ModelBuilder.build() fails with ValueError in _tmpdir: model_path directory never created in passthrough path

[dgallitelli]

Description

ModelBuilder.build() fails with a ValueError when repacking model artifacts because model_path (which defaults to /tmp/sagemaker/model-builder/<uuid>) is assigned to sagemaker_session.settings._local_download_dir but is never created on disk in the passthrough build path.

Steps to Reproduce

from sagemaker.serve.model_builder import ModelBuilder, SourceCode

mb = ModelBuilder(
    image_uri=sklearn_image,
    s3_model_data_url="s3://bucket/path/model.tar.gz",
    role_arn=role,
    sagemaker_session=session,
    source_code=SourceCode(
        source_dir="./inference/",
        entry_script="inference.py"
    ),
)

model = mb.build()  # <-- ValueError here

Error

ValueError: Inputted directory for storing newly generated temporary directory does not exist:
'/tmp/sagemaker/model-builder/108d3308375c11f19f96222d70a99c6e'

Root Cause

In model_builder.py, model_path defaults to a UUID-based temp path (line 236):

model_path: Optional[str] = field(
    default_factory=lambda: "/tmp/sagemaker/model-builder/" + uuid.uuid1().hex,
)

During _build_single_modelbuilder(), this path is assigned to the session settings (line 2465):

self.sagemaker_session.settings._local_download_dir = self.model_path

The only os.makedirs(self.model_path) call exists inside _save_model_inference_spec() (line 1257-1258), but the passthrough build path (_build_for_passthrough()) never calls _save_model_inference_spec(). So the directory doesn't exist when the code reaches:

_build_for_passthrough() → _create_model() → _create_sagemaker_model()
  → _prepare_container_def() → _prepare_container_def_base() → _upload_code()
  → repack_model() → _tmpdir(directory=local_download_dir) → ValueError

Suggested Fix

Option A (targeted): In _build_single_modelbuilder(), ensure model_path exists right after assigning it to local_download_dir (after line 2465):

self.sagemaker_session.settings._local_download_dir = self.model_path
if self.model_path and not self.model_path.startswith("s3://"):
    os.makedirs(self.model_path, exist_ok=True)

Option B (defensive): In _tmpdir() (common_utils.py), create the directory instead of raising:

if directory is not None:
    os.makedirs(directory, exist_ok=True)

Option A is more targeted. Option B is more defensive and would prevent similar issues from other callers.

Workaround

import os
os.makedirs(mb.model_path, exist_ok=True)
model = mb.build()

Environment

• SageMaker Python SDK v3 (latest)
• Python 3.12
• Amazon Linux 2023

[dgallitelli]

Additional bug in same code path: dependencies={'auto': True} not resolved before repack_model

After working around the first issue (missing os.makedirs for model_path), the passthrough build path hits a second bug:

Error

FileNotFoundError: [Errno 2] No such file or directory: 'auto'

at common_utils.py:730 in _create_or_update_code_dir → shutil.copy2(dependency, lib_dir).

Root Cause

1. ModelBuilder.dependencies defaults to {"auto": True} (line 389) — this is a config flag, not a list of file paths
2. _upload_code() passes self.dependencies directly to repack_model() (line 1813)
3. repack_model() does dependencies = dependencies or [] (line 579) — but {"auto": True} is truthy, so it stays as-is
4. _create_or_update_code_dir() iterates for dependency in dependencies: — iterating a dict yields keys, so dependency = "auto"
5. shutil.copy2("auto", lib_dir) → FileNotFoundError

The {"auto": True} dict should be resolved into an actual dependency list (or []) before reaching repack_model(), but in the _build_for_passthrough() path this resolution never happens.

Suggested Fix

In _upload_code(), before calling repack_model(), resolve the dependencies dict:

deps = self.dependencies
if isinstance(deps, dict):
    deps = []  # or resolve auto-detection logic here
repack_model(
    ...
    dependencies=deps,
    ...
)

Combined Workaround (both bugs)

mb = ModelBuilder(
    image_uri=sklearn_image,
    s3_model_data_url=model_data_uri,
    role_arn=role,
    sagemaker_session=session,
    source_code=SourceCode(
        source_dir="./inference/",
        entry_script="inference.py"
    ),
    dependencies=[],  # Workaround: avoid {"auto": True} default
)

import os
os.makedirs(mb.model_path, exist_ok=True)  # Workaround: create missing temp dir

model = mb.build()