Arm backend: Remove silu decomposition pass by AdrianLundell · Pull Request #17202 · pytorch/executorch

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

AdrianLundell · 2026-02-04T13:07:46Z

Instead handle silu similar to other table op activations

cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai

Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

pytorch-bot · 2026-02-04T13:07:50Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17202

📄 Preview Python docs built from this PR

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

❌ 6 New Failures

As of commit 0c9e364 with merge base 3b819a2 ():

NEW FAILURES - The following jobs have failed:

pull / test-samsung-models-linux / linux-job (gh)
RuntimeError: Command docker exec -t c30dc42888f1c2926798beafbed5bdeb135c7b6c3a785edc05eb3d540df63d45 /exec failed with exit code 1
pull / test-samsung-quantmodels-linux / linux-job (gh)
RuntimeError: Command docker exec -t 3502a27da7a8a79e62c17ceede409a5386b6e1ceef266797736de58b65e5be50 /exec failed with exit code 1
pull / unittest / macos / macos-job (gh)
examples/models/test/test_export.py::ExportTest::test_mv3_export_to_executorch
pull / unittest-editable / linux / linux-job (gh)
backends/xnnpack/test/ops/test_conv2d.py::TestConv2d::test_dq_conv2d_seq
pull / unittest-editable / macos / macos-job (gh)
export/tests/test_target_recipes.py::TestTargetRecipes::test_mv3_model
trunk / test-mcu-cortex-m-backend / linux-job (gh)
RuntimeError: Command docker exec -t ddfe38c7ac5369e66390fbee10597d062f1a6d166d1e7a5f22c2636fe4457590 /exec failed with exit code 1

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

zingo

Approved BUT if possible please wait to after #17200 is merged and rebase before merging (if rest is green) to prove Arm tests PASS.

zingo · 2026-02-04T13:41:55Z

Merged now, I will rebase to test.

Issue: 1. SiLU's quantized output was forced to reuse its input's quantization scale, wasting resolution since SiLU's output range is always narrower than its input's. 2. The composable quantizer and `silu_.default` was separately misclassified in the legacy annotator since pytorch#17202. Fix: Give SiLU its own output qspec. Also fixes an unrelated latent bug that this change exposed. Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

@kimishpatel

### Summary SiLU's quantized output is forced to reuse its input's quantization scale, which wastes resolution for no benefit. This commit gives SiLU its own output quantization spec. ### Where this happens: | Quantizer | Mechanism | `silu.default` | `silu_.default` | | --- | --- | --- | --- | | Composable (`_TOSAQuantizerV2`, default since #19758) | `TOSAQuantizationConfig. SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | shared (bug) | shared (bug) | | Legacy (`_TOSAQuantizerV1`) | `quantization_annotator.py` op sets | own qspec | shared (bug) | ### Why quantization sharing is wrong for SiLU: <img width="512" height="492" alt="image" src="https://github.com/user-attachments/assets/ea71702f-bbe0-402a-b2b0-595b25e0d170" /> - **SiLU non-linearly compresses and warps its input range:** - Unlike ReLU, which maintains a 1:1 linear mapping ($y = x$) for non-negative inputs, SiLU is non-linear across its entire domain. - For negative inputs, its output never goes below -0.278, no matter how negative the input gets. So its output range is always narrower than its input's. Reusing the input scale misaligns the quantization bins with the actual output distribution, **wasting resolution.** - **Reusing the input's scale doesn't save any computation at runtime:** SiLU runs as a lookup table (`class TableOps`) whose entries are precomputed at compile time, not runtime. ### Changes | File | Change | Why | | --- | --- | --- | | `quantizer/quantization_config.py` | Remove `silu.default`/`silu_.default` from `SHARED_OUTPUT_ACT_QSPEC_PATTERNS` | The accuracy loss as described above. | | `quantizer/quantization_annotator.py` | Move `silu_.default` into `_one_to_one`, next to `silu.default` | The legacy annotator classified the two variants differently. #17202 put `silu.default` in `_one_to_one` (own qspec) but `silu_.default` in `_one_to_one_shared_input_or_input_act_qspec` (shared qspec). | | `test/ops/test_silu.py` | `Silu.forward` clones its input | Latent bug found while working on this change: SiLU(inplace=True) was mutating the test pipeline's shared input tensor which the pipeline reuses. It stayed invisible because the buggy (too large) output scale made the comparison helper(`compare_rel_frobenius_and_cosine_similarity()`) treat every value as noise and skip the check entirely, so the test kept passing without validating anything. | | `test/misc/test_shared_qspecs.py` | Add `NonSharedQspecSilu` + `test_silu_does_not_share_input_qspec` | Check the emitted q/dq parameters, the same way the rest of `test_shared_qspecs.py` checks sharing. | ### Testing ```bash pytest backends/arm/test/misc/test_shared_qspecs.py # 17 passed pytest backends/arm/test/ops/test_silu.py -k "not vgf" # 34 passed, 32 xfailed ``` cc @kimishpatel @jerryzh168 @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani Signed-off-by: Youngsik Yang <vacu9708@gmail.com>

Arm backend: Remove silu decomposition pass
dcca5f5
Instead handle silu similar to other table op activations Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: If3253ef6a3e76bd6fd5887b9d43874af0a897705

AdrianLundell added the partner: arm For backend delegation, kernels, demo, etc. from the 3rd-party partner, Arm label Feb 4, 2026

AdrianLundell requested a review from digantdesai as a code owner February 4, 2026 13:07

AdrianLundell added ciflow/trunk release notes: none Do not include this in the release notes labels Feb 4, 2026

meta-claBot 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 Feb 4, 2026

zingo approved these changes Feb 4, 2026
View reviewed changes

Merge branch 'main' into change-1194162
0c9e364

AdrianLundell merged commit 4aacfd6 into pytorch:mainFeb 5, 2026
300 of 306 checks passed

AdrianLundell deleted the change-1194162 branch February 17, 2026 17:14

vacu9708 mentioned this pull request Jul 28, 2026
Arm backend: Give SiLU its own output quantization params #21437
Merged