Allow sharding fine tuned misaligned models #126
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
When loading large models, weights are sharded across a mesh of TPUs, splitting the original weights into smaller tensors, each one with the same shape. This is not possible, however, if the original weights shape is not divisible across the number of TPUs, because it results in a smaller tensor for the last TPU. This change pads the tensor with zeros, making it splittable across the TPUs.
baptistecolle
approved these changes
Dec 6, 2024
There was a problem hiding this comment.
LGTM but did you try RLHFlow/ArmoRM-Llama3-8B-v0.1 also? Should we add it to the tests? They mention it in the issue so it would be nice to check that it indeed work
|
@baptistecolle I could try it but RLHFlow/ArmoRM-Llama3-8B-v0.1 is a |
baptistecolle
pushed a commit
that referenced
this pull request
Dec 10, 2024
* fix(Makefile): re-add style target * feat(jetstream): pad weights to support unaligned sharding When loading large models, weights are sharded across a mesh of TPUs, splitting the original weights into smaller tensors, each one with the same shape. This is not possible, however, if the original weights shape is not divisible across the number of TPUs, because it results in a smaller tensor for the last TPU. This change pads the tensor with zeros, making it splittable across the TPUs.
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Outdated suggestions cannot be applied.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?
When loading large models, weights are sharded across a mesh of TPUs,
splitting the original weights into smaller tensors, each one with the
same shape.
This is not possible, however, if the original weights shape is not
divisible across the number of TPUs, because it results in a smaller
tensor for the last TPU.
This change pads the tensor with zeros, making it splittable across the
TPUs.
I expected this to cause a change in the model, but apparently that is not the case, and the model still has the same shape and produces a correct output. I suspect this might be due to the way the weight is loaded on the mesh, and the fact that the model figures out the part it should use, but I hadn't got the change to try many "misaligned" models,
the only one I used here it works now.
Fixes #67
Before submitting