Running test training fails on Macbook Pro (intel)

[Ishitori]

Hi team,

I have created a tiny graph and I want to run graphstorm locally on my Macbook Pro (intel processor) laptop to do link prediction. I don't have NVidia GPU, and I install dependencies using default pip commans (pyhon 3.11):

pip install graphstorm
pip install dgl

I use PyTorch as a backend (default, didn't change it. PyTorch version 2.0.1)

I have created DistDGL graph using provided tool, and trying to run training script. I opened ssh on my laptop and allowed to do password-less login, and it can login.

The problem starts when it tries to create a server via gsgnn_lp.py. It calls rpc.reset() at dgl/distributed/dist_context.py, line 266 and fails with NameError: name '_CAPI_DGLRPCReset' is not defined.

I think this suggests that I am missing a dependency, but I am not sure.
Could someone help me track it down, please?

Thank you.

[classicsong]

You can try to run GraphStorm in Standalone mode:

graphstorm/tests/end2end-tests/graphstorm-ec/test.sh

Lines 30 to 31 in 74c99ec

	echo "**************standalone"
	python3 $GS_HOME/python/graphstorm/run/gsgnn_ep/gsgnn_ep.py --part-config /data/movielen_100k_ec_1p_4t/movie-lens-100k.json --cf $GS_HOME/training_scripts/gsgnn_ep/ml_ec.yaml

[Ishitori]

Well, if I use this line on main branch, I get ImportError: cannot import name 'use_wholegraph' from 'graphstorm.utils' (/usr/local/lib/python3.11/site-packages/graphstorm/utils.py). But I don't think it is related...

I am also doing Link prediction, not classification or regression, so I guess I need to use gsgnn_lp.py, right? If so, I don't see how can I run it in standalone mode - it tries to initialize distributed context almost immedeately...

[classicsong]

Can you use the latest code?
WholeGraph support was recently added.

[Ishitori]

Thanks, I have updated the code and it looks like the problem is gone. Thank you!