Weight compression via Lora Correction Algorithm #2816
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
github-actions
bot
added
NNCF PT
ljaljushkin
commented
Jul 16, 2024
nncf/quantization/algorithms/weight_compression/openvino_backend.py
Outdated
Show resolved
Hide resolved
ljaljushkin
force-pushed
the
nl/lora_correct_prod_squash
branch
2 times, most recently
from
a355df6 to
24b1f9e
Compare
github-actions
bot
added
the
documentation
andreyanufr
reviewed
Jul 23, 2024
| for i in range(n_iters): | ||
| VX = Vr @ X | ||
| if not w_regularization: | ||
| sol = fns.linalg.lstsq(fns.transpose(VX), fns.transpose(dY)) |
There was a problem hiding this comment.
thanks, corrected
andreyanufr
reviewed
Jul 23, 2024
| if not w_regularization: | ||
| sol = fns.linalg.lstsq(fns.transpose(X), fns.transpose(dYU), driver="gelsy") | ||
| else: | ||
| Ind = fns.eye(Vr.shape[1], backend=Vr.backend, dtype=Vr.dtype) |
There was a problem hiding this comment.
May be also worth to add some math formulas in comment
There was a problem hiding this comment.
Done, also refactored to avoid too many transpose. please take a look at it as well
alexsu52
reviewed
Jul 24, 2024
nncf/quantization/algorithms/weight_compression/lora_correction.py
Outdated
Show resolved
Hide resolved
nncf/quantization/algorithms/weight_compression/lora_correction.py
Outdated
Show resolved
Hide resolved
nncf/quantization/algorithms/weight_compression/openvino_backend.py
Outdated
Show resolved
Hide resolved
nncf/quantization/algorithms/weight_compression/weight_lowering.py
Outdated
Show resolved
Hide resolved
ljaljushkin
force-pushed
the
nl/lora_correct_prod_squash
branch
from
470fbd8 to
a3b26c7
Compare
…nl/lora_correct_prod_squash
andreyanufr
reviewed
Aug 23, 2024
| for i in range(n_gs): | ||
| offset = i * gs | ||
| denum = fns.sum(s[offset : offset + gs]) | ||
| s[offset : offset + gs] = s[offset : offset + gs] / denum |
There was a problem hiding this comment.
I think we can skip first normalization
andreyanufr
approved these changes
Aug 26, 2024
alexsu52
reviewed
Aug 27, 2024
|
Latest build of conformance test with lora test case: |
alexsu52
approved these changes
Aug 28, 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.
Changes
Lora Correction algorithm for int4/nf4 weight compression.
Reason for changes
Method for improving accuracy by migrating quantization noise to “learnable” lora adapters.
Related tickets
135863
Tests