I have trained a model using
DistributedDataParallel. After training, I serialized the model like so where the model is wrapped using
Note that this serialization was performed in the launcher function which is typically passed to
torch.multiprocessing. My training setup consists of 4 GPUs.
Now when I am trying to load the checkpoint in my local inference setup (single GPU) the keys are not matching. The model, in this case, is not wrapped using
DistributedDataParallel. Any pointers would be useful.