[Bug] There is not any InfiniBand or NVLink in my 4-GPU machine, how can I use mscclpp to communicate?

[Maphsge4]

IndexError: IB transport out of range: 0 >= 0

Can I avoid using IB to communicate? Thanks for answering my question!

[chenhongyu2048]

For reference: use cudaipc

mscclpp::Transport transport = mscclpp::Transport::CudaIpc;

[chhwang]

For reference: use cudaipc

mscclpp::Transport transport = mscclpp::Transport::CudaIpc;

@Maphsge4 Did this help you? If you are working on PCIe, you should also make sure that your GPUs are peer-to-peer accessible, otherwise CudaIpc transport won't work as well.

[Lorenz5622]

@chhwang where should I add this code
mscclpp::Transport transport = mscclpp::Transport::CudaIpc;
, into a .cuda file or a python file?
I am trying to run splitwise (https://github.com/aashaka/vllm-oss/tree/splitwise-pr) , it will be great help if you could give me a solution in detail. Thanks!