Support extra ops modes for LLM Models

[Jiseong-oh]

Summary

Add Extra Ops for Exynos Backend

• Cos, GroupNorm, Index, Log, Pow, Rms_Norm, Sigmoid, sing, Splite_with_sizes_copy, sum_int_list, tanh and Topk
• All ops are verfied
• These ops will be used for supporting LLM models

Support Perf mode

• It can be supported Perf mode with experimental
• This mode MUST BE used to verify model on exynos device farm fristly before testing on the phone

cc @SS-JIA @digantdesai @kimishpatel

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18670

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 5 Pending

As of commit 2694d01 with merge base 4f689f3 ():
💚 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.

[SS-JIA]

@claude help review this PR

[claude]

Claude finished @SS-JIA's task in 4m 41s —— View job

---

Review of PR #18670: Support extra ops modes for LLM Models

• Gather context and understand the PR changes
• Review new op builders
• Review compose_rms_norm pass
• Review test files
• Review serialization and compile options changes
• Review example script changes and perf mode support
• Post comprehensive review

---

Overview

This PR adds 12 new op builders for the Samsung Exynos backend (Cos, GroupNorm, Index, Log, Pow, RmsNorm, Sigmoid, Sin, SplitWithSizesCopy, SumIntList, Tanh, TopK), a RecomposeRmsNorm pass, and performance mode support via a new PerformanceMode enum. Each new op has a corresponding unit test. All example scripts are updated to use PerformanceMode.HIGH_PERFORMANCE by default.

Overall the PR is well-structured and follows existing patterns. Below are the issues I found, organized by severity.

---

Issues

1. Bug: op_pow.py target vs test mismatch — backends/samsung/builders/op_pow.py:20

The builder registers target = "aten.pow.Tensor_Tensor", but the test in test_pow.py:36 checks for torch.ops.aten.pow.Tensor_Scalar (using x**2.0 which produces Tensor_Scalar). Additionally, the partitioner adds exir_ops.edge.aten.pow.Tensor_Scalar to SUPPORTED_OPS (enn_partitioner.py:42).

This means the builder only handles Tensor_Tensor pow, but the test and partitioner expect Tensor_Scalar. The Tensor_Scalar variant would match via the partitioner's SUPPORTED_OPS list and presumably be handled by the ReplaceOpsWithScalar pass, but the builder itself would never directly match Tensor_Scalar ops. If both variants need builder support, the builder's target should be a list including both. If only Tensor_Tensor is intended for the builder and Tensor_Scalar is handled through the scalar replacement pass, the test should be updated to test the Tensor_Tensor case that actually exercises the builder.

Fix this →

2. compose_rms_norm.py — potential _get_eps_node / _get_gamma_node returning None — backends/samsung/_passes/compose_rms_norm.py:22-33

def _get_eps_node(self, nodes):
    add_node = [n for n in nodes if hasattr(n, "name") and "add" in n.name][0]
    for a in add_node.args:
        if isinstance(a, float) or a.op != "call_function":
            return a

def _get_gamma_node(self, output_node):
    for a in output_node.args:
        if a.op != "call_function":
            return a

If no arg matches the condition, these methods implicitly return None, which would then be passed as the eps/gamma argument to rms_norm. Consider raising an error if the expected node is not found rather than silently passing None.

Fix this →

3. compose_rms_norm.py — fragile node name matching — backends/samsung/_passes/compose_rms_norm.py:24

add_node = [n for n in nodes if hasattr(n, "name") and "add" in n.name][0]

Matching by "add" in n.name is fragile — node names like "add_padding", "broadcast_add", etc. could be false positives. The Qualcomm version (backends/qualcomm/_passes/recompose_rms_norm.py) matches by checking node.target against known ATen add ops, which is more robust. Consider matching on node.target instead of node.name.

Fix this →

4. Hardcoded high-performance mode in all examples — examples/samsung/aot_compiler.py:81

Every example script now unconditionally uses PerformanceMode.HIGH_PERFORMANCE. The PR description warns "This mode MUST BE used to verify model on exynos device farm firstly before testing on the phone." But none of the scripts expose this as a CLI argument, so users can't opt out. The aot_compiler.py (the general-purpose compiler) especially should probably default to DEFAULT or expose --perf_mode as an argument.

5. Copyright header — backends/samsung/_passes/compose_rms_norm.py:1

The file has # Copyright (c) Qualcomm Innovation Center, Inc as the first copyright line. This file is derived from backends/qualcomm/_passes/recompose_rms_norm.py (same class name, similar approach), using a get_source_partitions-based implementation. The Qualcomm copyright is appropriate if code was derived from their work, but verify this dual-copyright header is intentional.

6. gen_samsung_backend_compile_spec type hint — backends/samsung/serialization/compile_options.py:76

def gen_samsung_backend_compile_spec(
    chipset: str,
    perf_mode: PerformanceMode = None,
):

The type hint says PerformanceMode but the default is None. Should be Optional[PerformanceMode] = None for correctness.

7. Inconsistent op name casing in define_op calls

Most existing op builders use ALL_CAPS for the op type string (e.g., "SQRT", "SIGMOID", "TANH", "LOG"), but some new ops use mixed casing:

• op_sin.py:31: "Sin" (vs "SIN")
• op_cos.py:31: "Cos" (vs "COS")
• op_topk.py:74: "TopK" (vs "TOPK")

The existing op_hardsigmoid.py also uses "HardSigmoid" (mixed case), so this may be intentional per the ENN backend's expected op names. Worth confirming these are the exact strings the backend library expects.

8. Minor: op_split_with_sizes_copy.py missing return type annotation — line 26

The define_node method is missing the -> None return type annotation, unlike all other new builders.

9. Minor: op_topk.py shadows built-in sorted — line 70

sorted = cast(bool, node.args[4])

Shadows the Python built-in sorted. Not a functional bug but not great practice.

---

Positive observations

• Tests cover multiple configurations (e.g., index on different axes, split with different chunk sizes, sum with/without keepdims, topk with different dims).
• The RecomposeRmsNorm pass correctly uses get_source_partitions with torch.nn.RMSNorm source matching, which is a clean approach.
• The FlatBuffers schema change for PerformanceMode is clean and backward-compatible (default = 0).
• The @experimental decorator on PerformanceMode is a good way to signal that this feature is not yet stable.

---

Summary

The main issues to address are:

1. The pow target mismatch (issue Add support for quantized LeakyReLU #1) — verify whether the builder covers the right variant and ensure the test exercises the builder path
2. Fragile RMS norm pass (issues Re-sync with internal repository #2, Rename _pt2e to pt2e #3) — _get_eps_node/_get_gamma_node could return None and name-based matching is fragile
3. Hardcoded high-performance mode in all examples (issue Add unlifting pass under private config #4) — consider making this configurable or defaulting to DEFAULT
4. Type hint correctness (issue Re-sync with internal repository #6) — Optional[PerformanceMode]

The rest are minor style/consistency observations.

---

[SS-JIA]

Stamp to unblock -- but the agent review comments are valid. Would highly recommend fixing them before landing!

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: Jiseong-oh / name: Jiseong-oh (b1de59b, fcb7e9b)
• ✅ login: Jiseong-oh / name: jiseong.oh (451de74, 7281f62, ad35d48, bde65b1, de3bae4)
• ✅ login: jingya-zhang / name: Jingya Zhang (bde65b1, de3bae4)
• ✅ login: xiongzhan-linghu / name: xz.linghu (ad35d48)

[github-actions]

Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as Stale.
Feel free to remove the Stale label if you feel this was a mistake.
If you are unable to remove the Stale label please contact a maintainer in order to do so.
If you want the bot to never mark this PR stale again, add the no-stale label.
Stale pull requests will automatically be closed after 30 days of inactivity.

[Jiseong-oh]

/easycla