GPTQ - move calibration of quantiztion params to after hessian calibration #25
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
Satrat
reviewed
Jul 18, 2024
Comment on lines
+261
to
+265
|
|
||
| # quantization scales and zp are already initialized but we do not | ||
| # want to calibrate wrt to these | ||
| self.model.apply(disable_quantization) | ||
|
|
There was a problem hiding this comment.
One thought here: in the case where we have activation quantization and GPTQ on the weights, we will end up running calibration twice and I don't think we need to anymore. Since we aren't using the scale/zp during the Hessian calculation maybe we should try and set them in the same calibration as the hessian.
Not a requirement for now but just something to think about for the future.
horheynm
approved these changes
Jul 18, 2024
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Outdated suggestions cannot be applied.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.
fixes a small bug in GPTQ implemetation where scales and zero points were calculated before hessian calibration passes. this likely is the cause of the small drift in results we notice in the implemetation
additionally this change is important for the activation reordering feature as delaying the update until after the hessian is computed allows us to set the scales and zero points easily wrt the permuted weight matrix