draft: add prototype for safer checkpoint format #3186
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Outdated suggestions cannot be applied.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.
This is a proof of concept for using a 'safer' checkpoint format in composer, mentioned in #3181.
This PR is intended for inspiration and is not expected to be merged.
pickleand therefore.ptand composer's.tarcheckpoint formats all have the possibility of executing arbitrary code during loading which is undesirable. The approach taken here to avoid this is to use 'inert' formats like.npy,.safetensorsandjsonto store data.This implementation patches the composer saving and loading functionality to use the custom format if the checkpoint is given the file extension:
.safe.When saving,
state_dictis traversed and anything not json serializable is extracted and replaced with some state that can be identified and restored during loading such as{"__safe_storage_obj": "timedelta", "seconds": 123}for atimedeltaobject, or{"__safe_storage_obj": "tensor", "id": "foo"}for a tensor (where the tensor data is stored separately).This proof of concept has lots of room for improvement:
.npyfiles instead of usingsavezso that the arrays aren't zipped twiceBefore submitting
pre-commiton your change? (see thepre-commitsection of prerequisites)