models icon indicating copy to clipboard operation
models copied to clipboard
verified publisher icon NVIDIA-Merlin

[QST] What is the proper way to load checkpoint to merlin.models.torch.DLRMModel

Open breadbread1984 opened this issue 1 year ago • 0 comments

❓ Questions & Help

DLRMModel cannot load trained checkpoint properly.

Details

the torch implement of DLRMModel has an interaction layer implement with a member defined in register_buffer in condition block (see https://github.com/NVIDIA-Merlin/models/blob/eb1e54196a64a70950b2a7e7744d2150e052d53e/merlin/models/torch/blocks/dlrm.py#L67). the newly created interaction has no such member until the forward is called. therefore, the checkpoint cannot be loaded properly. what is the recommended way of loading?

breadbread1984 avatar Nov 27 '24 01:11 breadbread1984