Scale estimation/rectification for int4 compression #2549
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 OpenVINO
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #2549 +/- ##
============================================
- Coverage 91.19% 29.95% -61.24%
============================================
Files 493 494 +1
Lines 45468 45775 +307
============================================
- Hits 41464 13713 -27751
- Misses 4004 32062 +28058
... and 319 files with indirect coverage changes
Flags with carried forward coverage won't be shown. Click here to find out more.
|
andreyanufr
changed the title
<style>
</style>
|
nncf/quantization/algorithms/weight_compression/openvino_backend.py
Outdated
Show resolved
Hide resolved
alexsu52
previously requested changes
Mar 20, 2024
nncf/quantization/algorithms/weight_compression/scale_estimation.py
Outdated
Show resolved
Hide resolved
2) Added docstrings.
2) Updated OV compression/decompression pieline.
alexsu52
reviewed
Apr 18, 2024
Comment on lines
+354
to
364
| model, | ||
| self._backend_entity.name_to_node_mapping, | ||
| all_weight_params, | ||
| nodes_to_compress, | ||
| activations, | ||
| awq_params.subset_size, | ||
| awq_params.percent_to_apply, | ||
| awq_params.alpha_min, | ||
| awq_params.alpha_max, | ||
| awq_params.steps, | ||
| ) |
There was a problem hiding this comment.
It looks like this is out of scope of the PR, but my opinion is __init__ parameters should be look like this:
Suggested change
| model, | |
| self._backend_entity.name_to_node_mapping, | |
| all_weight_params, | |
| nodes_to_compress, | |
| activations, | |
| awq_params.subset_size, | |
| awq_params.percent_to_apply, | |
| awq_params.alpha_min, | |
| awq_params.alpha_max, | |
| awq_params.steps, | |
| ) | |
| awq_params.subset_size, | |
| awq_params.percent_to_apply, | |
| awq_params.alpha_min, | |
| awq_params.alpha_max, | |
| awq_params.steps, | |
| ) |
This comment is something to think about.
nncf/quantization/algorithms/weight_compression/scale_estimation.py
Outdated
Show resolved
Hide resolved
|
|
||
| @staticmethod | ||
| def dump_parameters( | ||
| model: ov.Model, parameters: Dict, algo_name: Optional[str] = "quantization", path: Optional[List] = None | ||
| ) -> None: | ||
| dump_parameters(model, parameters, algo_name, path) | ||
|
|
||
| @staticmethod | ||
| def get_compress_decompress_pipeline( |
There was a problem hiding this comment.
IMHO:
Suggested change
| def get_compress_decompress_pipeline( | |
| def create_compress_decompress_fn( |
| return lambda w, s, zp: compiled_model([w, s, zp])[0] | ||
|
|
||
| @staticmethod | ||
| def get_compress_pipeline( |
There was a problem hiding this comment.
IMHO:
Suggested change
| def get_compress_pipeline( | |
| def create_compress_fn( |
nncf/quantization/algorithms/weight_compression/torch_backend.py
Outdated
Show resolved
Hide resolved
…stimation_pr_revert
|
Scale estimation algorithm doesn't work for group_size=-1 and fails with no clear message: |
ljaljushkin
approved these changes
Apr 24, 2024
alexsu52
reviewed
Apr 25, 2024
| return compressed_weights, scale, zero_point | ||
|
|
||
|
|
||
| def do_integer_quantization_with_fixed_scale( |
There was a problem hiding this comment.
Could you explain this refactoring? Let's discuss it offline.
Comment on lines
99
to
106
| def apply( | ||
| self, | ||
| model: TModel, | ||
| graph: NNCFGraph, | ||
| statistic_points: Optional[StatisticPointsContainer] = None, | ||
| dataset: Optional[Dataset] = None, | ||
| ) -> TModel: | ||
| """ |
There was a problem hiding this comment.
This function is 180 lines long and implements several critical steps at once (weights reshaping/preparing, rectification of initial scale, rectification of scale based on grid search). I believe stages separation in private functions with minimal description will not only improve readability, but allow to unit test separate parts independently
There was a problem hiding this comment.
@daniil-lyakhov Are you expecting something like this andreyanufr@2c99248 ?
There was a problem hiding this comment.
More like this fb883db, but looks like such refactoring is risky without comprehensive testing
There was a problem hiding this comment.
I suggest to create an issue for that
Comment on lines
99
to
106
| def apply( | ||
| self, | ||
| model: TModel, | ||
| graph: NNCFGraph, | ||
| statistic_points: Optional[StatisticPointsContainer] = None, | ||
| dataset: Optional[Dataset] = None, | ||
| ) -> TModel: | ||
| """ |
There was a problem hiding this comment.
More like this fb883db, but looks like such refactoring is risky without comprehensive testing
Co-authored-by: Daniil Lyakhov <[email protected]>
daniil-lyakhov
approved these changes
Apr 29, 2024
alexsu52
reviewed
Apr 30, 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
Added scale estimation for compression which minimizes L2 error between original MatMul and compressed one.
Reason for changes
Increases accuracy for compressed to 4 bit models.
Related tickets
CVS-129177
Tests
In process