You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, we only test compatibility with models as part of benchmarking. In order to catch issues before they appear there, we should have a basic training test that passes in output samples from cuGraph-PyG into a model. This would be mostly a smoke test; the other tests should verify that the output of cuGraph-PyG is correct, regardless.
The text was updated successfully, but these errors were encountered:
Consolidates various speed improvements tested while running performance benchmarks. Avoids copying batch data, removes redundant data loading code, simplifies and improves de-offsetting, even though that is now being bypassed entirely for homogeneous graphs. Removes extra host to device copy. Properly flips the src/dst columns in the returned `HeteroData` minibatch objects, avoid exposing this to the end user.
I've confirmed this cuts the MFG time by a factor of 4.
Closes#3807
Authors:
- Alex Barghi (https://github.com/alexbarghi-nv)
Approvers:
- Vibhu Jawa (https://github.com/VibhuJawa)
- Don Acosta (https://github.com/acostadon)
- Brad Rees (https://github.com/BradReesWork)
URL: #3795
Currently, we only test compatibility with models as part of benchmarking. In order to catch issues before they appear there, we should have a basic training test that passes in output samples from cuGraph-PyG into a model. This would be mostly a smoke test; the other tests should verify that the output of cuGraph-PyG is correct, regardless.
The text was updated successfully, but these errors were encountered: