[WIP] Refactored weight compression for further unification. #2181
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
force-pushed
the
nl/compress_ingored_scope
branch
from
089c389 to
3719612
Compare
ljaljushkin
changed the title
ljaljushkin
requested review from
l-bat,
alexsu52,
a team,
andreyanufr and
AlexKoff88
and removed request for
a team
ljaljushkin
commented
Oct 9, 2023
Comment on lines
+95
to
+97
| const_shape = nncf_node.layer_attributes.constant_attributes[weight_port_id]["shape"] | ||
| channel_axes = get_weight_channel_axes(nncf_node, weight_port_id) | ||
| axes = get_channel_agnostic_reduction_axes(channel_axes, const_shape) |
There was a problem hiding this comment.
@l-bat reduction axes are found differently using node and layer_attributes
Comment on lines
+82
to
+86
| weight_port_ids = nncf_node.layer_attributes.get_const_port_ids() | ||
| for weight_port_id in weight_port_ids: | ||
| weight_op_friendly_name = nncf_node.layer_attributes.constant_attributes[weight_port_id]["name"] | ||
| weight_node = friendly_name_to_op_map[weight_op_friendly_name] | ||
| if weight_node is None: |
There was a problem hiding this comment.
@l-bat please pay attention to the way of getting weight_node and manipulation with port_id
previously, it was different:
allowed_metatypes_to_const_port = {OVEmbeddingMetatype: [0], OVMatMulMetatype: [0, 1]}
for node in model.get_ordered_ops():
for const_port_id in allowed_metatypes_to_const_port[metatype]:
weight_node = get_operation_const_op(node, const_port_id)|
|
||
| @staticmethod | ||
| def is_node_with_weights(node: NNCFNode) -> bool: | ||
| return node.layer_attributes and node.layer_attributes.constant_attributes |
There was a problem hiding this comment.
@l-bat this is similar to SmoothQuant logic, but it was not presented in the original implementation of weight compression.
Comment on lines
+269
to
+271
| compression_algorithm = WeightCompression(mode, ratio, group_size) | ||
| graph = NNCFGraphFactory.create(model) | ||
| return compression_algorithm.apply(model, graph) |
There was a problem hiding this comment.
The main difference, that current implementation requires the nncf_graph.
But it shouldn't induce a lot of overhead. Will provide time and memory measurements later.
cc: @alexsu52
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## develop #2181 +/- ##
===========================================
+ Coverage 36.18% 36.36% +0.18%
===========================================
Files 479 478 -1
Lines 42960 42978 +18
===========================================
+ Hits 15546 15630 +84
+ Misses 27414 27348 -66
|
ljaljushkin
changed the title
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
Refactored weight compression for OpenVINO and Torch to have a single entry point -
WeightCompressionalgorithm.Now it's based on the NNCF graph and has a common method that finds nodes for compression.
Reason for changes
The refactoring allows to easily support ignored scope without code duplication
Related tickets
122223
Tests
weight compression tests