Register buffer init callbacks in llama #32428
Draft
Conversation
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
kwen2501
changed the title
Contributor
|
Hi @kwen2501, thanks for opening this PR! cc'ing @muellerzr as this likely overlaps with accelerate's handling of weights loading and @ArthurZucker re Llama |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Collaborator
ArthurZucker
left a comment
ArthurZucker
left a comment
There was a problem hiding this comment.
Thanks! I remember this is required for some specific usage with torch that we talked about! Let's iterate 🤗
Comment on lines
+1137
to
+1141
| # Create buffer init callbacks by extending the one from `LlamaModel`, | ||
| # i.e. appending a prefix to all buffer FQNs. | ||
| for key, val in self.model.buf_init_callbacks.items(): | ||
| new_key = ".".join(["model", key]) | ||
| self.buf_init_callbacks[new_key] = val |
Collaborator
There was a problem hiding this comment.
IMO this should go in the LlamaPreTrainedModel at best!
| self.original_inv_freq = self.inv_freq | ||
|
|
||
| # Save buffer init callback | ||
| LlamaModel.buf_init_callbacks.setdefault("rotary_emb.inv_freq", init_inv_freq) |
Collaborator
There was a problem hiding this comment.
can this not be called in the LlamaModel directly, it's a bit weird for us to register something like this at the class level
5 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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Outdated suggestions cannot be applied.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.
What does this PR do?
Non-persistent buffers is not saved in state dict.
In the case of meta init, while loading state dict from checkpoint can fill in parameters and persistent buffers, we need a way to initialize non-persistent buffers.
This PR does so by registering a buffer's init function against the buffer's FQN, and attaching such a callback dict to the model.
For how the init callbacks can be used, please refer to the
init_buffersutility in this PR:pytorch/PiPPy#1135
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.