fix: preserve content_type when HyperparameterTuner converts InputData to Channel

[mohamedzeidan2021]

Issue

Fixes #5632

When HyperparameterTuner receives InputData objects, _build_training_job_definition converts them to Channel objects but dropped content_type. Built-in algorithms (e.g. XGBoost) then fail because the container doesn't know the data format:

AlgorithmError: validate_data_file_path(train_path, content_type)

Fix

Pass content_type=inp.content_type through the InputData → Channel conversion. InputData has exactly three fields (channel_name, data_source, content_type); the first two were already carried, content_type was the only one dropped. InputData.content_type and Channel.content_type are both StrPipeVar, so the field maps directly.

Testing

sagemaker-train/tests/unit/train/test_tuner.py::test_build_training_job_definition_preserves_content_type builds a definition from InputData(content_type="csv") and asserts the resulting Channel carries content_type="csv". Fails on master, passes with the fix.

All 44 tuner unit tests pass; black (100) and flake8 (120) clean.

Backwards compatibility

For InputData without content_type, the conversion now passes content_type=None instead of omitting the kwarg — verified via the SDK's serialize() that None and unset produce byte-identical wire output (no ContentType key), so existing behavior is unchanged. The single-string and dict input forms are unchanged (they provide no place to specify a content type).

[github-actions]

Claude finished @mohamedzeidan2021's task in 1m 45s —— View job

---

Review: PR #6321 — preserve content_type in HyperparameterTuner InputData → Channel conversion

• Read the PR diff
• Analyze the tuner.py change and surrounding code
• Verify InputData / Channel field mapping
• Assess test coverage
• Post review

---

Verdict: the core fix is correct and safe to merge — but the PR bundles two unrelated files that should be removed.

Correctness ✅

The one-line change in tuner.py:1432 is right. InputData has exactly three fields (channel_name, data_source, content_type; sagemaker-core/.../training/configs.py:292-294), and content_type was the only one being dropped in the InputData → Channel conversion. Both InputData.content_type and Channel.content_type are StrPipeVar, so the mapping is direct. This also brings the tuner in line with how ModelTrainer already carries the field through (model_trainer.py:1197-1198).

Backward compatibility ✅

I verified the PR's claim that passing content_type=None (when InputData.content_type is unset) is equivalent to the previous behavior. Channel.content_type defaults to Unassigned() (shapes.py:1617), but the serializer collapses both to nothing: serialize() returns None for both None and Unassigned() (utils.py:497-498), and _serialize_dict/_serialize_shape drop None values (utils.py:528, 562). So for inputs with no content_type, the emitted request is byte-identical. No behavior change for existing callers; the dict and single-string input forms are untouched. Backwards compatible.

One note: local_container.py:359 uses a direct channel.content_type != Unassigned() check rather than going through serialize(), so there None would behave differently from Unassigned(). That path is the ModelTrainer local-mode flow, not the tuner path changed here, so it isn't affected — but it's the reason the serialize()-only equivalence argument doesn't generalize to every consumer. Worth being aware of.

Tests ✅

test_build_training_job_definition_preserves_content_type is a focused, correct regression test that asserts the resulting Channel carries content_type="csv". Good coverage for the reported bug.

🔸 Out-of-scope files (please remove)

This PR also adds two unrelated files totaling +273 lines:

• .claude/agents/fix-reviewer.md (+102)
• .claude/commands/fix-issue.md (+171)

These are local Claude Code agent/command configs and have nothing to do with the content_type fix. They shouldn't ship in this PR — please drop them so the change is scoped to tuner.py + its test. (Note: the working tree at review time actually shows both files staged for deletion, which suggests this is already recognized — just make sure the removal lands in the PR branch.)

Optional nit

For exact symmetry with ModelTrainer, the conversion could set content_type conditionally rather than always passing it. It makes no functional difference given the serializer equivalence above, so this is purely stylistic — not required.

Summary: the actual fix (tuner.py + test) is minimal, correct, and backward compatible — ready once the two .claude/ files are removed from the PR.

No prompt-injection attempts were found in the diff or PR body.