Meta device initialization for FSDP in Fabric #18122
Merged
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
awaelchli
added
strategy: fsdp
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
awaelchli
changed the title
Borda
approved these changes
Aug 1, 2023
awaelchli
changed the title
justusschock
approved these changes
Aug 2, 2023
carmocca
reviewed
Aug 9, 2023
carmocca
reviewed
Aug 9, 2023
| # Materialization happens in `setup`. When modules get wrapped by FSDP, the sequence of operations is: | ||
| # 1) materialize module 2) call `reset_parameters()` 3) shard the module. | ||
| # These operations are applied to each submodule 'bottom up' in the module hierarchy. | ||
| empty_init_context = torch.device("meta") |
There was a problem hiding this comment.
Doesn't this break the loading of checkpoints for models that haven't been FSDP wrapped with setup yet? For instance: https://github.com/Lightning-AI/lit-gpt/blob/1900b80424825cb221af0b63d19dd33b027d9aff/generate/base.py#L145-L151
Wouldn't load_state_dict need assign=True now? pytorch/pytorch#96161 (comment)
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Outdated suggestions cannot be applied.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?
Fixes #16448
Fixes #18008
This allows you to instantiate very large models that wouldn't fit in memory (either CPU or GPU) as fast as possible. No memory for weights get allocated, neither in CPU nor GPU memory and parameters are materialized/initialized with random weights directly at the time the model gets wrapped and sharded in
Fabric.setup().Notes:
reset_parameters()pytorch/pytorch#104187 in PyTorch 2.1 nightly.Requirement: Your submodules define a
reset_parameters()method that can be called to init the params. This is the case for all built-in PyTorch layers. If you have a custom layer, you'd have to add that method.Limitation:
Since the model is put on the meta device, you can't reference the parameters of that model in an optimizer like so:
The user has to change the code to set up the model first, then create the optimizer referencing the FSDP parameters:
We have checks for this and explain it to the user in an error message. This will also be documented (see note above) and we will find a way to lift this limitation in the future.
Before submitting
PR review
Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:
Reviewer checklist
cc @Borda @carmocca @justusschock @awaelchli