[DCP] Removes Checkpoint Wrapped Prefix from state dict fqns #118119
Closed
+23
−2
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/118119
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 3bbd358 with merge base abe3c55 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
github-actions
bot
added
the
oncall: distributed
|
You can check https://github.com/pytorch/pytorch/blob/main/test/distributed/fsdp/test_fsdp_optim_state.py#L620. But remember NOT use FSDP to wrap the model as FSDP will handle the prefix for you. So you won't be able to test the logic if FSDP is used. |
fegin
approved these changes
Jan 25, 2024
Comment on lines
462
to
464
| for model, name in model_names: | ||
| for fqn in _get_fqns(model, name): | ||
| self.assertNotIn(_CHECKPOINT_WRAPPED_MODULE, fqn) |
There was a problem hiding this comment.
We should just call get_state_dict() and compare its keys with the original model without activation checkpoint (you can deepcopy the original model).
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Merge failedReason: 1 jobs have failed, first few of them are: Check mergeability and dependencies for ghstack prs / ghstack-mergeability-check Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
jeffdaily
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Feb 8, 2024
…#118119) Fixes pytorch#117399 ~~Soliciting some early feedback here.~~ ~~Do we happen to know if there already some tests that cover this case or would it make sense to add? @fegin , @wz337~~ Edit: Added tests Pull Request resolved: pytorch#118119 Approved by: https://github.com/fegin
pytorchmergebot
pushed a commit
that referenced
this pull request
Apr 24, 2024
Fixes #124546 When setting `use_orig_params = False` and using activation checkpointing, the FQN mapping as retrieved by the `_get_fqns` function is incorrect because the prefix that is added to the name of each activation checkpointed module, `_checkpoint_wrapped_module`, can still be present. I think this is an edge case with the `_get_fqns` function that was not addressed by this previous commit #118119. Without the change, the list of object names for an activation checkpointed module with FSDP (and `use_orig_params=False`) can be something like: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_checkpoint_wrapped_module', '_flat_param'] ``` Which will incorrectly return just one FQN, `{'model.transformer.blocks.0._flat_param'}`, when all the FQNs of the parameters of the transformer block should be returned. With the change, the list of object names will now have `_checkpoint_wrapped_module` removed: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_flat_param'] ``` And the FQNs are correctly retrieved and returned in `_get_fqns` when [this condition](https://github.com/pytorch/pytorch/blob/ea61c9cb299b6dfebc57dc9d8821c34321d568ab/torch/distributed/checkpoint/state_dict.py#L168) is satisfied. The correct FQNs are: ``` {'model.transformer.blocks.0.attn.Wqkv.bias', 'model.transformer.blocks.0.ffn.up_proj.bias', 'model.transformer.blocks.0.attn.out_proj.weight', 'model.transformer.blocks.0.norm_2.weight', 'model.transformer.blocks.0.ffn.down_proj.weight', 'model.transformer.blocks.0.attn.Wqkv.weight', 'model.transformer.blocks.0.norm_2.bias', 'model.transformer.blocks.0.ffn.up_proj.weight', 'model.transformer.blocks.0.ffn.down_proj.bias', 'model.transformer.blocks.0.norm_1.bias', 'model.transformer.blocks.0.norm_1.weight', 'model.transformer.blocks.0.attn.out_proj.bias'} ``` Pull Request resolved: #124698 Approved by: https://github.com/Skylion007
pytorchmergebot
pushed a commit
to xuhancn/pytorch
that referenced
this pull request
Apr 24, 2024
Fixes pytorch#124546 When setting `use_orig_params = False` and using activation checkpointing, the FQN mapping as retrieved by the `_get_fqns` function is incorrect because the prefix that is added to the name of each activation checkpointed module, `_checkpoint_wrapped_module`, can still be present. I think this is an edge case with the `_get_fqns` function that was not addressed by this previous commit pytorch#118119. Without the change, the list of object names for an activation checkpointed module with FSDP (and `use_orig_params=False`) can be something like: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_checkpoint_wrapped_module', '_flat_param'] ``` Which will incorrectly return just one FQN, `{'model.transformer.blocks.0._flat_param'}`, when all the FQNs of the parameters of the transformer block should be returned. With the change, the list of object names will now have `_checkpoint_wrapped_module` removed: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_flat_param'] ``` And the FQNs are correctly retrieved and returned in `_get_fqns` when [this condition](https://github.com/pytorch/pytorch/blob/ea61c9cb299b6dfebc57dc9d8821c34321d568ab/torch/distributed/checkpoint/state_dict.py#L168) is satisfied. The correct FQNs are: ``` {'model.transformer.blocks.0.attn.Wqkv.bias', 'model.transformer.blocks.0.ffn.up_proj.bias', 'model.transformer.blocks.0.attn.out_proj.weight', 'model.transformer.blocks.0.norm_2.weight', 'model.transformer.blocks.0.ffn.down_proj.weight', 'model.transformer.blocks.0.attn.Wqkv.weight', 'model.transformer.blocks.0.norm_2.bias', 'model.transformer.blocks.0.ffn.up_proj.weight', 'model.transformer.blocks.0.ffn.down_proj.bias', 'model.transformer.blocks.0.norm_1.bias', 'model.transformer.blocks.0.norm_1.weight', 'model.transformer.blocks.0.attn.out_proj.bias'} ``` Pull Request resolved: pytorch#124698 Approved by: https://github.com/Skylion007
petrex
pushed a commit
to petrex/pytorch
that referenced
this pull request
May 3, 2024
Fixes pytorch#124546 When setting `use_orig_params = False` and using activation checkpointing, the FQN mapping as retrieved by the `_get_fqns` function is incorrect because the prefix that is added to the name of each activation checkpointed module, `_checkpoint_wrapped_module`, can still be present. I think this is an edge case with the `_get_fqns` function that was not addressed by this previous commit pytorch#118119. Without the change, the list of object names for an activation checkpointed module with FSDP (and `use_orig_params=False`) can be something like: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_checkpoint_wrapped_module', '_flat_param'] ``` Which will incorrectly return just one FQN, `{'model.transformer.blocks.0._flat_param'}`, when all the FQNs of the parameters of the transformer block should be returned. With the change, the list of object names will now have `_checkpoint_wrapped_module` removed: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_flat_param'] ``` And the FQNs are correctly retrieved and returned in `_get_fqns` when [this condition](https://github.com/pytorch/pytorch/blob/ea61c9cb299b6dfebc57dc9d8821c34321d568ab/torch/distributed/checkpoint/state_dict.py#L168) is satisfied. The correct FQNs are: ``` {'model.transformer.blocks.0.attn.Wqkv.bias', 'model.transformer.blocks.0.ffn.up_proj.bias', 'model.transformer.blocks.0.attn.out_proj.weight', 'model.transformer.blocks.0.norm_2.weight', 'model.transformer.blocks.0.ffn.down_proj.weight', 'model.transformer.blocks.0.attn.Wqkv.weight', 'model.transformer.blocks.0.norm_2.bias', 'model.transformer.blocks.0.ffn.up_proj.weight', 'model.transformer.blocks.0.ffn.down_proj.bias', 'model.transformer.blocks.0.norm_1.bias', 'model.transformer.blocks.0.norm_1.weight', 'model.transformer.blocks.0.attn.out_proj.bias'} ``` Pull Request resolved: pytorch#124698 Approved by: https://github.com/Skylion007
mvpatel2000
pushed a commit
to mvpatel2000/pytorch
that referenced
this pull request
May 17, 2024
Fixes pytorch#124546 When setting `use_orig_params = False` and using activation checkpointing, the FQN mapping as retrieved by the `_get_fqns` function is incorrect because the prefix that is added to the name of each activation checkpointed module, `_checkpoint_wrapped_module`, can still be present. I think this is an edge case with the `_get_fqns` function that was not addressed by this previous commit pytorch#118119. Without the change, the list of object names for an activation checkpointed module with FSDP (and `use_orig_params=False`) can be something like: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_checkpoint_wrapped_module', '_flat_param'] ``` Which will incorrectly return just one FQN, `{'model.transformer.blocks.0._flat_param'}`, when all the FQNs of the parameters of the transformer block should be returned. With the change, the list of object names will now have `_checkpoint_wrapped_module` removed: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_flat_param'] ``` And the FQNs are correctly retrieved and returned in `_get_fqns` when [this condition](https://github.com/pytorch/pytorch/blob/ea61c9cb299b6dfebc57dc9d8821c34321d568ab/torch/distributed/checkpoint/state_dict.py#L168) is satisfied. The correct FQNs are: ``` {'model.transformer.blocks.0.attn.Wqkv.bias', 'model.transformer.blocks.0.ffn.up_proj.bias', 'model.transformer.blocks.0.attn.out_proj.weight', 'model.transformer.blocks.0.norm_2.weight', 'model.transformer.blocks.0.ffn.down_proj.weight', 'model.transformer.blocks.0.attn.Wqkv.weight', 'model.transformer.blocks.0.norm_2.bias', 'model.transformer.blocks.0.ffn.up_proj.weight', 'model.transformer.blocks.0.ffn.down_proj.bias', 'model.transformer.blocks.0.norm_1.bias', 'model.transformer.blocks.0.norm_1.weight', 'model.transformer.blocks.0.attn.out_proj.bias'} ``` Pull Request resolved: pytorch#124698 Approved by: https://github.com/Skylion007
atalman
pushed a commit
that referenced
this pull request
May 24, 2024
…26559) Fixes #124546 When setting `use_orig_params = False` and using activation checkpointing, the FQN mapping as retrieved by the `_get_fqns` function is incorrect because the prefix that is added to the name of each activation checkpointed module, `_checkpoint_wrapped_module`, can still be present. I think this is an edge case with the `_get_fqns` function that was not addressed by this previous commit #118119. Without the change, the list of object names for an activation checkpointed module with FSDP (and `use_orig_params=False`) can be something like: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_checkpoint_wrapped_module', '_flat_param'] ``` Which will incorrectly return just one FQN, `{'model.transformer.blocks.0._flat_param'}`, when all the FQNs of the parameters of the transformer block should be returned. With the change, the list of object names will now have `_checkpoint_wrapped_module` removed: ``` ['model', '_fsdp_wrapped_module', 'transformer', 'blocks', '0', '_fsdp_wrapped_module', '_flat_param'] ``` And the FQNs are correctly retrieved and returned in `_get_fqns` when [this condition](https://github.com/pytorch/pytorch/blob/ea61c9cb299b6dfebc57dc9d8821c34321d568ab/torch/distributed/checkpoint/state_dict.py#L168) is satisfied. The correct FQNs are: ``` {'model.transformer.blocks.0.attn.Wqkv.bias', 'model.transformer.blocks.0.ffn.up_proj.bias', 'model.transformer.blocks.0.attn.out_proj.weight', 'model.transformer.blocks.0.norm_2.weight', 'model.transformer.blocks.0.ffn.down_proj.weight', 'model.transformer.blocks.0.attn.Wqkv.weight', 'model.transformer.blocks.0.norm_2.bias', 'model.transformer.blocks.0.ffn.up_proj.weight', 'model.transformer.blocks.0.ffn.down_proj.bias', 'model.transformer.blocks.0.norm_1.bias', 'model.transformer.blocks.0.norm_1.weight', 'model.transformer.blocks.0.attn.out_proj.bias'} ``` Pull Request resolved: #124698 Approved by: https://github.com/Skylion007 Co-authored-by: Saaketh <narayan.saaketh@gmail.com>
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.
Fixes #117399
Soliciting some early feedback here.Do we happen to know if there already some tests that cover this case or would it make sense to add? @fegin , @wz337Edit: Added tests
cc @mrshenli @pritamdamania87 @zhaojuanmao @satgera @rohan-varma @gqchen @aazzolini @osalpekar @jiayisuse @H-Huang @kwen2501 @awgu @penguinwu @fegin @XilunWu @wanchaol @fduwjj @wz337 @tianyu-l @wconstab @yf225