GPUCachedFeature does not currently support updating the feature with different dimensions when initialized. Our inferencing loop uses the update function. Since layers have differing hidden dimensions, we get the error below.
We need to update GPUCachedFeature so that it reconstructs the GPUCache again when the feature dimension is changed.
This bug prevents us from using GPUCachedFeature in our single GPU examples where we have the inferencing loop.
This could be a release blocker. Now that we are aware of the issue, it would be really nice if next immediate release fixed it. I moved my place today so I will work on the fix tomorrow.
🔨Work Item
IMPORTANT:
Project tracker: https://github.com/orgs/dmlc/projects/2
Description
https://github.com/dmlc/dgl/blob/67a897fa56353f766fc84303ff99b14c79c6a896/examples/sampling/graphbolt/node_classification.py#L224
GPUCachedFeature does not currently support updating the feature with different dimensions when initialized. Our inferencing loop uses the update function. Since layers have differing hidden dimensions, we get the error below.
We need to update GPUCachedFeature so that it reconstructs the GPUCache again when the feature dimension is changed.
This bug prevents us from using GPUCachedFeature in our single GPU examples where we have the inferencing loop.