Arm backend: Propagate node info from quantizer to backend #15300
Conversation
This file contains hidden or 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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15300
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (6 Unrelated Failures)As of commit 5e375e5 with merge base c8e9684 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
oscarandersson8218
added
partner: arm
meta-cla
bot
added
the
CLA Signed
| node.meta[Q_ANNOTATION_KEY]._annotated = True | ||
| meta_custom = node.meta.get("custom", {}) | ||
| meta_custom[ArmAnnotationInfo.CUSTOM_META_KEY] = annotation_info | ||
| node.meta["custom"] = meta_custom |
Collaborator
Author
There was a problem hiding this comment.
@digantdesai What are your thoughts on this conceptually?
Collaborator
|
@digantdesai I have a feeling the MCU backend will need something similar to this since mixed int/fp graphs will likely be more common there. |
Collaborator
|
Adding @SS-JIA as @digantdesai is out of office. |
oscarandersson8218
force-pushed
the
node_quant_metadata
branch
from
c9f432e to
ef42e7f
Compare
Contributor
|
Apologies for the delay, I will take a look at this PR tomorrow |
AdrianLundell
added a commit
to AdrianLundell/executorch
that referenced
this pull request
Nov 7, 2025
A number of ops only handles shape/meta-data without changing the dynamic range. In these cases, no rescaling needs to be performed and the int8 portable_ops kernel can be used directly. A new test is added to ensure this behaviour, as well as a test showing how operators which does change the dynamic range (SUB) are not supported. To support quantization of graphs with no-rescale ops in the beginning/ end of the graph, two new quantizers InputQuantizer and OutputQuantizer are introduced. By explicitly stating the dtype of the input/output, no-rescale ops inherit dtypes from them as with any other op. This change exposes the issue of mixing dtypes within the graph, which adds back xfails for the broadcasted add and mul tests. This can be fixed in a future patch after pytorch#15300 is resolved. Signed-off-by: Adrian Lundell <[email protected]> Change-Id: I8f79b86b633f9ad8d9f183c914754b0ee2f7a87c
oscarandersson8218
force-pushed
the
node_quant_metadata
branch
from
ef42e7f to
b3c3da5
Compare
Use the Node meta 'custom' field to propagate information from quantizer to partitioner using a new ArmAnnotationInfo data class. This allows us to track quantized node reliably which is useful in order to track which nodes should 'fold' it's quantization parameter and which should be kept in fp when mixing integer and float in a sub-graph. Co-authored-by: Per Åstrand <[email protected]> Signed-off-by: Oscar Andersson <[email protected]> Change-Id: I31309d65cac50e497318eae8678880684ec77cda
oscarandersson8218
force-pushed
the
node_quant_metadata
branch
from
b3c3da5 to
5e375e5
Compare
Contributor
|
@oscarandersson8218 unfortunately, this diff is causing a lot of failures internally, and we will have to revert. Sorry about that, I should have imported the changes first to make sure everything is ok. There were two types of failures - First, BUCK build failures which I was able to fix with #15759 The second is a bit trickier. Turns out we have some internal tests that try to serialize the exported model via When this happens, we see the error I believe the easiest fix for this would be to make Would you mind submitting a new PR with these changes? For the new one I will be sure to import the changes internally first. |
SS-JIA
added a commit
that referenced
this pull request
Nov 11, 2025
Collaborator
Author
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
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.
Use the Node meta 'custom' field to propagate information from quantizer to partitioner using a new ArmAnnotationInfo data class. This allows us to track quantized node reliably which is useful in order to track which nodes should 'fold' it's quantization parameter and which should be kept in fp when mixing integer and float in a sub-graph.
cc @freddan80 @per @zingo @digantdesai