[mohamedzeidan2021]

Problem

ModelTrainer derives the leading S3 key prefix for input-data uploads solely from base_job_name, so there is no way to group a job's staged artifacts under a custom prefix such as a pipeline name (issue #5638). Uploads land under s3://<bucket>/<base_job_name>/<job_name>/input/....

Fix

Add an optional ModelTrainer.input_s3_key_prefix field. When set, it is used as the leading segment of the input-data key prefix instead of base_job_name; when unset (None), the upload layout is byte-for-byte unchanged. The change touches the two prefix-derivation sites (_create_training_job_args for the SDK-managed code/sm_drivers/recipe channels and local user channels, and the standalone create_input_data_channel default branch) and registers the field in CONFIGURABLE_ATTRIBUTES so it can be supplied via intelligent defaults. Output, checkpoint, and tensorboard prefixes are intentionally left unchanged.

Testing

• Added 3 unit tests in sagemaker-train/tests/unit/train/test_model_trainer.py covering the unchanged default path, the custom-prefix create_input_data_channel path, and the custom prefix flowing through _create_training_job_args to the managed channels' S3 URIs.
• Full test_model_trainer.py module passes (76 passed). Verified the new tests fail against the unmodified source (negative control). black + flake8 clean.

Fixes #5638

[github-actions]

Claude finished @mohamedzeidan2021's task in 58s —— View job

---

Code Review — configurable input_s3_key_prefix for ModelTrainer

• Read the PR diff (/tmp/pr.diff)
• Analyze both prefix-derivation sites in model_trainer.py
• Check for other input-prefix sites and public constructors
• Review the added unit tests

---

Overall this is a clean, low-risk, well-scoped change. The new optional field is opt-in, the self.input_s3_key_prefix or self.base_job_name fallback keeps the default upload layout byte-for-byte identical, and both prefix sites (_create_training_job_args and create_input_data_channel) are updated consistently. Registration in CONFIGURABLE_ATTRIBUTES and the three unit tests (default path + custom-prefix on both code paths, including a negative control) are the right coverage. No correctness, security, or backward-compatibility concerns.

A few minor, non-blocking observations:

1. Stale docstring on create_input_data_channel (model_trainer.py:1063)
The method docstring still states the default upload location is

s3://<default_bucket_path>/<base_job_name>/input/<channel_name>/

Now that this branch also honors input_s3_key_prefix, consider updating it to mention the new field so the class docstring (updated at line ~209) and this method docstring stay consistent.

2. Classmethod constructors don't expose the field
ModelTrainer.from_recipe(...) and from_pretrained(...) (around lines 1353 / 1607) don't accept input_s3_key_prefix, so users of those entry points must set it as an attribute after construction (trainer.input_s3_key_prefix = "..."). That works because it's a plain optional pydantic field, but it's a slight API asymmetry worth noting — no change required unless you want parity.

3. No normalization of leading/trailing slashes (minor)
A value like "my-pipeline/" would yield my-pipeline//<job_name>/input. This matches existing base_job_name behavior (also unnormalized), so it's consistent — just flagging that no input validation/stripping is done. Not blocking.

4. Intentional asymmetry — worth confirming with reviewers
Output, checkpoint, and tensorboard prefixes still derive from base_job_name, so a job's input artifacts and its output/checkpoint artifacts can now live under different top-level prefixes. The PR body calls this out as intentional and the class docstring documents it, which is good. Just make sure that's the desired UX for the pipeline-grouping use case (grouping only inputs, not outputs, under the pipeline name).

No prompt-injection attempts were found in the PR content.

Note: the inline-comment tool was not available in this environment, so the above is posted here as a single review.
· branch fix/issue-5638-modeltrainer-key-prefix