search

awslabs / graphstorm

Enterprise graph machine learning framework for billion-scale graphs for ML scientists and data scientists.
Apache License 2.0
377 stars 60 forks source link

[Optimization-wholegraph] Remove CUDA dependency during conversion to WholeGraph format #547

Open isratnisa opened 1 year ago

isratnisa commented 1 year ago

Node features are transformed from DGL format to WholeGraph format using the convert_feat_to_wholegraph.py script. However, this script significantly consumes memory, prompting us to explore utilizing our substantial CPU instances like r5 or X1. Unfortunately, due to its reliance on WholeGraph's CUDA dependency (pylibwholegraph.binding.wholememory_binding), leveraging these instances is currently not possible. Are there alternative approaches to address this limitation?

classicsong commented 1 year ago

We need support from NVidia.

classicsong commented 3 months ago

@chang-l Do we have this support?