Experimental support for 4-bit compression with LUT per layer and per block#3684
Conversation
| if reduction_axis == 0: | ||
| weight = fns.transpose(weight) | ||
| reduction_axis = 1 |
There was a problem hiding this comment.
A test for this condition would be nice
There was a problem hiding this comment.
@andreyanufr please add to the ticket with leftovers
|
|
||
| @pytest.mark.parametrize("value_type", [None, TensorDataType.float16, TensorDataType.f8e4m3, TensorDataType.int8]) | ||
| @pytest.mark.parametrize("group_size", [-1, 4]) | ||
| def test_adaptive_codebooks(value_type, group_size): |
There was a problem hiding this comment.
-
Suggest to add tests in separate file
tests/openvino/native/quantization/weights_compression/test_adaptive_codebook.py -
Added tests don’t increase coverage or add new functionality.
The current tests only verify that the function doesn’t fail with different arguments and check operation type types.
There was a problem hiding this comment.
@andreyanufr please add to the ticket with leftovers
| weighted_importance: Tensor | None = None | ||
|
|
||
|
|
||
| class KMeansWeighted: |
There was a problem hiding this comment.
Suggest to move it in separate code, and add tests for algorithm.
There was a problem hiding this comment.
@andreyanufr please add to the ticket with leftovers
2) Removed unused dataset from example.
ljaljushkin
left a comment
ljaljushkin
left a comment
There was a problem hiding this comment.
no blocking comments from my side
|
before merge
before release
All remaining comments should be appended to the ticket as leftovers to address once the algorithm is transitioned from experimental status. |
2) Added codebook estimation test duration.
MaximProshin
changed the title
MaximProshin
changed the title
MaximProshin
changed the title
6 participants
Changes
Implemented computation of codebook based on k-means algorithm.
Reason for changes
Related tickets
CVS-169609
CVS-180243 for leftovers
Tests
https://github.com/openvinotoolkit/nncf/actions/runs/21363309569