs and whatnot. We keep the API the same but just change...

pietern

Two computers, one is ubantu and the other is win11. Pytorch 2.3.0 is used for distributed training model. Since win11 does not support nccl mode, gloo is used, but the...

soundmemories

Hi, Can Gloo use libfabric ? I see it has ibverbs to be used as transport? why not libfabric to allow all types of transport?

niyazmurshed

[vcpkg] Fix the order of structures to adapt to the linu kernel

4

Hi, all. I am the maintainer of vcpkg. Recently we received a build error regarding gloo. https://github.com/microsoft/vcpkg/issues/38852 **Reproduce**: ``` git clone https://github.com/microsoft/vcpkg cd vcpkg/ ./vcpkg install gloo:x64-linux ``` **Error**: ```...

FrankXie05

cstdint for uint8_t need to be included explicitly when compiling with GCC 15

atupone