Enable sharded checkpoint save and load (support local, sharded, and full state dicts for FSDP) #1902
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
eracah
changed the title
bcui19
reviewed
Feb 6, 2023
bcui19
reviewed
Feb 6, 2023
dakinggg
reviewed
Feb 8, 2023
There was a problem hiding this comment.
Nice! Please add unit tests :) For manual test, could you also test a resumption run? i.e. i'd like to see that training for 100 batches is the same as training for 50 batches, checkpointing, and then resuming and training for another 50. Also would be good to check with all three types.
|
Ok, @dakinggg, I added unit tests and docs. Give it another look when you get the chance. |
Co-authored-by: Daniel King <[email protected]>
Co-authored-by: Daniel King <[email protected]>
Co-authored-by: Daniel King <[email protected]>
* check non-rank0 ranks only hold a shard for "full" state dict modified
* "local" state dicts have flattened shards
* "sharded" state dicts have unflattened shards
Co-authored-by: Daniel King <[email protected]>
|
@dakinggg, okay I added the tests you asked for. Can you give it one more quick once over and then I'll merge it? |
dakinggg
approved these changes
Feb 16, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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?
Enables configuring what type of state dict to user for models and optimizers in
StateWe enable the user to set fsdp_config['state_dict_type'] to configure these types of state dicts. The types are:
'full'(akatorch.distributed.fsdp.StateDictType.FULL_STATE_DICT) - the full unflattened, unsharded state dict materialized only on rank 0 with cpu offloading'local'(akatorch.distributed.fsdp.StateDictType.LOCAL_STATE_DICT- materializes just the local flattened shard of the state_dict on each rank'sharded'(akatorch.distributed.fsdp.StateDictType.SHARDED_STATE_DICT) - returns the unflattened, sharded state_dict.when a user specifies
Trainer(model=my_model, optimiziers=my_optimizer, fsdp_config={..., 'state_dict_type': 'local'}, ...), then they enable local sharding for their model parameters and their optimizer states.In order to enable this functionality this PR:
State.state_dict()for model parameters and optim parameters to support the three FSDP state dict typesState.load_state_dict()for model and optim that supports state dict typessave_checkpointto allow non-zero ranks to save checkpointsload_checkpointto allow non-zero ranks to load checkpointsTodo:
Tests
did a bunch of manual tests
Also did manual tests to make sure proper error raised if:
What issue(s) does this change relate to?
fix CO-1433
fix CO-1682