Feat (mx): unpadding during dequantization #1134
Open
+109
−77
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
Giuseppe5
commented
Dec 19, 2024
| @@ -28,6 +28,7 @@ def apply_input_view(self, x): | |||
| return x.flatten(start_dim, start_dim + 1) | |||
|
|
|||
| def create_quant_tensor(self, qt_args: Tuple[Any]) -> GroupwiseFloatQuantTensor: | |||
| shape = self.tracked_parameter_list[0].shape | |||
There was a problem hiding this comment.
We don't support weight quant sharing for groupwise anyway, so this is safe, but it is ugly.
Giuseppe5
commented
Dec 19, 2024
| new_zp = self.zero_point_ | ||
|
|
||
| return new_value, new_scale, new_zp | ||
| from brevitas.utils.quant_utils import groupwise_dequant |
There was a problem hiding this comment.
Also ugly, maybe the function should live somewhere else?
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Outdated suggestions cannot be applied.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.
Reason for this PR
Groupwise quantization requires padding when the input channel shape is not divisible by groupsize.
Padding works well until it doesn't, and there are important edge cases that were not covered by the previous implementation.
(e.g., weight only quantization where padding was required. Until now, we also had to force activation quantization because otherwise we had shape mismatch).
Changes Made in this PR
With the current implementation, we un-pad when dequantizing, taking care of all the edge cases
Few todos:
Testing Summary
Risk Highlight
Checklist
devbranch.