Models With Tied Weights Need Re-Tieing After FSDP Param Init #3154
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Currently in
FullyShardedDataParallelPlugin
, theparam_init_fn
is set when sync_module_states=True. This is required byFSDP
to initialize the shards (i.e. rank > 0) params in a variety of situations, including the important one where the parameter was ontorch.device("meta")
becauselow_cpu_mem_mode
was used.However
FullyShardedDataParallelPlugin.param_init_fn
is now set towhich causes problems when there are tied weights. Consider the following scenario
A
andB
that share a tied weightA.weight = B.weight
A.to_empty()
is called, thenA.weight
will be reassigned to a new tensor.B.empty()
is called, and then nowA.weight != B.weight
.This is observed to cause problems in the
low_cpu_mem_mode=True
case, because now when getting themanaged_params
in FSDP, see here,param_init_fn
is not called, becauselow_cpu_mem_mode
will load weights in this shard. Somanaged_params
will not have duplicates if weights are tied.param_init_fn
is called andmanaged_params
may have duplicates.Then when FSDP calls
_sync_module_params_and_buffers
, thetorch.distributed._broadcast_coalesced
will be trying to communicate different number of tensors. This is not as intended and causes unexpected behaviors.Fixes inconsistency in current logic in general, in particular is required to fix huggingface/trl#2089 to completion.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@muellerzr