[ml service] Refactor: Decompose ml_single_open_custom function for better maintainability

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.

[songgot]

This commit refactors the monolithic ml_single_open_custom function (200+ lines) into smaller, focused helper functions to improve code readability and maintainability. The main function is now structured as a clear sequence of steps, each handled by a dedicated helper function.

[hj210]

This patch refactors a 200+ line function into five smaller functions,
improving readability and maintainability. 👍

Additional recommendations:
a. Avoid the '__' prefix—it’s reserved for system/compiler use.
b. Split _configure_nnfw_tensors by NNFW types to adhere to SRP.

[myungjoo]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

[songgot]

From cursor. (both look legit)

1. c/src/ml-api-inference-single.c:L1154

_configure_nnfw_tensors()가 info->nnfw만 참조합니다. 그런데 직전에 호출한 _validate_and_determine_nnfw()는 검증 결과를 info->nnfw에 다시 써주지 않아서, 사용자가 ML_NNFW_TYPE_ANY(또는 잘못된 값)을 넘기면 여기 switch가 default로 빠집니다. TensorFlow/NNFW 등에서는 in/out 텐서 메타데이터 설정이 필수인데 이 단계가 생략돼 런타임 오류가 발생합니다. 검증된 nnfw를 info->nnfw에 저장하거나 이 함수가 nnfw 인자를 직접 받아 사용하도록 조정이 필요해 보입니다.

2. c/src/ml-api-inference-single.c:L1181

framework_name/framework 옵션을 통해 info->fw_name에 값을 넣을 수 있지만, _configure_filter_properties()에서는 항상 fw_name 인자(자동 감지 결과)만 사용합니다. 그래서 사용자가 명시적으로 프레임워크 문자열을 지정해도 더 이상 반영되지 않아 기존 기능이 회귀된 상태입니다. info->fw_name이 존재하면 이를 우선 사용하도록 복원해 주세요.

Thank you for your review.
I will properly revise this PR after completing some minor refactoring.

[myungjoo-bot]

Automated review (transcribed from an AI review agent's report; please verify before acting).

Summary: The PR splits ml_single_open_custom() in c/src/ml-api-inference-single.c into eight static helpers, renames __setup_in_out_tensors -> _setup_in_out_tensors, and reformats type * var -> type *var across ~25 unrelated hunks. I traced every path of the original (validation order, goto error -> ml_single_close, list_models / hw_name / converted_models release, *single = NULL init, return codes, NNTR_INF block) against the decomposed version. Cleanup order and resource handling are preserved, but two data-flow changes alter behavior, and the PR conflicts with current main.

1. [High] Resolved nnfw type not used for tensors setup — _configure_nnfw_tensors (PR ~line 1154, called at ~1283) switches on info->nnfw (user value) instead of the local nnfw rewritten by _ml_validate_model_file() for ML_NNFW_TYPE_ANY. With ML_NNFW_TYPE_ANY and a .pb / .pt model the switch hits default, the mandatory in/out info push is skipped, and open fails with ML_ERROR_STREAMS_PIPE even when valid infos were supplied (previously: success, or a clear ML_ERROR_INVALID_PARAMETER). ml_single_create_handle (nnfw) and fw_name do use the resolved value, so handle and tensors config disagree. Fix: pass nnfw as a parameter and switch (nnfw) (do not write back into caller-owned info).
2. [High] framework_name option silently ignored — _configure_filter_properties (~1181): the original used info->fw_name when set, else _ml_get_nnfw_subplugin_name (nnfw). The PR always passes the auto-derived name, so the explicit framework_name from ml_single_open_with_option() (still populated at ~1412-1414) is dead. Public-API regression. Fix: const char *framework = info->fw_name ? info->fw_name : fw_name; in g_object_set.
3. [High] Merge conflict with main — git merge-tree reports a content conflict in this file. Since the PR base, main added "latency", info->latency_mode to the g_object_set call (507be1d; would be dropped by a naive rebase into _configure_filter_properties), rewrote __setup_in_out_tensors (d53de78), and changed ml_single_close (472be75). Also, your own open #661 replaces exactly the _configure_general/armnn/nnfw_tensors region; the two PRs conflict with each other and should be rebased into one series.
4. [Medium] Existing tests cannot catch #1 or #2 — the ML_NNFW_TYPE_ANY tests use tflite (falls into default in both versions), and the only framework_name test sets the name equal to the auto-detected one. Please add (a) under ENABLE_TENSORFLOW, ml_single_open of the .pb model with ML_NNFW_TYPE_ANY + in/out infos expecting ML_ERROR_NONE, and (b) an ml_option test with a framework_name different from the default subplugin name, asserting via ml_single_get_property (single, "framework", ...).
5. [Medium] Android CI failed on all four ABIs (logs expired). _ml_convert_predefined_entity has an __ANDROID__-only branch and is now called twice (see #7). Please re-run after rebase and confirm green.
6. [Low]_configure_handle (~1045) is a single assignment that can only return ML_ERROR_NONE, yet the caller (~1277-1280) has a dead goto error. Fold it into _configure_async_settings (as void) or inline both.
7. [Low]_ml_convert_predefined_entity (info->models) is now called twice (~1010 and ~1178). Convert once in ml_single_open_custom and pass the result to both helpers.
8. [Low] Unrelated churn — 25 of 30 hunks are type * var -> type *var reformatting in untouched functions, contradicting the repo-wide gst-indent convention (.github/workflows/static.check.scripts/indent.sh) and being the main cause of the conflict. The __ -> _ rename is also partial (__invoke, __process_output, __ml_validate_model_file untouched). Please drop the reformatting / rename from this PR.
9. [Low] Doxygen — remove the "Single Responsibility Principle" editorializing in _configure_nnfw_tensors; document @param[out] determined_nnfw / @return in _validate_and_determine_nnfw; char *hw_name in _configure_filter_properties should be const char *.

No back-door or suspicious behavior found.