multi-algorithm HPO tuning via HyperparameterTuner.create()

[ucegbe]

Issue

Multi-algorithm HPO via HyperparameterTuner.create() is broken

When using HyperparameterTuner.create() with a model_trainer_dict containing multiple trainers, calling tuner.tune() fails with:

AttributeError: 'NoneType' object has no attribute 'training_image'

Root cause: _start_tuning_job unconditionally calls _build_training_job_definition, which accesses self.model_trainer. For multi-algo tuners created via HyperparameterTuner.create(), self.model_trainer is None — only self.model_trainer_dict is populated.

Additionally, the SageMaker CreateHyperParameterTuningJob API expects multi-algo jobs to use the TrainingJobDefinitions parameter (a list) rather than TrainingJobDefinition (singular). The existing code only ever passes the singular form.

Changes

All changes are in sagemaker-train/src/sagemaker/train/tuner.py.

_start_tuning_job

• Added detection of single vs. multi-algo mode by checking whether self.model_trainer is None and self.model_trainer_dict is populated.
• For single-algo: behavior is unchanged, passes training_job_definition (singular).
• For multi-algo: calls new _build_training_job_definitions and passes training_job_definitions (plural) in the tuning request. The underlying HyperParameterTuningJob.create in sagemaker.core.resources already accepts both parameters.

_build_training_job_definitions (new method)

• Iterates over self.model_trainer_dict and builds a list of HyperParameterTrainingJobDefinition objects, one per trainer.
• Each definition includes:
  • definition_name — the trainer key from the dict
  • tuning_objective — per-trainer objective from objective_metric_name_dict
  • hyper_parameter_ranges — per-trainer ranges from _hyperparameter_ranges_dict
  • static_hyper_parameters — per-trainer static params from static_hyperparameters_dict
  • metric_definitions — per-trainer metrics from metric_definitions_dict
  • Full passthrough of OutputDataConfig (including compression_type), ResourceConfig, StoppingCondition, environment variables, and VPC config from each ModelTrainer.
• Input data config handling supports all input types: str, dict, list[Channel], list[InputData], and includes internal ModelTrainer channels (code, sm_drivers).

Testing

Manual validation: Verified that multi-algorithm HPO tuning with HyperparameterTuner.create() using two ModelTrainer instances (XGBoost and LightGBM) successfully launches a tuning job via tuner.tune(). Single-algo tuning behavior is unchanged.

Unit tests: Added tests/unit/train/test_tuner_multi_algo.py (30 tests) covering:

• TestStartTuningJobBranching (4 tests): Verifies _start_tuning_job routes to _build_training_job_definition (singular) for single-algo and _build_training_job_definitions (plural) for multi-algo, and that the correct key (training_job_definition vs training_job_definitions) is passed in the API request.
• TestBuildTrainingJobDefinitions (20 tests): Covers the new multi-algo method — one definition per trainer, correct definition_name, per-trainer training images/objectives/HP ranges/static HPs/resource config/stopping condition/role, all input types (string, dict, InputData list, Channel list), internal channel inclusion, deduplication, metric definitions, VPC config, and environment passthrough.
• TestCompressionTypePassthrough (7 tests): Verifies compression_type (NONE, GZIP) is correctly carried through in both single-algo and multi-algo OutputDataConfig, and that MagicMock values do not leak through the isinstance guard.

All 30 new tests pass. All pre-existing tuner tests continue to pass.