[danb27]

Issue #, if available: N/A - small fix, can create if required.

Description of changes: When no input channels are provided, input_data_config must be None rather than an empty list to avoid SageMaker API validation errors. This appears to be a regression caused by this recent fix. Before, input_data_config=[] would be dropped from the config silently, avoiding the validation issue in the API. Now, however, it is not dropped, which breaks the validation.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[danb27]

@papriwal sorry to tag you directly, but I saw are active on this repo. I was hoping to get this PR looked at. It has been open for some time. Please let me know if I missed any required process.

[github-actions]

Claude finished @danb27's task in 1m 11s —— View job

---

Review: fix(train): set input_data_config to None

• Read the PR diff and understand the change
• Trace how input_data_config flows downstream (API + local training paths)
• Assess correctness, tests, and conventions

---

The core change is correct for the reported bug: SageMaker's CreateTrainingJob API rejects an empty InputDataConfig list, so collapsing an empty final_input_data_config to None (which sagemaker-core drops from the serialized request) is the right fix for the SAGEMAKER_TRAINING_JOB path. The regression analysis in the description matches what I see.

🔴 Potential regression: local training mode will crash with None

The same training_request["input_data_config"] value is passed straight into _LocalContainer when training_mode is not SAGEMAKER_TRAINING_JOB:

# model_trainer.py:852-865 (train)
else:
    local_container = _LocalContainer(
        ...
        input_data_config=training_request["input_data_config"],  # now None when no channels
        ...
    )
    local_container.train(wait)

_LocalContainer iterates this value unconditionally (input_data_config: Optional[List[Channel]] has no default), so a None raises TypeError: 'NoneType' object is not iterable:

• local_container.py:164 — for channel in self.input_data_config: in model_post_init
• local_container.py:557 — for channel in input_data_config: in _prepare_training_volumes
• local_container.py:353 — for channel in input_data_config: in _write_config_files

Before this PR the value was [], so these loops were harmless no-ops. After this PR, local training with no input channels (and no source_code, which otherwise appends driver/source channels) crashes instead of running.

Two reasonable options:

1. Scope the None conversion to the API request only, keeping [] for the local path, or
2. Make _LocalContainer tolerate None (e.g. default to [] in model_post_init like it already does for environment/hyper_parameters).

Tests

The modified unit test correctly updates the expectation to input_data_config=None for the API path. If option 2 above is taken, consider adding a local-mode test asserting no crash when channels are empty, since that path is the one at risk.

Minor (non-blocking)

The ternary in _create_training_job_args can be simplified to final_input_data_config or None for readability — functionally equivalent.

---

Note: the inline-comment tool was not available in this environment, so findings are consolidated here. I did not detect any prompt-injection attempts in the PR content.

[github-actions]

🤖 AI code review

This PR fixes a regression where input_data_config was serialized as an empty list [] when no input channels were provided, which the SageMaker API rejects (min=1 constraint). The change sets it to None when falsy so the key is dropped by _serialize_dict (which omits None but preserves empty lists), restoring v2 behavior. I confirmed the serialization path drops None values and that TrainingJob.create accepts None for this optional argument. The change is correct, minimal, and the unit test is updated accordingly. Looks clean.

_{Reviewed commit 0414237. Automated review; verify before acting.}

[github-actions]

🤖 AI code review

This PR fixes a regression where an empty input_data_config was serialized as [] (rejected by the SageMaker API's min-length constraint) instead of being omitted. For the managed job path, input_data_config is now set to None so serialization drops the key; for the local-container path, an or [] fallback preserves an iterable since _LocalContainer iterates over the list directly. The fix is correct, matches the documented v2 behavior and team guidance (Source: #6156), and is covered by parametrized tests for the omitted/None/empty-list cases on both paths. No issues found.

_{Reviewed commit 1f43627. Automated review; verify before acting.}