Samsung Exynos AI LiteCore - Support yolo26

[Jiseong-oh]

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]

🔗 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 9d35e34 with merge base dba8a83 ():
💚 Looks good so far! There are no failures yet. 💚

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

[github-actions]

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.

[psiddh]

Wrong axes on out_shape — scale factor is off for every size=-style upsample.

out_shape = get_shape(node)
scale_factor = [
    out_shape[0] * 1.0 / in_shape[-2],
    out_shape[1] * 1.0 / in_shape[-1],
]

get_shape returns the full tensor shape, so for NCHW out_shape[0] is N and out_shape[1] is C. The output_size this replaced was a 2-element [H, W] list, which is why indices [0]/[1] used to be correct.

Reproduced with the module already in backends/samsung/test/ops/test_upsample_nearest2d.py:

args: (x, [32, 32], None)
in_shape: [1, 4, 16, 16]   out_shape: [1, 4, 32, 32]
this PR : [0.0625, 0.25]
expected: [2.0, 2.0]

args[2] (scale_factors) is None whenever size= is used, so the override on line 40 does not rescue it. This is the common path.

Fix: out_shape[-2] / out_shape[-1].

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

The guard is and-chained, and in the only reachable case the indices output is silently dropped.

users = list(node.users.keys())
if (
    len(users) != 1
    and users[0].target.__name__ == "getitem"
    and users[0].args[1] != 0
):
    return False

Two issues that compound:

1. To bail out on anything other than "exactly one user, a getitem(0)" these clauses need or with negated comparisons. As written, len(users) == 1 short-circuits the guard to False no matter what that user is — and users[0].target.__name__ raises AttributeError for a non-call_function user (e.g. output, whose target is the string "output").

2. RemoveGetItemPass runs before serialization (_passes/enn_pass_manager.py:79) and rewrites single-consumer aten.max.dim into exir_ops.edge.aten.amax.default, raising for any count other than 1 or 2 (backends/transforms/remove_getitem_op.py:30-46). So this visitor only ever sees max.dim with exactly two users — values and indices. It then defines only output_idx=0, registers only vals_to_ids[users[0]], and emits a single-output ReduceMax; the indices getitem never gets a tensor id. And if users[0] is the indices getitem, all three clauses are true and the build fails.

Either emit both outputs the way op_topk.py now does, or return False when the indices output is live.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

copied_indices is appended unconditionally, so copied_indices[idx] only lines up by accident.

copied_indices.append(copied_idx)        # runs for every (output_idx, user) pair
if copied_idx == output_idx:
    output_id = self.define_tensor(user, enn_graph, vals_to_ids)
    all_output_tensors.append(output_id)

copied_indices ends up with len(args[1]) * num_getitem_users entries while all_output_tensors has num_getitem_users. Line 53's copied_indices[idx] therefore reads the first N entries — getitem indices in node.users iteration order — whereas all_output_tensors[idx] is ordered by ascending output_idx.

Those orders coincide only when the getitems were created in ascending index order. With x = s[2] written before y = s[1], copied_indices[0] is 2 while all_output_tensors[0] is index 1's tensor, and the emitted STRIDEDSLICE bounds are swapped.

Fix: move the append inside the if copied_idx == output_idx: branch.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

Two smaller things in this block:

• Line 52, end = in_shape aliases the list from get_shape; end[axis] = ... mutates it in place, and every iteration's params["end"] is the same object. Harmless today because define_op copies eagerly, but end = list(in_shape) costs nothing.
• "shrink_axis_mask": pow(2, axis) — a split_with_sizes_copy output keeps the split axis with size points[i]; shrinking it would drop the dimension entirely. Is that what ENN's STRIDEDSLICE expects here, or should this be 0?

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

Output order follows node.users iteration order rather than the getitem index.

outputs is passed straight to define_op(..., all_output_tensors, ...) alongside output="both". When both getitems exist but getitem(1) was created first — e.g. t = torch.topk(x, k); i = t.indices; v = t.values — outputs comes out as [indices, values] and the two ENN outputs are transposed. The previous code appended value then index unconditionally, so the ordering was structural.

test_topk.py returns tuple(torch.topk(...)), which always produces ascending order, so it won't catch this.

Separately: a topk whose result tuple is consumed by a non-getitem user hits the continue on line 62, leaving outputs empty while output_type stays "both" — a zero-output op.

Fix: collect into {idx: tensor_id} and emit by index rather than by append order.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

The model is lowered twice.

edge_prog = to_edge_transform_and_lower_to_enn(pt_model, example_args, compile_specs=compile_specs)
edge = edge_prog.to_backend(EnnPartitioner(compile_specs))

to_edge_transform_and_lower_to_enn (backends/samsung/utils/export_utils.py:78-82) already calls to_edge_transform_and_lower with EnnPartitioner, so this re-runs partitioning over an already-delegated program. Drop this line and call edge_prog.to_executorch(...) directly.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

Two argparse foot-guns in this block:

• --dump uses type=bool with nargs="?", so --dump False evaluates to bool("False") == True. action="store_true" is what's meant.
• --input_dims (line 314) uses type=eval, which evaluates arbitrary input; ast.literal_eval or nargs=2, type=int is safer. Relatedly, the args.input_dims != [640, 640] check on line 264 rejects a tuple literal such as (640,640).

Also: line 159 constructs RuntimeExecutor(exec_prog, input_tensor) inside the per-batch loop, so the .pte is re-pushed to the device once per image (128× for coco128). And the usage examples on lines 28-31 say python test_yolo26.py, but the file is yolo26_validate.py.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

Copyright (c) Intel Corporation — presumably carried over from OpenVINO's export_and_validate.py, which this mirrors. Worth confirming that's the intended attribution for a file in examples/samsung/.

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

This branch is unreachable as the pass is currently wired.

Both const-materialisation branches are gated on is_x_scalar and is_edge / is_y_scalar and is_edge. The pass is registered only in transform_for_export_pass (_passes/enn_pass_manager.py:63), which runs on the pre-edge ATen graph, so is_edge is always False and create_const_node never fires. I confirmed the ATen path works fine without it — scalars pass through as raw Python values and floor_divide dispatches correctly, including for remainder.Scalar_Tensor.

Beyond the dead code: if the pass were moved to the edge stage, create_const_node would register_buffer and insert a get_attr into an ExportedProgram's graph without updating graph_signature or state_dict. torch.export lifts buffers to placeholders, so a bare get_attr isn't valid there.

Either drop the scalar/const machinery, or register the pass where it's needed and extend the helper to update the program signature.

(The decomposition itself checks out — I ran it standalone over Scalar, Scalar_Tensor, Tensor, and integer inputs; max abs error ≤ 2.4e-7.)

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.

[psiddh]

FP32 tolerance loosened 3× with no explanation.

The harness default is atol=1e-3 (backends/test/harness/tester.py:321). This is the FP32 path of a plain elementwise add, where near-exact agreement is the expectation, and the change isn't mentioned in the PR description.

If an FP32 add genuinely started needing 3e-3 during this work, that's worth understanding rather than absorbing — the annotate_2in1out change in quantizer/annotator.py is a candidate, though it shouldn't affect the FP32 path. What regressed?

---

Generated by an AI reviewer (Claude Code). Please verify before acting on it.