How to run GraphGym with ROCm ?

[OscarRoussel]

Hi,

I followed the installation steps for PyG ROCm but I'm struggling to find info on how to now run graph gym next.

Thank you.

[Looong01]

Just use my pre-build wheels and install them.
https://github.com/Looong01/pyg-rocm-build/releases

Then:

git clone https://github.com/snap-stanford/GraphGym
cd GraphGym
pip install -r requirements.txt
pip install -e .  # From latest verion
pip install graphgym # (Optional) From pypi stable version

It should work

[OscarRoussel]

Hi, sorry for the delay.

I used the prebuilt wheels you provided.

Here is additional information I should have provided in my first message :

rocm version : 6.3.2-66
pytorch : 2.6.0 for rocm 6.2.4
pytorch geometric : provided pyg-rocm-build whl

installed the other wheels provided.

Followed the installation procedure of GraphGym (same as what you described).

Here is the issue :

When running bash run_single_pyg.sh from GraphGym/run

I get the error :

Traceback (most recent call last):
  File "/home/oscar/Documents/GraphGym/run/main_pyg.py", line 35, in <module>
    auto_select_device()
  File "/home/oscar/Documents/GraphGym/graphgym/utils/device.py", line 56, in auto_select_device
    memory_raw = get_gpu_memory_map()
  File "/home/oscar/Documents/GraphGym/graphgym/utils/device.py", line 13, in get_gpu_memory_map
    result = subprocess.check_output([
  File "/usr/lib/python3.10/subprocess.py", line 421, in check_output
    return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
  File "/usr/lib/python3.10/subprocess.py", line 503, in run
    with Popen(*popenargs, **kwargs) as process:
  File "/usr/lib/python3.10/subprocess.py", line 971, in __init__
    self._execute_child(args, executable, preexec_fn, close_fds,
  File "/usr/lib/python3.10/subprocess.py", line 1863, in _execute_child
    raise child_exception_type(errno_num, err_msg, err_filename)
FileNotFoundError: [Errno 2] No such file or directory: 'nvidia-smi'

Which is caused by the get_gpu_memory_map() function of GraphGym to attempt a call at nvidia-smi.

Used some LLM to overload the function in question. Now seems to correctly use the GPU (still have some issues on my end to solve).

However a specialized PR of GraphGym to consider a rocm use of the function seems useless as the repository is not maintained.

[Looong01]

Hi, sorry for the delay.

I used the prebuilt wheels you provided.

Here is additional information I should have provided in my first message :

rocm version : 6.3.2-66 pytorch : 2.6.0 for rocm 6.2.4 pytorch geometric : provided pyg-rocm-build whl

installed the other wheels provided.

Followed the installation procedure of GraphGym (same as what you described).

Here is the issue :

When running from GraphGym/runbash run_single_pyg.sh

I get the error :

Traceback (most recent call last):
  File "/home/oscar/Documents/GraphGym/run/main_pyg.py", line 35, in <module>
    auto_select_device()
  File "/home/oscar/Documents/GraphGym/graphgym/utils/device.py", line 56, in auto_select_device
    memory_raw = get_gpu_memory_map()
  File "/home/oscar/Documents/GraphGym/graphgym/utils/device.py", line 13, in get_gpu_memory_map
    result = subprocess.check_output([
  File "/usr/lib/python3.10/subprocess.py", line 421, in check_output
    return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
  File "/usr/lib/python3.10/subprocess.py", line 503, in run
    with Popen(*popenargs, **kwargs) as process:
  File "/usr/lib/python3.10/subprocess.py", line 971, in __init__
    self._execute_child(args, executable, preexec_fn, close_fds,
  File "/usr/lib/python3.10/subprocess.py", line 1863, in _execute_child
    raise child_exception_type(errno_num, err_msg, err_filename)
FileNotFoundError: [Errno 2] No such file or directory: 'nvidia-smi'

Which is caused by the function of GraphGym to attempt a call at nvidia-smi.get_gpu_memory_map()

Used some LLM to overload the function in question. Now seems to correctly use the GPU (still have some issues on my end to solve).

However a specialized PR of GraphGym to consider a rocm use of the function seems useless as the repository is not maintained.

It is the problem of GraphGym, not torch_geometric's. The fundamental functions of GraphGym should work, including use AMD GPUs, cuz there is no CUDA codes inside GraphGym source codes. But its function like monitoring GPU's info may not work, cuz it should use rocm-smi or amd-smi. I think it is easy to change it if anyone want to add AMD GPU support for GraphGym, just change this line of code "GraphGym/graphgym/utils/device.py, line 13, in get_gpu_memory_map" only and it should work.