In the Llama recipes repository, there is a function called freeze_transformer_layers. This function can be triggered when sharding the model via FSDP.
However, the nightly (or main) FSDP docs, at time of writing, states:
FSDP has some constraints on freezing parameters (i.e. setting
param.requires_grad=False). Foruse_orig_params=False, each FSDP instance must manage parameters that are all frozen or all non-frozen. Foruse_orig_params=True, FSDP supports mixing frozen and non-frozen, but we recommend not doing so since then the gradient memory usage will be higher than expected (namely, equivalent to not freezing those parameters). This means that ideally, frozen parameters should be isolated into their ownnn.Modules and wrapped separately with FSDP.
Why are we able to freeze / unfreeze certain parameters in this way? Should we be wrapping the decoder blocks in FSDP separately?