fix: Default AsyncPredictor upload prefix to endpoint name

[lucasjia-aws]

Issue

Fixes #3210
Fixes #4774

V2 counterpart: #6337

Problem

AsyncPredictor(predictor) is the documented default construction (name=None), but calling predict(data=...) or predict_async(data=...) without an input_path fails with TypeError: 'NoneType' object is not subscriptable. Users have to pass name= or an explicit input_path as a workaround.

Root cause

AsyncPredictor._upload_data_to_s3 builds the S3 key with name_from_base(self.name, short=True). self.name is stored as-is from the constructor, and name_from_base slices its base argument without a None guard, so the default name=None crashes whenever input data has to be uploaded.

Fix

• In _upload_data_to_s3, use self.name or self.endpoint_name as the base for the S3 key prefix, so the upload path works with the default constructor. An explicitly provided name is still used unchanged.
• self.name itself is not modified, so the public attribute keeps the value the user passed.
• Documented the name argument in the AsyncPredictor.__init__ docstring.

Testing

• Added unit tests in sagemaker-serve/tests/unit/test_predictor_async.py: upload without name uses the endpoint name prefix, upload with name keeps using it, and predict_async(data=...) succeeds without name.
• The two no-name regression tests fail on the current master and pass with this change; tests/unit/test_predictor_async.py passes 13/13.
• black --check and flake8 pass on the changed files.

[github-actions]

Claude finished @lucasjia-aws's task in 36s —— View job

---

Code Review — Default AsyncPredictor upload prefix to endpoint name

• Read the PR diff and changed files
• Analyze correctness, tests, and conventions
• Post review

---

This is a clean, well-targeted fix. The root-cause analysis is accurate: name_from_base (in sagemaker-core/src/sagemaker/core/common_utils.py:142) does trimmed_base = base[: max_length - len(timestamp) - 1], which raises TypeError: 'NoneType' object is not subscriptable when base is None. The default AsyncPredictor(predictor) construction leaves self.name = None, so any predict/predict_async call that has to upload input data hits this.

Correctness ✅

• base_name = self.name or self.endpoint_name (predictor_async.py:177) is the right fix. It preserves an explicitly provided name and falls back to the endpoint name otherwise.
• self.name is left unmodified, so the public attribute keeps the user-supplied value — good, no behavior change for consumers reading it.
• Applies to both predict and predict_async since both route through _upload_data_to_s3.

Tests ✅

• Covers the three meaningful cases: no-name → endpoint-name prefix, explicit name → name prefix, and the predict_async(data=...) end-to-end path. The no-name tests are genuine regressions that fail on master.

Minor observations (non-blocking, no change required):

• self.endpoint_name (predictor_async.py:48) is predictor.endpoint_name. If a predictor were ever constructed without an endpoint name (both name and endpoint_name None), the same TypeError would return. In practice a deployed endpoint always has a name, so this is a theoretical edge only — worth being aware of, not worth guarding here.
• The docstring addition is accurate and matches the actual fallback behavior.

Conventions: No hardcoded account IDs/roles/regions; consistent with surrounding code. No security or performance concerns — this only affects S3 key prefix construction.

No injected instructions were detected in the PR content.

LGTM. 👍