Support more ops and verify more models by Jiseong-oh · Pull Request #14106 · pytorch/executorch

Jiseong-oh · 2025-09-09T05:34:12Z

Expands more ops and Verify more models

Summary

Added for more OPs currently supported by Exynos
Added functionality for models for Showcase
Additional Module for Optimization
Fixed ci issue

Test plan

models="ic3, ic4, resnet18, resnet50, mv3, edsr, dl3, w2l, vit, mobilebert
python -m executorch.examples.samsung.aot_compiler --model_name=$model -c E9955

cc @SS-JIA @digantdesai @kimishpatel

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

pytorch-bot · 2025-09-09T05:34:15Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14106

📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 1 Cancelled Job, 37 Pending, 2 Unrelated Failures

As of commit 3437e4c with merge base a89b858 ():

NEW FAILURE - The following job has failed:

pull / test-moshi-linux / linux-job (gh)
RuntimeError: Command docker exec -t 7fa09c9475ebe7705a957c74e40f24cec13617faeecc5cde8f66ea22c9bc41d0 /exec failed with exit code 1

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

pull / test-openvino-linux / linux-job (gh)

FLAKY - The following job failed but was likely due to flakiness present on trunk:

pull / test-setup-linux-gcc / linux-job (gh) (similar failure)
##[error]The operation was canceled.

BROKEN TRUNK - The following job failed but was present on the merge base:

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

pull / test-binary-size-linux-gcc / linux-job (gh) (trunk failure)
##[error]The operation was canceled.

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

- Add ops in op builders - batchmatmul, div, maximun, minimun, rsqrt, slice_copy, sqrt, to_copy Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: jingya-zhang <jingya.zhang@samsung.com>

- add squeeze, sub ops Co-authored-by: chong-chen <chong.chen@samsung.com>

1. Propagate constant and remove constant ops. 2. Prevent some ops from decomposing.(HardSwish, linear in mv3) 3. Add models name to support list (ic4, edsr enabled at the same time) in aot_compiler Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com>

Add prelu, softmax, layer_norm, upsample builders. Add these ops to list which keep ops not to decompose Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: xz-linghu <xz.linghu@samsung.com>

1. Add necessary ops - gelu, expand, etc. 2. Preprocess mul/add/sub scalar ops, add ReplaceScalarOps pass. Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com>

Convert conv1d to conv2d and support logsoftmax op. w2l can be enabled in samsung enn backend Co-authored-by: Jonghun Cha <jhbb.cha@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com>

Support the ops of bert float model (finetune) Co-authored-by: chong-chen <chong.chen@samsung.com>

Implement test for each op supported in builders Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: Jonghun Cha <jhbb.cha@samsung.com> Co-authored-by: jingya-zhang <jingya.zhang@samsung.com>

add more model tests to test workflow. Like mv3, dl3, vit, etc.

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

…utorch into expands_more_ops

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

SS-JIA · 2025-09-10T23:29:36Z

examples/samsung/aot_compiler.py

    model = model.eval()
    outputs = model(*example_inputs)

+    print("start start ...")


it's removed.

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

SS-JIA · 2025-09-10T23:32:14Z

examples/samsung/scripts/mobilebert_finetune.py

+# Output from pretrained model exceeds the representation scale of half-float.
+# Finetune bert model on specific task and make output more reasonable for hardware.
+# Here is an example.
+class MobileBertFinetune:


very cool/useful to have this script as an example, by the way

SS-JIA

LGTM. Please remove the print statement and I will merge.

swolchok · 2025-09-11T16:12:19Z

backends/samsung/builders/op_constant_pad_nd.py

+    def __init__(self, *args) -> None:
+        super().__init__(*args)


I think __init__ definitions like this can just be omitted

swolchok · 2025-09-11T16:16:41Z

backends/samsung/builders/op_bmm.py

+        input1 = node.args[0]
+        input_id_1 = self.define_tensor(input1, enn_graph, vals_to_ids)
+        input2 = node.args[1]
+        input_id_2 = self.define_tensor(input2, enn_graph, vals_to_ids)
+
+        # output
+        output_id = self.define_tensor(node, enn_graph, vals_to_ids)
+
+        enn_graph.define_op(
+            node.name, "BATCH_MATMUL", [input_id_1, input_id_2], [output_id]
+        )


it seems like there's a lot of boilerplate in these visitor definitions. You could package up a few helper subclasses like UnaryOpVisitor, BinaryOpVisitor, etc. that get the operator name ("BATCH_MATMUL" etc.) from a class property similar to the existing target property, and then also accommodate the ones with params by having the helper subclass call self.get_params() (default implementation that returns None on the helper subclass) and pass the result to define_op if it isn't None.

swolchok · 2025-09-11T16:20:26Z

backends/samsung/test/ops/test_log_softmax.py

+class LogSoftmax(torch.nn.Module):
+    def __init__(self, dim) -> None:
+        super().__init__()
+        self.module = torch.nn.LogSoftmax(dim=dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.module(x)


this looks like it could be deleted and replaced with using torch.nn.LogSoftmax directly

@SS-JIA

Expands more ops and Verify more models ### Summary - Added for more OPs currently supported by Exynos - Added functionality for models for Showcase - Additional Module for Optimization - Fixed ci issue ### Test plan models="ic3, ic4, resnet18, resnet50, mv3, edsr, dl3, w2l, vit, mobilebert python -m executorch.examples.samsung.aot_compiler --model_name=$model -c E9955 cc @SS-JIA @digantdesai @kimishpatel --------- Signed-off-by: jiseong.oh <jiseong.oh@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: jingya-zhang <jingya.zhang@samsung.com> Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: Jonghun Cha <jhbb.cha@samsung.com>

change docker image for using android ndk

6d929ae

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

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 Sep 9, 2025

Jiseong-oh added partner: samsung For backend delegation, kernels, demo, etc. from the 3rd-party partner, Samsung release notes: exynos labels Sep 9, 2025

Jiseong-oh and others added 9 commits September 10, 2025 07:01

update batchmul, div, max/min/rsqrt/slice_copy/sqrt/to_copy op

93dff05

- Add ops in op builders - batchmatmul, div, maximun, minimun, rsqrt, slice_copy, sqrt, to_copy Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: jingya-zhang <jingya.zhang@samsung.com>

add squeeze, sub ops

49345f8

- add squeeze, sub ops Co-authored-by: chong-chen <chong.chen@samsung.com>

Add more ops not to decompose

3058d75

Add prelu, softmax, layer_norm, upsample builders. Add these ops to list which keep ops not to decompose Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: xz-linghu <xz.linghu@samsung.com>

Support torchvision VIT float model

d643e22

1. Add necessary ops - gelu, expand, etc. 2. Preprocess mul/add/sub scalar ops, add ReplaceScalarOps pass. Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com>

Support w2l float model

9172064

Convert conv1d to conv2d and support logsoftmax op. w2l can be enabled in samsung enn backend Co-authored-by: Jonghun Cha <jhbb.cha@samsung.com> Co-authored-by: chong-chen <chong.chen@samsung.com>

Support bert float model (finetune)

35b2072

Support the ops of bert float model (finetune) Co-authored-by: chong-chen <chong.chen@samsung.com>

Support ops test

5a909e1

Implement test for each op supported in builders Co-authored-by: chong-chen <chong.chen@samsung.com> Co-authored-by: xz-linghu <xz.linghu@samsung.com> Co-authored-by: Jonghun Cha <jhbb.cha@samsung.com> Co-authored-by: jingya-zhang <jingya.zhang@samsung.com>

Add model test

09d10d6

add more model tests to test workflow. Like mv3, dl3, vit, etc.

Jiseong-oh requested review from SS-JIA, digantdesai, kimishpatel and mergennachin September 10, 2025 08:51

Jiseong-oh added 5 commits September 10, 2025 10:11

fix lint errors

49e54c6

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

enable bmm op test

5f1734e

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

Merge branch 'main' into expands_more_ops

ec28b28

fix lint issue for url-issue

4ae638b

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

Merge branch 'expands_more_ops' of https://github.com/Jiseong-oh/exec…

37ab8ec

…utorch into expands_more_ops

Jiseong-oh mentioned this pull request Sep 10, 2025

change docker image for using android ndk #14105

Closed

fix lint error

20780f2

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

SS-JIA reviewed Sep 10, 2025

View reviewed changes

apply review comments

3437e4c

Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

SS-JIA reviewed Sep 10, 2025

View reviewed changes

SS-JIA approved these changes Sep 11, 2025

View reviewed changes

SS-JIA merged commit 63481e3 into pytorch:main Sep 11, 2025
271 of 277 checks passed

swolchok reviewed Sep 11, 2025

View reviewed changes

Jiseong-oh deleted the expands_more_ops branch October 10, 2025 04:26

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support more ops and verify more models#14106

Support more ops and verify more models#14106
SS-JIA merged 17 commits intopytorch:mainfrom
Jiseong-oh:expands_more_ops

Jiseong-oh commented Sep 9, 2025 •

edited by pytorch-bot bot

Loading

Uh oh!

pytorch-bot bot commented Sep 9, 2025 •

edited

Loading

Uh oh!

SS-JIA Sep 10, 2025

Uh oh!

Jiseong-oh Sep 11, 2025

Uh oh!

SS-JIA Sep 10, 2025

Uh oh!

SS-JIA left a comment

Uh oh!

Uh oh!

swolchok Sep 11, 2025 •

edited

Loading

Uh oh!

swolchok Sep 11, 2025

Uh oh!

swolchok Sep 11, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

Jiseong-oh commented Sep 9, 2025 • edited by pytorch-bot bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

pytorch-bot bot commented Sep 9, 2025 • 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/14106

❌ 1 New Failure, 1 Cancelled Job, 37 Pending, 2 Unrelated Failures

Uh oh!

SS-JIA Sep 10, 2025

Choose a reason for hiding this comment

Uh oh!

Jiseong-oh Sep 11, 2025

Choose a reason for hiding this comment

Uh oh!

SS-JIA Sep 10, 2025

Choose a reason for hiding this comment

Uh oh!

SS-JIA left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

swolchok Sep 11, 2025 • 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!

swolchok Sep 11, 2025

Choose a reason for hiding this comment

Uh oh!

swolchok Sep 11, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Jiseong-oh commented Sep 9, 2025 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Sep 9, 2025 •

edited

Loading

swolchok Sep 11, 2025 •

edited

Loading