Avoid random weights initialization when quantizing #291
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
The device parameter is added to qcreate, and scale buffers are created on the same device as the weights.
4 tasks
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Outdated suggestions cannot be applied.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?
As raised by @latentCall145, there is a useless random weights initialization when quantizing a module.
The solution suggested in #290 is correct but makes the low-level quantization API depend on
accelerate, which is only an optional dependency used by the high-level model API.This is more or less the same implementation, but more explicitly using the
metadevice. Note that we need to explicitly preserve the scales, since unlikeaccelerate, pytorch does not distinguish between parameters and buffers when skipping initialization.