Add Automicrobatching for Non-Powers-of-2 + Fixes to FSDP deadlocks using Adaptive Sync Hooks #3503
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
mvpatel2000
reviewed
Jul 30, 2024
mvpatel2000
reviewed
Jul 31, 2024
Comment on lines
+134
to
+144
| @no_type_check | ||
| def unshard(self): | ||
| """ | ||
| Run the unshard logic. | ||
| This is an unpatched method from pytorch, meant to be reverted to | ||
| whenever automicrobatching turns off its hooks for increased throughput. | ||
| This includes all-gathering the flat parameter | ||
| and switching to using the unsharded flat parameter. If the handle does | ||
| not need unsharding, then this only switches to using the unsharded | ||
| flat parameter. For ``NO_SHARD``, this is a no-op. | ||
| If FSDP is in :meth:`summon_full_params` and the handle uses parameter |
There was a problem hiding this comment.
This should probably be in the if torch 2.3.1 section
Comment on lines
+567
to
+568
|
|
||
| if auto_microbatching: |
There was a problem hiding this comment.
can you add a comment on what this is doing?
Comment on lines
+326
to
+327
| def _double_device_train_microbatch_size(state: State): | ||
| """Double device_train_microbatch_size when automicrobatching searches upward for a higher non-OOM microbatch size. |
There was a problem hiding this comment.
should this go into automcirobatching utils folder?
| num_consecutive_thrashes = 0 | ||
| return num_consecutive_thrashes | ||
|
|
||
| def _handle_downward_search_in_automicrobatching(state: State, lowest_oom_microbatch_size: int, highest_non_oom_microbatch_size: int, lower_bound_microbatch_size: int, num_search_steps: int, max_search_steps: int): |
There was a problem hiding this comment.
same comment on moving to utils?
| @@ -1251,6 +1421,7 @@ def __init__( | |||
| if parallelism_config is not None: | |||
| # Patch PyTorch to fix distributed bugs | |||
| patch_pytorch() | |||
| patch_unshard_for_automicrobatching(self.auto_microbatch_size_found) | |||
There was a problem hiding this comment.
this should be just part of patch_pytorch to simplify interface
There was a problem hiding this comment.
we need to pass in a boolean variable telling it how to patch this one specific method though - i feel like it would be less readable if we passed self.auto_microbatch_size_found directly into patch_pytorch
Comment on lines
+2976
to
+2977
| # Sync for OOMs | ||
| found_cuda_oom = _found_ooms_across_ranks(self.state, found_cuda_oom) |
There was a problem hiding this comment.
this block is really complicated. lets move to a helper fn
| with torch.no_grad(), model_eval_mode(self.state.model): | ||
| if self.state.fsdp_enabled and self.first_batch_complete: | ||
| print("readd hooks for eval") |
| @@ -8,6 +8,18 @@ | |||
| convert_nested_dict_to_flat_dict, | |||
| extract_hparams, | |||
| ) | |||
| from composer.utils.automicrobatching import ( | |||
| # _create_sync_hook, | |||
| @@ -164,4 +176,14 @@ | |||
| 'validate_credentials', | |||
| 'build_remote_backend', | |||
| 'RemoteFilesExistingCheckStatus', | |||
| # '_create_sync_hook', | |||
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.
No description provided.