Since torchrec register a state_hook. The path of is always prepended a long prefix:
for example, if I have a module structure like
A.B.C.ec = EmbeddingCollection() # A is not DMP wrapped, instead, ec is DMP wrapped.
# then I'll get a state dict key
A.B.C.ec.A.B.C.ec.embeddings.weight
For some reason, I can use DMP for my whole model. Is there any thing I can do?
Since torchrec register a state_hook. The path of is always prepended a long prefix:
for example, if I have a module structure like
For some reason, I can use DMP for my whole model. Is there any thing I can do?