MP cleanup and partial refactoring by irenaby · Pull Request #1391 · SonySemiconductorSolutions/mct-model-optimization

irenaby · 2025-03-16T17:16:50Z

Pull Request Description:

Simplify resource utilization constraints construction - compute total weights/bops directly instead of per node and aggregating at a later stage. Use RUCalculator to compute all targets at once and use the results directly. Remove separate computation for non-configurable nodes. Remove all intermediate methods per target and the double api for conf/non-conf which are no longer needed. Remove special hacky handling of total ru metric, which is now obtained automatically by RUCalculator along with the rest.
Convert LP solver module functions into class, remove dependency on MPSearchManager, move all code that is not related to LP to MPSearchManager.
Move more of MP search responsibility to MPSearchManager and make it the entry point for the search. The manager precomputes all metrics (sensitivity and ru) and uses LPSolver only for building and solving LP. Move virtual graph creation and restoring final config to MPSearchManager, and only create virtual graph for BOPS if both weights and activations are configurable (previously was created in any case of BOPS).

Checklist before requesting a review:

I set the appropriate labels on the pull request.
I have added/updated the release note draft (if necessary).
I have updated the documentation to reflect my changes (if necessary).
All function and files are well documented.
All function and classes have type hints.
There is a licenses in all file.
The function and variable names are informative.
I have checked for code duplications.
I have added new unittest (if necessary).

…ager

…, call LP from MPSearchManager with precomputed metrics

irenaby · 2025-03-17T07:44:40Z

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py

+                             f"following targets: {unsatisfiable_targets}")
+        return rel_target_ru
+
+    def _build_sensitivity_mapping(self, eps: float = EPS) -> Dict[int, Dict[int, float]]:


Moved from lp file, no change in implementation

irenaby · 2025-03-17T07:54:47Z

model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

+             in self.layer_to_indicator_vars_mapping.values()]
+        ).flatten()
+
+        return config.tolist()


Config reconstruction was moved to MP manager

irenaby · 2025-03-17T07:55:39Z

model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

+        return config.tolist()
+
+    @staticmethod
+    def _init_problem_vars(layer_to_metrics_mapping: Dict[int, Dict[int, float]]) -> Tuple[


No functional changes in this method

irenaby · 2025-03-17T07:55:54Z

model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

+
+        return layer_to_indicator_vars_mapping, layer_to_objective_vars_mapping
+
+    def _formalize_problem(self) -> LpProblem:


No functional changes in this method

irenaby · 2025-03-17T07:56:31Z

model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

+
+        return lp_problem
+
+    def _add_ru_constraints(self, lp_problem: LpProblem):


This method was modified.

irenaby · 2025-03-17T07:58:57Z

tests/keras_tests/non_parallel_tests/test_lp_search_bitwidth.py

@@ -57,144 +57,122 @@ def reconstruct_config_from_virtual_graph(self,


 class MockMixedPrecisionSearchManager:


Fixed to pass. Will be replaced by new tests.

elad-c

In general for tests: write a comment on each stating what that test tests.

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py

irenaby · 2025-03-23T09:19:13Z

In general for tests: write a comment on each stating what that test tests.

In general I agree, but no new tests were added, so I'm not sure what exactly you refer to.

irenab added 10 commits March 16, 2025 15:37

move virtual graph creation and search method call inside MP manager

fb3a2e0

fix test

d663f28

remove unused imports

2132708

move sensitivity computation from linear_programming to MP search man…

617e476

…ager

remove separate computation for non-configurable nodes

266f3f9

fix tests

5c619ad

simplify ru constraints construction in mixed precision

6fdd3d6

fix test

91f2aa5

tiny updates

5666542

convert LP functions into class, remove dependency on MPSearchManager…

42aeae8

…, call LP from MPSearchManager with precomputed metrics

github-actions bot added auto:core auto:tests labels Mar 16, 2025

fix lp_search test

07c9486

irenaby commented Mar 17, 2025

View reviewed changes

irenaby requested review from elad-c and ofirgo and removed request for ofirgo March 17, 2025 08:00

irenaby marked this pull request as ready for review March 17, 2025 08:00

ofirgo removed their request for review March 17, 2025 08:32

elad-c reviewed Mar 17, 2025

View reviewed changes

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py Show resolved Hide resolved

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py Outdated Show resolved Hide resolved

add missing type hints

ab69efb

elad-c approved these changes Mar 23, 2025

View reviewed changes

irenaby merged commit 20c321c into main Mar 23, 2025
36 checks passed

ofirgo deleted the mp_refactor3 branch July 2, 2025 13:03

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

MP cleanup and partial refactoring#1391

MP cleanup and partial refactoring#1391
irenaby merged 12 commits intomainfrom
mp_refactor3

irenaby commented Mar 16, 2025 •

edited

Loading

Uh oh!

irenaby Mar 17, 2025

Uh oh!

irenaby Mar 17, 2025

Uh oh!

irenaby Mar 17, 2025

Uh oh!

irenaby Mar 17, 2025

Uh oh!

irenaby Mar 17, 2025

Uh oh!

irenaby Mar 17, 2025

Uh oh!

elad-c left a comment

Uh oh!

Uh oh!

Uh oh!

irenaby commented Mar 23, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants


		return layer_to_indicator_vars_mapping, layer_to_objective_vars_mapping

		def _formalize_problem(self) -> LpProblem:


		return lp_problem

		def _add_ru_constraints(self, lp_problem: LpProblem):

		@@ -57,144 +57,122 @@ def reconstruct_config_from_virtual_graph(self,


		class MockMixedPrecisionSearchManager:

Conversation

irenaby commented Mar 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Description:

Checklist before requesting a review:

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

irenaby Mar 17, 2025

Choose a reason for hiding this comment

Uh oh!

elad-c left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

irenaby commented Mar 23, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

irenaby commented Mar 16, 2025 •

edited

Loading