FSDP sample fails with CUDA initialization error on HyperPod EKS

[shimomut]

When running the FSDP sample app on HyperPod EKS cluster, I got this error.

[W CUDAFunctions.cpp:108] Warning: CUDA initialization: Unexpected error from cudaGetDeviceCount(). Did you run some cuda functions before calling NumCudaDevices() that might have already set an error? Error 803: system has unsupported display driver / cuda driver combination (function operator())
Traceback (most recent call last):
  File "/fsdp/train.py", line 281, in <module>
    main(args)
  File "/fsdp/train.py", line 144, in main
    dist.init_process_group()
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/c10d_logger.py", line 86, in wrapper
    func_return = func(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/distributed_c10d.py", line 1184, in init_process_group
    default_pg, _ = _new_process_group_helper(
  File "/usr/local/lib/python3.10/dist-packages/torch/distributed/distributed_c10d.py", line 1339, in _new_process_group_helper
    backend_class = ProcessGroupNCCL(
ValueError: ProcessGroupNCCL is only supported with GPUs, no GPUs found!

Cluster spec:

• ml.g5.8xlarge x 8
• ml.g5.12xlarge x 2

[sean-smith]

looks like it's missing the nvidia gpu device plugin or you didn't specify the number of devices in your yaml like so:

resources:
        limits:
          nvidia.com/gpu: 1 # requesting 1 GPU

[iankouls-aws]

Reproduced the error with 2 x g5.8xlarge nodes. I believe the root cause is:
"system has unsupported display driver / cuda driver combination"

[mhuguesaws]

Related to #475

[mhuguesaws]

@shimomut please add link to test case. There are multiple FSDP test cases in the repo.

[shimomut]

This is the FSDP sample I faced the issue:
https://github.com/aws-samples/awsome-distributed-training/blob/main/3.test_cases/10.FSDP/README-EKS.md

Updated the original post as well.

[mhuguesaws]

Removing the Cuda compat package will likely resolve this.