Error: Peer-to-peer access is unsupported on this platform

[Youhe-Jiang]

Hi, the problem is:

(ParaWorker pid=1955026) Error: Peer-to-peer access is unsupported on this platform.
(ParaWorker pid=1955026) In the current version of distserve, it is necessary to use a platform that supports GPU P2P access.
(ParaWorker pid=1955026) Exiting...

I face a problem like this, but I actually checked the P2P connection between the two GPUs, and I tried the following codes for testing the P2P connection between GPUs:

tensor_a = torch.randn(10, device="cuda:0")
try:
    # Attempt to directly copy tensor_a from GPU 0 to GPU 1
    tensor_b = tensor_a.to("cuda:1")
    print("Successfully copied tensor from GPU 0 to GPU 1 using P2P.")
except RuntimeError as e:
    print("Failed to copy tensor from GPU 0 to GPU 1 using P2P. Error:", e)

and the output is:

Successfully copied tensor from GPU 0 to GPU 1 using P2P.

and the GPU topo is:

can you provide any suggestions?

thank you!

[Youhe-Jiang]

The GPUs I use are RTX 4090s

[Youhe-Jiang]

Yeah I solved the problem, seems that RTX 4090/PCIe connections cannot support DistServe, we run DistServe successfully on A100 machines with NVLINKs.