[FX] Support weight quantization for operations where weight_port_id != 1
#3334
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
anzr299
requested changes
Mar 10, 2025
anzr299
requested changes
Mar 11, 2025
Comment on lines
+464
to
+467
| if is_weight: | ||
| channel_axes = self._backend_entity.get_weight_quantization_axes(node, target_point, len(shape)) | ||
| else: | ||
| channel_axes = (1,) |
There was a problem hiding this comment.
This is not required, it can be reverted back to the old code.
|
|
||
| # Weight statistics is constant, so only one collection is enough. | ||
| range_estimator_params = self._get_range_estimator_parameters(target_point, qconfig) |
| ) | ||
| channel_idx = channel_axes[0] if channel_axes else 0 | ||
|
|
||
| if is_weights and not channel_axes: |
There was a problem hiding this comment.
Suggested change
| if is_weights and not channel_axes: | |
| if not len(channel_axes): |
to cover the case of vector weights which are being quantized per channel
| scale_shape = tuple( | ||
| get_scale_shape(input_shape, is_weights=is_weights, per_channel=per_channel, channel_idx=channel_idx) | ||
| ) | ||
| channel_idx = channel_axes[0] if channel_axes else 0 |
There was a problem hiding this comment.
Suggested change
| channel_idx = channel_axes[0] if channel_axes else 0 |
Since channel axes is already being checked and handled in the if-else block below. channel_axes[0] can directly be passed to channel_idx parameter of get_scale_shape
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Outdated suggestions cannot be applied.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
Updated the FX backend’s
_get_input_scale_shapeto use the FX insertion point shape and, when available, the actual weight tensor’s shape to compute the per‑channel scale shape.Adjusted statistics collector in
_get_stat_collectorso that the reduction and aggregation axes are derived using the same channel axes as used for scale shape computation.Related tickets
Issue #3206
Tests
All tests run successfully