use maxcut for total ru #1324
Merged
use maxcut for total ru #1324
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
irenaby
force-pushed
the
ru_refactor_total
branch
from
e718734 to
274f77d
Compare
irenaby
commented
Jan 13, 2025
| @@ -67,13 +67,19 @@ def __init__(self, | |||
| self.compute_metric_fn = self.get_sensitivity_metric() | |||
| self._cuts = None | |||
|
|
|||
| self.ru_metrics = target_resource_utilization.get_restricted_metrics() | |||
| # To define RU Total constraints we need to compute weights and activations even if they have no constraints | |||
There was a problem hiding this comment.
I couldn't just replace Total target computation without ugly hacks, so I changed the way it's handled altogether. I'm not thrilled about this implementation, but I think it's the lesser evil, as now we can actually reuse activations and weights utilization without recomputing it for Total and it simplifies the code around utilization matrix (and no need for the hack).
irenaby
commented
Jan 13, 2025
| @@ -510,21 +508,22 @@ def compare(self, quantized_model, float_model, input_x=None, quantization_info: | |||
| activation_bits = [layer.activation_holder_quantizer.get_config()['num_bits'] for layer in holder_layers] | |||
| # TODO maxcut: restore activation_bits == [4, 4] and unique_tensor_values=16 when maxcut calculates tensor sizes | |||
| # of fused nodes correctly. | |||
| self.unit_test.assertTrue((activation_bits == [4, 8])) | |||
| # TODO: maxcut Test updated but lowered activation ru (how can 4000 enforce 4,4??). Not sure what the fused nodes | |||
There was a problem hiding this comment.
@elad please comment on my comment to your comment :)
irenaby
commented
Jan 13, 2025
| @@ -174,13 +174,12 @@ def compare(self, quantized_model, float_model, input_x=None, quantization_info= | |||
| # test with its current setup (therefore, we don't check the input layer's bitwidth) | |||
| self.unit_test.assertTrue((activation_bits == [4, 8])) | |||
|
|
|||
| # TODO maxcut: restore this test after total_memory is fixed to be the sum of weight & activation metrics. | |||
There was a problem hiding this comment.
I restored some of maxcut TODOs but there are more remaining that I wasn't sure about
elad-c
approved these changes
Jan 14, 2025
| @@ -510,21 +508,22 @@ def compare(self, quantized_model, float_model, input_x=None, quantization_info: | |||
| activation_bits = [layer.activation_holder_quantizer.get_config()['num_bits'] for layer in holder_layers] | |||
| # TODO maxcut: restore activation_bits == [4, 4] and unique_tensor_values=16 when maxcut calculates tensor sizes | |||
| # of fused nodes correctly. | |||
| self.unit_test.assertTrue((activation_bits == [4, 8])) | |||
| # TODO: maxcut Test updated but lowered activation ru (how can 4000 enforce 4,4??). Not sure what the fused nodes | |||
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.
Pull Request Description:
Use max cut activation method for total resource utilization target.
Compute total target from weights and activations utilization instead of as a separate metric.
Simplify mixed precision search manger and linear programming to support 2d utilization matrix only instead of >=2.
Checklist before requesting a review: