fix(serve): repack source_code for image_uri/ModelTrainer builds

[lucasjia-aws]

ModelBuilder classified any image_uri build with a model artifact and source_code (no model/inference_spec) as passthrough and dropped the inference code, so build() produced a model.tar.gz without code/ and register() yielded a package that could not serve. This regressed the classic v2 Model(model_data, entry_point, source_dir) repack behavior.

In _build_for_passthrough(), distinguish a pure image-only deployment from an "image + model artifact + source_code" build. For the latter, keep the source code and bridge model_path to s3_model_data_url so the existing _upload_code(repack=True) path repacks the code into the model artifact, producing a self-contained tarball. Pure-image (Nova/BYOC) passthrough behavior is unchanged. ModelTrainer builds are covered too since they are normalized to model_path before validation.

Also pass script_dependencies (a list derived from SourceCode) to repack_model instead of the deprecated self.dependencies auto-detect dict, which repack_model would otherwise iterate as filesystem paths.

Issue #, if available:

Description of changes:

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

[lucasjia-aws]

related to issue #6105

[github-actions]

Claude finished @lucasjia-aws's task in 6m 29s —— View job

---

Code Review — PR #6112: repack source_code for image_uri/ModelTrainer builds

• Read PR diff (/tmp/pr.diff)
• Trace passthrough → _create_model → _prepare_container_def(_base) → _upload_code(repack=True) paths
• Analyze correctness, best practices, tests
• Post review

---

Overall this is a well-reasoned, well-commented fix. The logic is sound: bridging model_path → s3_model_data_url, clearing s3_upload_path, and letting is_repack()/_upload_code(repack=True) produce a self-contained model.tar.gz correctly restores the classic v2 repack behavior, and the model-data resolution (repacked_model_data or s3_upload_path or s3_model_data_url at model_builder.py:2437 / :2588) picks up repacked_model_data as intended. The script_dependencies fix is also correct — _upload_code's non-repack branch already relies on self.script_dependencies (model_builder.py:2280), and self.dependencies is the deprecated auto-detect dict ({"auto": ...}) that repack_model would wrongly iterate as filesystem paths.

A few points worth addressing:

1. No tests added for the new behavior (most important)

All 46 additions are in model_builder.py; there are no test changes. This PR fixes a regression and adds a new decision branch in _build_for_passthrough() plus changes repack dependency handling, yet nothing guards against re-regression. There's already a good home for these in sagemaker-serve/tests/unit/test_model_builder_methods.py (test_build_for_passthrough_*). Please add coverage for:

• image_uri + s3 model_path + source_code → asserts _upload_code(repack=True) runs and repacked_model_data drives ModelDataUrl (code ends up under code/).
• model_path → s3_model_data_url bridging and s3_upload_path reset to None.
• pure image-only passthrough (Nova/BYOC) → unchanged (regression guard for the existing behavior at model_builder.py:1920-1929).
• repack_dependencies → repack_model is called with the script_dependencies list, not self.dependencies.

2. has_source_code requires both entry_point and source_dir (model_builder.py:1894-1896)

_initialize_script_mode_variables can set entry_point while leaving source_dir=None (model_builder.py:1659-1662) when SourceCode.source_dir is unset but entry_script is a standalone file. In that case has_source_code is False, so this new branch is skipped and the fall-through path at :1926-1927 silently nulls entry_point/source_dir — dropping the inference code, i.e. the exact regression this PR fixes, just for the single-file layout. This does mirror is_repack() (:2246), so it may be an acceptable known limitation, but if single-file entry points are a supported input it's worth handling or at least logging a warning rather than silently dropping code.

3. Minor: inconsistent str() guarding

model_builder.py:1887 uses str(self.model_path).startswith("s3://") while the pre-existing fall-through at :1920 uses self.model_path.startswith(...). Both are guarded by a truthiness check so neither errors, but the defensive str() in the new code suggests model_path might sometimes be non-str — if that's a real concern it applies to :1920 too; if not, the str() is unnecessary. Not blocking.

Nothing here is a correctness blocker for the happy path — the primary ask is test coverage (#1). Points #2 and #3 are for your judgment.

Note: the inline-comment tool wasn't available in this session, so I've consolidated the review here with file:line references instead of inline anchors.

[github-actions]

Claude finished @lucasjia-aws's task in 8m 52s —— View job

---

Code Review — PR #6112

• Read PR diff (/tmp/pr.diff)
• Trace passthrough classification, _build_for_model_server validation, and the repack=True path
• Analyze correctness, best practices, tests
• Post review

---

The current diff takes a different approach than the PR description. The body describes bridging model_path → s3_model_data_url inside _build_for_passthrough(), but this diff doesn't touch that method. Instead it (1) excludes source_code from passthrough classification (model_builder.py:1854, :1865) so a source_code build falls through to the real builders, and (2) relaxes the _build_for_model_server validation to accept s3_model_data_url (model_builder_servers.py:105-114), plus the script_dependencies fix. The mechanics of what's here are sound, but please reconcile the description with the code.

1. Validation + repack key on s3_model_data_url, but ModelTrainer/model_path sets model_path — the scenario the body claims to cover may still fail (most important)

The PR body states "ModelTrainer builds are covered too since they are normalized to model_path before validation." But the new validation checks s3_model_data_url, not model_path:

if (not self.model and not mlflow_path and not self.inference_spec
        and not self.s3_model_data_url):
    raise ValueError(...)

For a ModelTrainer (or TrainingJob), build() sets self.model_path and nulls self.model (model_builder.py:3021-3041) — it does not populate self.s3_model_data_url. I searched the module and found no path that copies model_path → s3_model_data_url (it's only ever read from the user param / serve_settings, or reset to None). So:

• ModelTrainer + source_code + model_server → reaches _build_for_model_server, all four conditions falsy → still raises ValueError.
• Even if it passed, the repack itself gates on s3_model_data_url (_upload_code, model_builder.py:2249), so with only model_path set, repack=True would skip repacking and no code/ would land in the tarball.

Net: the fix only actually works when the user passes s3_model_data_url= explicitly. The model_path/ModelTrainer regression named in the body doesn't appear to be covered by this diff. Please confirm with an end-to-end test, or bridge model_path → s3_model_data_url (as the description says) so both inputs work.

2. Behavior change for non-1P image_uri + source_code with no model_server

With source_code now excluded from passthrough, a non-1P image_uri + source_code build with no model_server set falls through to _passthrough = False and then hits the existing guard at model_builder.py:1871-1875 ("Model_server must be set when non-first-party image_uri is set"). Previously this was treated as passthrough (image-only, code silently dropped). Erroring is arguably the correct new behavior, but it is a behavior change from "silently deploys image-only" to "raises" — worth a changelog note and a test so it's intentional, not incidental.

3. Test coverage is only negative-path (still the gap flagged in the prior review)

All six test additions just add mock_builder.s3_model_data_url = None to keep the existing "missing required params" assertions passing. There is no positive coverage for the new behavior:

• s3_model_data_url set (no model/mlflow/inference_spec) → validation passes and routes to the correct server builder.
• source_code present → passthrough is not selected (_passthrough = False) — regression guard for model_builder.py:1854/:1865.
• repack=True calls repack_model with the script_dependencies list rather than self.dependencies (the deprecated auto-detect dict) — guard for model_builder.py:2289/2305.

Please add at least these three; #1 above especially needs an end-to-end assertion that code ends up under code/.

Minor

• script_dependencies fix (model_builder.py:2287-2289, :2305) is correct — repack_model treats dependencies as a list of filesystem paths (sagemaker-core/.../common_utils.py:567), so passing the {"auto": ...} dict would have been wrong. Good catch.
• The two passthrough branches (:1848, :1859) are now identical except for the is_1p_image_uri sign; not new here, but they could be collapsed.

Nothing here is a happy-path blocker for the explicit s3_model_data_url case, which is implemented correctly. The main asks are: reconcile the description, verify/close the model_path/ModelTrainer gap (#1), and add positive-path tests (#3).

Note: the inline-comment tool wasn't available in this session, so I've consolidated the review here with file:line references.

[github-actions]

Code Review — PR #6112

• Read PR diff (/tmp/pr.diff)
• Trace passthrough → _create_model → _upload_code(repack=True) paths
• Analyze correctness, tests
• Post review

View job run