datamixing recipe path fix

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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

[mujtaba1747]

Description

build_hyperpod_datamix_recipe_from_context() writes the generated data-mixing
recipe to the HyperPod CLI recipes directory, then returned a path relative to
that directory with the .yaml extension stripped (a CLI --recipe reference).

Both SFTTrainer and CPTTrainer store that return value in self._recipe_path,
which get_resolved_recipe() → _load_user_recipe() later opens as a filesystem
file via os.path.isfile(). Because the value was neither absolute/CWD-relative nor
carried the .yaml extension, the file could never be found and the job failed
before submission.

Fix

Return the absolute path (with .yaml intact) of the recipe file that was just
written, so the recipe resolver can load it as a user recipe. The value's only
consumer is the resolver; the launch recipe is computed separately via
get_hyperpod_recipe_path(), so this has no impact on the CLI submission path.

Changes

• sagemaker-train/src/sagemaker/train/common_utils/data_mixing_utils.py — return
  the absolute recipe path from build_hyperpod_datamix_recipe_from_context();
  updated the docstring to describe the new return contract.
• sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py —
  updated the return-value test to assert an absolute, .yaml-suffixed path.
• sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py — aligned
  cluster/instance config with the working setup and extended the test to poll the
  S3 manifest to confirm end-to-end completion instead of passing on submission.
• Renamed test_nova_hyperpod.py → test_nova_sft_hyperpod.py.

Testing

• Unit: test_data_mixing_utils, test_data_mixing_properties, test_sft_trainer,
  and test_cpt_trainer_data_mixing suites pass (135 tests).
• End-to-end: submitted a Nova Micro LoRA SFT + data-mixing job to a HyperPod
  cluster; the job reached Succeeded, confirming the recipe now resolves and the
  job runs to completion

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