Samsung Exynos AI LiteCore - Support yolo26 by Jiseong-oh · Pull Request #22960 · pytorch/executorch

Jiseong-oh · 2026-09-20T23:59:28Z

Summary

Decompose remainder to floor_div+mul+sub
add op builder for amax, floor, floor_div, gather, max_dim
Refactor op builder for div, index, aplit, topk, upsample_nearest2d
add test and validate for yolo26
update annotation algorithm to cover graph topology from yolo26 model

Test plan

python test_yolo26.py -c E9955 -m yolo26s -d /path/to/images -p A8W8 --validate coco128.yaml

cc @SS-JIA @digantdesai @kimishpatel

pytorch-bot · 2026-09-20T23:59:31Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22960

📄 Preview Python docs built from this PR

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

✅ No Failures

As of commit 0e1184c with merge base 2c85103 ():
💚 Looks good so far! There are no failures yet. 💚

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

github-actions · 2026-09-21T00:00:17Z

This PR needs a `release notes:` label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

psiddh

Automated review of the yolo26 enablement changes. I checked the branch out into a worktree and verified several of these by execution rather than inspection (noted inline).

Highest priority: the op_upsample_nearest2d.py scale-factor change reads out_shape[0]/out_shape[1] (N and C) instead of [-2]/[-1]; I reproduced [0.0625, 0.25] instead of [2.0, 2.0] using the module already in test_upsample_nearest2d.py. Close behind: op_max_dim.py's guard is and-chained and drops the indices output in the only state RemoveGetItemPass can leave it in, and op_split_with_sizes_copy.py's copied_indices bookkeeping only lines up when getitem users happen to be in ascending order.

Worth noting on the positive side: removing DecomposeScaledDotProductAttention from quantize_module is correct — prepare_pt2e already runs it via EnnQuantizer.transform_for_annotation → transform_for_annotation_pass. And routing transform_for_export_pass through to_edge_transform_and_lower_to_enn fixes a real inconsistency with samsung_tester.py.

One process note: this adds five op builders and a new pass with no new tests under backends/samsung/test/ops/. An op-level test would have caught the upsample bug directly.

This review was generated by an AI reviewer (Claude Code). Findings are offered as starting points — please verify each one against your own understanding of the ENN backend before acting.

mergennachin

Checked this commit with Python reproductions and captured builder output. Device execution was unavailable because the ENN SDK bindings are missing on my machine.

Jiseong-oh · 2026-10-02T06:12:31Z

@psiddh @mergennachin Could you let me know if you have any comments? I am waiting for this pr to approval.

Updated saven_tensors code. this code will be merged in #22960 Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

psiddh · 2026-10-02T16:29:29Z

            _impl(user, res_list)
        return res_list

+    def _walk_qdq_chain_to_terminals(self, cur: Node) -> List[Node]:


Could we add focused graph-level tests for _walk_qdq_chain_to_terminals and the backward qparam propagation path? In particular: silu -> split/chunk, Q/DQ fanout, clone/contiguous, and a negative case where multi-input ops are not backward-propagated. If not in this PR, please track it as follow-up since any issues here could become silent accuracy regressions.

psiddh · 2026-10-02T16:34:10Z

+            return False
+
+        axis = len(indices) - 1
+        target_indices_node = indices[axis]


This still appears to assume the single non-None index is the last entry in indices. For a case like x[index, None], target_indices_node = indices[-1]would beNone. Could we find the actual non-None index and use its position as axis`, while still rejecting multiple non-None indices?

silu which is followed by split operator is not properly annotated to have qparam. detect activation function by backward propagation search Co-Authored-By: Jintech Noh <jintech.noh@samsung.com> Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

split_with_sizes_copy's getitem users may only consume part of the op's output, so propagate quantization parameters per-output instead of assuming every output is used, and support having output branches with differing quant params. Co-Authored-By: Jintech Noh <jintech.noh@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

Regarding remainder op, decompose it into the equivalent div/floor/mul/sub sequence before lowering and Wire the new pass into EnnPassManager. Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

Register the four new NodeVisitors in builders/__init__.py, and add aten.floor_divide.default to EnnPartitioner.ops_to_not_decompose so the new op_floor_divide visitor actually sees the op instead of its decomposition. Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

Refactor op_topk.py's output/dim handling and simplify op_index.py. Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

Wire EnnPassManager's transform_for_export_pass into to_edge_transform_and_lower_to_enn, drop the now-redundant DecomposeScaledDotProductAttention call from quantize_module, and rewrite examples/samsung/utils.py's save_tensors to recursively walk arbitrary nested tensor structures instead of a flat list. Add the yolo26 model test and validation example script, and loosen test_add's atol for the new coverage. Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

- op_upsample_nearest2d: Replaced the 2-element output_size argument with get_shape(node) - op_max_dim : The guard was and-chained, so len(users) == 1 short-circuited it to False and it never rejected anything - op_split_with_sizes_copy : Copied_indices was appended for every (output_idx, user) pair scanned rather than only on a match - op_topk : output tensors were appended in node.users iteration order rather than by getitem index - yolo26_validate : Updated according to review comments Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

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

- locate the index tensor by position in aten.index.Tensor - Save nested tensor structures when dumping LLM inputs and outputs Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

psiddh · 2026-10-03T08:05:47Z

lgtm, only thing can we have qparam propagation tests as a follow-up if they are not landing in this PR?

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 20, 2026

Jiseong-oh deployed to cadence September 20, 2026 23:59 — with GitHub Actions Active

Jiseong-oh added partner: samsung For backend delegation, kernels, demo, etc. from the 3rd-party partner, Samsung module: samsung labels Sep 21, 2026

Jiseong-oh force-pushed the support-yolo26 branch from c9195e5 to 9d35e34 Compare September 21, 2026 00:01

Jiseong-oh deployed to cadence September 21, 2026 00:02 — with GitHub Actions Active

Jiseong-oh deployed to cadence September 21, 2026 00:22 — with GitHub Actions Active

Jiseong-oh requested a review from psiddh September 21, 2026 00:48

psiddh reviewed Sep 22, 2026

View reviewed changes

Jiseong-oh force-pushed the support-yolo26 branch from 9d35e34 to 42b8a63 Compare September 28, 2026 10:11

Jiseong-oh had a problem deploying to cadence September 28, 2026 10:11 — with GitHub Actions Failure

mergennachin reviewed Sep 28, 2026

View reviewed changes

Jiseong-oh force-pushed the support-yolo26 branch from 42b8a63 to 462c341 Compare September 29, 2026 08:30

Jiseong-oh deployed to cadence September 29, 2026 08:31 — with GitHub Actions Active

Jiseong-oh deployed to cadence September 29, 2026 08:48 — with GitHub Actions Active

Jiseong-oh added a commit that referenced this pull request Oct 2, 2026

updated save_tensors in utils.py

ab4f077

Updated saven_tensors code. this code will be merged in #22960 Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

psiddh reviewed Oct 2, 2026

View reviewed changes

Jiseong-oh and others added 5 commits October 3, 2026 04:23

Rework topk and index builders

d70c21e

Refactor op_topk.py's output/dim handling and simplify op_index.py. Co-Authored-By: Jingya Zhang <jingya.zhang@samsung.com> Signed-off-by: Jiseong Oh <jiseong.oh@samsung.com>

Jiseong-oh and others added 4 commits October 3, 2026 04:23

Fix second round of review comment

87b94dd

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

Fixed third round of review comment

0e1184c

- locate the index tensor by position in aten.index.Tensor - Save nested tensor structures when dumping LLM inputs and outputs Signed-off-by: jiseong.oh <jiseong.oh@samsung.com>

Jiseong-oh force-pushed the support-yolo26 branch from 462c341 to 0e1184c Compare October 3, 2026 04:24

Jiseong-oh deployed to cadence October 3, 2026 04:24 — with GitHub Actions Active

Jiseong-oh deployed to cadence October 3, 2026 04:40 — with GitHub Actions Active

psiddh approved these changes Oct 3, 2026

View reviewed changes

Conversation

Jiseong-oh commented Sep 20, 2026 • 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 20, 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/22960

✅ No Failures

Uh oh!

github-actions Bot commented Sep 21, 2026

This PR needs a release notes: label

Uh oh!

psiddh left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mergennachin left a comment • 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!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Jiseong-oh commented Oct 2, 2026

Uh oh!

psiddh Oct 2, 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!

psiddh Oct 2, 2026

Choose a reason for hiding this comment

Uh oh!

psiddh commented Oct 3, 2026

Uh oh!

This branch was successfully deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Jiseong-oh commented Sep 20, 2026 •

edited by pytorch-bot Bot

Loading

pytorch-bot Bot commented Sep 20, 2026 •

edited

Loading

This PR needs a `release notes:` label

mergennachin left a comment •

edited

Loading

psiddh Oct 2, 2026 •

edited

Loading