fully_shard load state_dict#90945
Closed
rohan-varma wants to merge 9 commits intogh/rohan-varma/627/basefrom
Closed
fully_shard load state_dict#90945rohan-varma wants to merge 9 commits intogh/rohan-varma/627/basefrom
rohan-varma wants to merge 9 commits intogh/rohan-varma/627/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/90945
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 298681f: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
rohan-varma
commented
Dec 15, 2022
| [test_name_mapping[str(s)] if s is not None else "none" for s in args] | ||
| ) | ||
|
|
||
| def _broadcast_state_dict(rank, state_dict): |
Contributor
Author
There was a problem hiding this comment.
not strictly needed right now but will be used in composable tests.
[ghstack-poisoned]
[ghstack-poisoned]
awgu
approved these changes
Dec 16, 2022
Collaborator
awgu
left a comment
There was a problem hiding this comment.
LGTM! There a few to-dos you left for yourself. Feel free to address those before landing.
| @skip_if_lt_x_gpu(2) | ||
| def test_state_dict_save_load_flow(self): | ||
| """ | ||
| E2E test of save + load with rank0_only + CPU offload for TransformerWithSharedParams |
Collaborator
There was a problem hiding this comment.
In the future, will this test include different state dict types and subtest the different configs?
| buffers, buffer_dtypes, fsdp_state.compute_device | ||
| if buffers: | ||
| mixed_precision_enabled_for_buffers = ( | ||
| fsdp_state._mixed_precision_enabled_for_buffers() if not _is_composable(fsdp_state) |
Collaborator
There was a problem hiding this comment.
To-do: We can make _mixed_precision_enabled_for_buffers() not be a method of FullyShardedDataParallel to make this not have to if / else here. We would be able to just check fsdp_state.mxied_precision.buffer_dtype is not None -- flexible whether that is in its own function or written like that every time.
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. [ghstack-poisoned]
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. [ghstack-poisoned]
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. [ghstack-poisoned]
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. [ghstack-poisoned]
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. [ghstack-poisoned]
awgu
reviewed
Dec 19, 2022
| param.zero_() | ||
| if zero_buffers: | ||
| for buffer in model.buffers(): | ||
| ctx = FSDP.summon_full_params(model) if summon_full else suppress() |
Collaborator
There was a problem hiding this comment.
Should we include any to-do or issue for following up on this? Or, could you remind me what the current status on this is?
Contributor
Author
There was a problem hiding this comment.
yeah, let me file an issue summarizing it.
Contributor
Author
|
@pytorchbot merge -f "CI passed" |
Collaborator
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
awgu
pushed a commit
to awgu/pytorch
that referenced
this pull request
Dec 20, 2022
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. Pull Request resolved: pytorch#90945 Approved by: https://github.com/awgu
This was referenced Dec 20, 2022
awgu
pushed a commit
to awgu/pytorch
that referenced
this pull request
Dec 20, 2022
Ensures that load_state_dict for fully_shard works: - Don't add back FSDP prefix - Small fix to ensure mixed precision check for buffers work Follow ups: - state_dict_type does not work, blocking rank0_only and CPU offload as well as other state dict implementations - No testing when wrapped with AC, using mixed precision, integration with distributed checkpoint, etc. Pull Request resolved: pytorch#90945 Approved by: https://github.com/awgu ghstack-source-id: 1fd8b50
This was referenced Dec 20, 2022
This was referenced Jan 5, 2023
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Stack from ghstack (oldest at bottom):
Ensures that load_state_dict for fully_shard works:
Follow ups: