Arm backend: Set fixed qparams for acos, asin and atanh by martinlsm · Pull Request #18288 · pytorch/executorch

martinlsm · 2026-03-18T11:55:51Z

acos, asin and atanh have shown to be problematic to quantize properly
due to their limited input range between [-1, 1] (inclusively for asin
and acos and exclusively for atanh). Before this patch, the approach for
quantizing these ops was to use the quantization spec from the
quantization config (typically a HistogramObserver if using deafult
symmetric quantization config). This caused problems when calibrating
the model with inputs close to -1 or 1, because they could land outside
the valid range of the operator. When this happened, the resulting TABLE
op set the output of these outliers to zero, which is not ideal.

To mitigate this problem, use fixed quantization params for these ops by
statically defining them in quantization_annotator.py. With this
solution, we potentially lose a bit of numerical precision because the
ops won't be affected by quantization calibration at all, but the
resulting TABLE ops won't set any zeros since there is no input that
can be outside the [-1, 1] interval anymore.

Also, do some general improvements:

In quantization_annotator.py, declare groups of ops as sets instead of
lists. The reason is that they are only used for lookup, so this change
improves the time complexity for such lookups.
Move down main section in analyze_output_utils.py

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell

The main section of a python file should be at the bottom to ensure everything from the file itself has been defined before running it. Signed-off-by: Martin Lindström <Martin.Lindstroem@arm.com> Change-Id: Ia534109ff87c9966641a9932c9fff8f6de5050fc

In quantization_annotator.py, declare groups of ops as sets instead of lists. The reason is that they are only used for lookup, so this change improves the time complexity for such lookups. Signed-off-by: Martin Lindström <Martin.Lindstroem@arm.com> Change-Id: I93428108b863c3dd2d6f80feea08859c059c187a

acos, asin and atanh have shown to be problematic to quantize properly due to their limited input range between [-1, 1] (inclusively for asin and acos and exclusively for atanh). Before this patch, the approach for quantizing these ops was to use the quantization spec from the quantization config (typically a HistogramObserver if using deafult symmetric quantization config). This caused problems when calibrating the model with inputs close to -1 or 1, because they could land outside the valid range of the operator. When this happened, the resulting TABLE op set the output of these outliers to zero, which is not ideal. To mitigate this problem, use fixed quantization params for these ops by statically defining them in quantization_annotator.py. With this solution, we potentially lose a bit of numerical precision because the ops won't be affected by quantization calibration at all, but the resulting TABLE ops won't set any zeros since there is no input that can be outside the [-1, 1] interval anymore. Signed-off-by: Martin Lindström <Martin.Lindstroem@arm.com> Change-Id: Id55156be5ca7fcfbf9f9c3f8ae88fb075509ce0c

pytorch-bot · 2026-03-18T11:55:55Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18288

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 Awaiting Approval, 1 New Failure, 1 Cancelled Job, 1 Pending, 3 Unrelated Failures

As of commit c420531 with merge base b1373e8 ():

AWAITING APPROVAL - The following workflows need approval before CI can run:

Claude Code (gh)
periodic (gh)

NEW FAILURE - The following job has failed:

pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model

CANCELLED JOB - The following job was cancelled. Please retry:

trunk / test-models-macos-cpu (llama, portable) / macos-job (gh)
##[error]The operation was canceled.

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

pull / unittest / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.
pull / unittest-editable / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.
trunk / unittest-release / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

martinlsm · 2026-03-18T11:56:04Z

@pytorchbot label ciflow/trunk

martinlsm · 2026-03-18T11:56:12Z

@pytorchbot label "partner: arm"

martinlsm · 2026-03-18T11:56:27Z

@pytorchbot label "release notes: none"

Copilot

Pull request overview

This PR updates the Arm backend quantization annotator to use fixed input quantization parameters for select domain-restricted ops (acos/asin/atanh) and adjusts corresponding operator tests, aiming to reduce out-of-range issues and stabilize INT test behavior.

Changes:

Add fixed-qparams input annotation path in quantization_annotator.py for aten.acos/asin/atanh.
Update acos/asin/atanh INT tests to rely on default similarity thresholds (and skip original-model comparison for atanh on invalid-domain inputs).
Move __main__ example block in analyze_output_utils.py to the bottom of the file.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
backends/arm/quantizer/quantization_annotator.py	Adds fixed-qparams quantization specs for acos/asin/atanh and refactors pattern containers to sets.
backends/arm/test/ops/test_atanh.py	Adjusts test vectors and conditionally skips original-model comparison at/over domain bounds.
backends/arm/test/ops/test_asin.py	Removes relaxed similarity thresholds to use pipeline defaults.
backends/arm/test/ops/test_acos.py	Removes relaxed Frobenius threshold to use pipeline defaults.
backends/arm/test/tester/analyze_output_utils.py	Relocates the `__main__` example block to the end of the module.

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

backends/arm/quantizer/quantization_annotator.py

+_fixed_input_qspec_ops: dict[Any, dict[int, _QParams]] = {
+    # acos has a valid range of [-1, 1]
+    torch.ops.aten.acos.default: {
+        8: _QParams((1.0 - (-1.0)) / (1 << 8), 0),
+        16: _QParams((1.0 - (-1.0)) / (1 << 16), 0),
+    },
+    # asin has a valid range of [-1, 1]
+    torch.ops.aten.asin.default: {
+        8: _QParams((1.0 - (-1.0)) / (1 << 8), 0),
+        16: _QParams((1.0 - (-1.0)) / (1 << 16), 0),
+    },
+    # atanh has a valid range of (-1, 1) (excluding -1 and 1).
+    torch.ops.aten.atanh.default: {
+        8: _QParams((0.999 - (-0.999)) / (1 << 8), 0),
+        16: _QParams((0.99999 - (-0.99999)) / (1 << 16), 0),


backends/arm/quantizer/quantization_annotator.py

        quant_properties.quant_output = _QuantProperty(0, output_act_qspec)
+    elif node.target in _fixed_input_qspec_ops:
+        num_bits = torch.iinfo(input_act_qspec.dtype).bits
+        qparams = _fixed_input_qspec_ops[node.target][num_bits]


backends/arm/test/tester/analyze_output_utils.py

+
+
+if __name__ == "__main__":
+    """This is expected to produce the example output of print_diff."""


backends/arm/test/ops/test_atanh.py

    )
+    if torch.any(test_data >= 1) or torch.any(test_data <= -1):
+        # The quantized model will saturate to max/min values while the
+        # original model will return inf/-inf, so comparison wont be valid here.


Martin Lindström added 3 commits March 18, 2026 09:24

martinlsm requested a review from digantdesai as a code owner March 18, 2026 11:55

Copilot AI review requested due to automatic review settings March 18, 2026 11:55

meta-cla bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Mar 18, 2026

pytorch-bot bot added the ciflow/trunk label Mar 18, 2026

pytorch-bot bot added the partner: arm For backend delegation, kernels, demo, etc. from the 3rd-party partner, Arm label Mar 18, 2026

pytorch-bot bot added the release notes: none Do not include this in the release notes label Mar 18, 2026

Copilot started reviewing on behalf of martinlsm March 18, 2026 11:56 View session

Copilot AI reviewed Mar 18, 2026

View reviewed changes

zingo approved these changes Mar 18, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Arm backend: Set fixed qparams for acos, asin and atanh#18288

Arm backend: Set fixed qparams for acos, asin and atanh#18288
martinlsm wants to merge 3 commits intopytorch:mainfrom
martinlsm:frobenius

martinlsm commented Mar 18, 2026 •

edited by pytorch-bot bot

Loading

Uh oh!

pytorch-bot bot commented Mar 18, 2026 •

edited

Loading

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants



		if __name__ == "__main__":
		"""This is expected to produce the example output of print_diff."""

Conversation

martinlsm commented Mar 18, 2026 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18288

❌ 2 Awaiting Approval, 1 New Failure, 1 Cancelled Job, 1 Pending, 3 Unrelated Failures

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

martinlsm commented Mar 18, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

martinlsm commented Mar 18, 2026 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Mar 18, 2026 •

edited

Loading