[TorchAO] Migrate torch.ao -> torchao by anzr299 · Pull Request #3854 · openvinotoolkit/nncf

anzr299 · 2026-01-22T17:14:48Z

Changes

Migration of torch.ao to torchao.

Notable differences in torch.ao and torchao implementation which affect test references:

Constant folding in torch.ao produced the new constant with the name _frozen_param{index}. In the new torchao implementation it keeps the same name as the node being replaced
The old deprecated XNNPACK Quantizer was being used before which is now changed to point to the new XNNPACKQuantizer implementation in https://github.com/pytorch/executorch/tree/main/backends/xnnpack/quantizer. This implementation now explicitly quantizes the conv_transpose2D operator and its weights. Earlier only the activations were quantized for this op

Reference graphs changed due to above:

tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/mobilenet_v3_small.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/resnet18.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/swin_v2_t.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/synthetic_transformer.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/unet.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/vit_b_16.dot
tests/executorch/data/ao_export_quantization_OpenVINOQuantizer/yolo11n_sdpa_block.dot
tests/executorch/data/XNNPACKQuantizer/unet.dot (Due to difference Number 2)
tests/executorch/data/XNNPACKQuantizer/unet_ref_qconfig.json ((Due to difference Number 2))

Reason for changes

Related tickets

Tests

PTQ-791 - Pass
Weights Compression Conformance - https://github.com/openvinotoolkit/nncf/actions/runs/21260554857 - Pass
WC Examples - https://github.com/openvinotoolkit/nncf/actions/runs/21277782722 - Pass
Install Tests - https://github.com/openvinotoolkit/nncf/actions/runs/21260085794 - Pass

JacobSzwejbka · 2026-01-22T18:27:27Z

Oh nice! Thank you! I was just about to put up this change myself. Its blocking pytorch/executorch#16752

anzr299 · 2026-01-22T18:34:12Z

Oh nice! Thank you! I was just about to put up this change myself. Its blocking pytorch/executorch#16752

Great! Are you aware of the release schedule for executorch?
generally we upgrade the pytorch version and at the same time torchvision.
Can we expect Executorch to also be the same? Since I cannot see any tags for a new version

JacobSzwejbka · 2026-01-22T19:09:16Z

Oh nice! Thank you! I was just about to put up this change myself. Its blocking pytorch/executorch#16752

Great! Are you aware of the release schedule for executorch? generally we upgrade the pytorch version and at the same time torchvision. Can we expect Executorch to also be the same? Since I cannot see any tags for a new version

Yeah we release around the same time as PyTorch/PyTorch iirc. We cut a branch recently for 1.1 pytorch/executorch#16365

mergennachin · 2026-01-22T22:09:12Z

constraints.txt

 # Pytorch
 torch==2.9.0
 torchvision==0.24.0
+torchao==0.14.0


does it work with torchao 0.15?

we just pinned the main line and release branch to 0.15

We usually upgrade the torch and related dependency versions on nncf develop branch in a follow up PR after release branch is prepared which is soon.

Although, I suppose for Executorch release which uses torch 2.10 torch.ao used in nncf latest commits should be fine?

The nncf commit used here pytorch/executorch#16752 doesn't depend on torchao at all and instead uses torch.ao. This could be a problem with 2.11 I agree. This PR should unblock that.

daniil-lyakhov · 2026-01-26T14:22:34Z

constraints.txt

 # Pytorch
 torch==2.9.0
 torchvision==0.24.0
+torchao==0.14.0


Aamir, as we discussed it with Alexander Suslov, the main idea was to get rid of torch.ao and do not use any external dependencies (like torchao)
The task is to remove torch.ao imports and replace them with code in the nncf codebase, no external dependencies

Ah I see, since I couldn't see a ticket for the migration, it was a side-effect requirement for CVS-176783 which requires testing OVQuantizer from Executorch.
I will instead port torch.ao helper functions to NNCF.

…he which eats up memory

Copilot

Pull request overview

This PR migrates the codebase from using torch.ao to the newer standalone torchao library for quantization functionality.

Changes:

Updated all import statements from torch.ao.quantization.* to torchao.quantization.pt2e.* and executorch.backends.xnnpack.quantizer for XNNPACK
Added torchao as a dependency across requirements files
Updated test reference files to reflect behavior changes in the new library (constant naming and XNNPACK quantization scope)

Reviewed changes

Copilot reviewed 36 out of 36 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/torch/requirements.txt	Added torchao and executorch dependencies
tests/torch/fx/test_quantizer.py	Updated imports from torch.ao to torchao/executorch
tests/torch/fx/test_model_transformer.py	Updated imports and added device parameter to create_getattr_from_value calls
src/nncf/torch/quantization/strip.py	Updated observer imports to use torchao
src/nncf/quantization/algorithms/min_max/torch_fx_backend.py	Updated observer imports to use torchao
src/nncf/experimental/torch/fx/transformations.py	Updated imports, added device parameter handling, and implemented _get_model_device helper
src/nncf/experimental/torch/fx/quantization/quantizer/*.py	Updated imports and documentation from torch.ao to torchao
tests/torch/data/fx/*/.dot	Updated reference graphs for constant naming changes
tests/torch/data/fx/*/_ref_qconfig.json	Updated reference configs for operation ordering changes
examples/**/requirements.txt	Added torchao==0.14.0 dependency
constraints.txt	Added torchao==0.14.0 constraint

Comments suppressed due to low confidence (1)

tests/torch/fx/test_model_transformer.py:1

Corrected spelling of 'neccesary' to 'necessary'.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-17T08:19:49Z

src/nncf/experimental/torch/fx/transformations.py


    with graph.inserting_before(sorted_consumer_nodes[0]):
-        new_const = create_getattr_from_value(model, graph, node_name, value)
+        # Passing device is neccesary to avoid large models to be cached by torchao.


Corrected spelling of 'neccesary' to 'necessary'.

Suggested change

# Passing device is neccesary to avoid large models to be cached by torchao.

# Passing device is necessary to avoid large models to be cached by torchao.

Copilot · 2026-02-17T08:19:50Z

src/nncf/experimental/torch/fx/transformations.py

                # TODO(dlyakhov): maybe need more complex attr name here
-                qparam_node = create_getattr_from_value(model, graph, target_node.name + key, value_or_node)
+                tensor_device = value_or_node.device
+                # Passing device is neccesary to avoid large models to be cached by torchao.


Corrected spelling of 'neccesary' to 'necessary'.

Suggested change

# Passing device is neccesary to avoid large models to be cached by torchao.

# Passing device is necessary to avoid large models to be cached by torchao.

anzr299 requested a review from a team as a code owner January 22, 2026 17:14

github-actions bot added NNCF PT Pull requests that updates NNCF PyTorch NNCF PTQ Pull requests that updates NNCF PTQ labels Jan 22, 2026

init

f64a0f4

github-actions bot added the API Public API-impacting changes label Jan 22, 2026

add torchao to torch test requirements

b6275b7

anzr299 force-pushed the an/migrate_to_torchao branch from 650f120 to f64a0f4 Compare January 22, 2026 17:33

anzr299 added 3 commits January 22, 2026 21:39

add executorch to requirements

a193bca

revert openvino quantizer imports

9e671a5

revert some more changes

f41a16a

revert

ef50f87

mergennachin reviewed Jan 22, 2026

View reviewed changes

anzr299 added 2 commits January 23, 2026 11:01

add torchao in example requirements

d84f54f

update refs

a119db0

anzr299 force-pushed the an/migrate_to_torchao branch from edcf8fa to 431d9bd Compare January 26, 2026 07:40

This was referenced Jan 26, 2026

[DEBUG] Debug CI Issue for TorchAO With Upgrade #3859

Closed

[DEBUG] Debug CI Issue for TorchAO; Collect Data on Develop #3858

Closed

anzr299 force-pushed the an/migrate_to_torchao branch from 0a45b85 to 349b56f Compare January 26, 2026 11:48

daniil-lyakhov reviewed Jan 26, 2026

View reviewed changes

github-actions bot removed the API Public API-impacting changes label Jan 26, 2026

anzr299 marked this pull request as draft January 26, 2026 15:15

github-actions bot added the API Public API-impacting changes label Jan 27, 2026

anzr299 added 2 commits January 28, 2026 13:44

pass device also with create_getattr_from_value to avoid infinite cac…

d388f8f

…he which eats up memory

remove dead code; clean

e138d70

anzr299 force-pushed the an/migrate_to_torchao branch from d23af0b to d388f8f Compare January 28, 2026 09:52

anzr299 marked this pull request as ready for review January 28, 2026 10:39

Merge branch 'openvinotoolkit:develop' into an/migrate_to_torchao

2e22915

Merge branch 'develop' into an/migrate_to_torchao

5964281

Copilot AI review requested due to automatic review settings February 17, 2026 08:14

anzr299 marked this pull request as draft February 17, 2026 08:15

Copilot AI reviewed Feb 17, 2026

View reviewed changes

Copilot started reviewing on behalf of anzr299 February 17, 2026 08:44 View session

JacobSzwejbka mentioned this pull request Feb 18, 2026

Bump torch pin to 2.11.0.dev20260215 pytorch/executorch#17511

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[TorchAO] Migrate torch.ao -> torchao#3854

[TorchAO] Migrate torch.ao -> torchao#3854
anzr299 wants to merge 12 commits intoopenvinotoolkit:developfrom
anzr299:an/migrate_to_torchao

anzr299 commented Jan 22, 2026 •

edited

Loading

Uh oh!

JacobSzwejbka commented Jan 22, 2026

Uh oh!

anzr299 commented Jan 22, 2026 •

edited

Loading

Uh oh!

JacobSzwejbka commented Jan 22, 2026

Uh oh!

mergennachin Jan 22, 2026

Uh oh!

anzr299 Jan 23, 2026 •

edited

Loading

Uh oh!

daniil-lyakhov Jan 26, 2026

Uh oh!

anzr299 Jan 26, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Feb 17, 2026

Uh oh!

Copilot AI Feb 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Comments

	# Passing device is neccesary to avoid large models to be cached by torchao.
	# Passing device is necessary to avoid large models to be cached by torchao.

Conversation

anzr299 commented Jan 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Reason for changes

Related tickets

Tests

Uh oh!

JacobSzwejbka commented Jan 22, 2026

Uh oh!

anzr299 commented Jan 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

JacobSzwejbka commented Jan 22, 2026

Uh oh!

mergennachin Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

anzr299 Jan 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

daniil-lyakhov Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

anzr299 Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 17, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 17, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Comments

anzr299 commented Jan 22, 2026 •

edited

Loading

anzr299 commented Jan 22, 2026 •

edited

Loading

anzr299 Jan 23, 2026 •

edited

Loading