[v3] FrameworkProcessor and ModelTrainer: 4 regressions (including dropping CodeArtifact support) from v2 migration

从 sagemaker/core/resources.py 开始了解会话处理，从 sagemaker-core/src/sagemaker/core/processing.py 查看 _package_code 和 _generate_framework_script，并从 sagemaker.train.templates 查看 INSTALL_REQUIREMENTS。比较 PR #4145 中 v2 的 CodeArtifact 行为与 training-toolkit 的行为。完成的标准是，四条回归路径都保留提供的会话、遵循 code_location，并通过 CodeArtifact 安装私有 requirements。

PySDK Version

• PySDK V2 (2.x)
• PySDK V3 (3.x)

Additional context

The following four issues were found during a migration from sagemaker==2.257.1 to sagemaker==3.7.1, moving training jobs from sagemaker.pytorch.PyTorch to sagemaker.train.ModelTrainer and processing jobs from sagemaker.pytorch.processing.PyTorchProcessor to sagemaker.core.processing.FrameworkProcessor.

---

Describe the bug

Four regressions found when migrating from v2 to v3, specifically around FrameworkProcessor (processing jobs) and ModelTrainer (training jobs). All four worked correctly in v2.

System information

• SageMaker Python SDK version: 3.7.1
• Framework name: PyTorch
• Framework version: 2.10
• Python version: 3.13
• CPU or GPU: Both
• Custom Docker image: N

---

Bug 1: wait=True does not respect sagemaker session

Affects: ModelTrainer.train(wait=True) and FrameworkProcessor.run(wait=True)

ProcessingJob.refresh() and TrainingJob.refresh() use Base.get_sagemaker_client() — a global/default client — instead of the sagemaker_session passed to the processor/trainer. This fails with NoCredentialsError when using assumed-role sessions (via STS).

To reproduce:

import boto3
from sagemaker.core.helper.session_helper import Session
from sagemaker.core.processing import FrameworkProcessor
from sagemaker.train import ModelTrainer
from sagemaker.core.training.configs import Compute, SourceCode

# Assumed-role session
sts = boto3.client("sts")
assumed = sts.assume_role(RoleArn="arn:aws:iam::123456789:role/MyRole", RoleSessionName="test")
creds = assumed["Credentials"]
assumed_session = boto3.Session(
    aws_access_key_id=creds["AccessKeyId"],
    aws_secret_access_key=creds["SecretAccessKey"],
    aws_session_token=creds["SessionToken"],
    region_name="us-west-2",
)
sm_session = Session(boto_session=assumed_session)

# FrameworkProcessor — job created OK, wait fails
processor = FrameworkProcessor(
    image_uri="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:2.10-cpu-py313",
    command=["python3"],
    role="arn:aws:iam::123456789:role/MyRole",
    instance_count=1,
    instance_type="ml.m5.xlarge",
    sagemaker_session=sm_session,
)
processor.run(code="my_script.py", source_dir="src", wait=True)
# → NoCredentialsError

# ModelTrainer — same issue
trainer = ModelTrainer(
    training_image="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:2.10-cpu-py313",
    role="arn:aws:iam::123456789:role/MyRole",
    source_code=SourceCode(entry_script="train.py", source_dir="src"),
    compute=Compute(instance_type="ml.m5.xlarge", instance_count=1),
    sagemaker_session=sm_session,
)
trainer.train(wait=True)
# → NoCredentialsError

Root cause (sagemaker/core/resources.py):

# ProcessingJob.refresh() / TrainingJob.refresh()
client = Base.get_sagemaker_client()  # ← ignores the session
response = client.describe_processing_job(**operation_input_args)

v2 behaviour — ProcessingJob.wait() used the session directly:

def wait(self, logs=True):
    if logs:
        self.sagemaker_session.logs_for_processing_job(self.job_name, wait=True)
    else:
        self.sagemaker_session.wait_for_processing_job(self.job_name)

---

Bug 2: FrameworkProcessor.code_location is accepted but ignored

FrameworkProcessor.__init__ accepts code_location and stores it as self.code_location. The docstring states it controls where code is uploaded. However, _package_code ignores it and always uploads to self.sagemaker_session.default_bucket().

To reproduce:

processor = FrameworkProcessor(
    image_uri="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:2.10-cpu-py313",
    command=["python3"],
    role="arn:aws:iam::123456789:role/MyRole",
    instance_count=1,
    instance_type="ml.m5.xlarge",
    code_location="s3://my-custom-bucket",  # ← ignored
)
processor.run(code="my_script.py", source_dir="src", wait=False)
# Code uploads to s3://sagemaker-us-west-2-123456789/... instead of s3://my-custom-bucket/...

Root cause (sagemaker-core/src/sagemaker/core/processing.py, _package_code):

s3_uri = s3.s3_path_join(
    "s3://",
    self.sagemaker_session.default_bucket(),  # ← always uses default bucket
    self.sagemaker_session.default_bucket_prefix or "",
    job_name, "source", "sourcedir.tar.gz",
)

self.code_location is never referenced.

v2 behaviour — FrameworkProcessor delegated to an estimator that honored code_location:

# v2 FrameworkProcessor._create_estimator
return self.estimator_cls(
    ...
    code_location=self.code_location,
    ...
)

Note: ModelTrainer does not offer code_location at all — it always uses session.default_bucket(). Suggested fix: either remove code_location from FrameworkProcessor to align with ModelTrainer, or update _package_code to use it when set.

---

Bugs 3 and 4 are both about losing the ability to install requirements.txt dependencies from a CodeArtifact repository.

1. PyTorch training and inference containers supported this via the CA_REPOSITORY_ARN environment variable (see https://github.com/aws/deep-learning-containers/issues/2509 for details)
2. https://github.com/aws/sagemaker-python-sdk/pull/4145 extended that support to processing jobs by exposing codeartifact_repo_arn on FrameworkProcessor.run()

For context, the new ray-based inference containers are also adding codeartifact support via CA_REPOSITORY_ARN environment variable as can be seen in https://github.com/aws/deep-learning-containers/blob/0fc07f317a4db68ff728274070fbe332dde1ca26/scripts/ray/sagemaker_serve.py#L68

---

Bug 3: CodeArtifact support missing from FrameworkProcessor

In v2, FrameworkProcessor.run() accepted codeartifact_repo_arn (added in PR #4145). This configured pip inside the container to authenticate with CodeArtifact before installing requirements.txt.

In v3, FrameworkProcessor.run() does not accept codeartifact_repo_arn, and _generate_framework_script has no CodeArtifact support. The generated runproc.sh runs pip install -r requirements.txt without authentication, failing for packages hosted on private CodeArtifact repositories.

To reproduce:

processor = FrameworkProcessor(
    image_uri="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:2.10-cpu-py313",
    command=["python3"],
    role="arn:aws:iam::123456789:role/MyRole",
    instance_count=1,
    instance_type="ml.m5.xlarge",
)

# v2 supported: processor.run(..., codeartifact_repo_arn="arn:aws:codeartifact:us-west-2:123:repository/domain/repo")
# v3 does not — parameter doesn't exist

processor.run(code="my_script.py", source_dir="src", wait=False)
# Container fails: pip install -r requirements.txt → package not found (private CodeArtifact repo)

v2 behaviour — _generate_framework_script injected CodeArtifact login into runproc.sh:

if [[ -f 'requirements.txt' ]]; then
    if ! hash aws 2>/dev/null; then
        echo "AWS CLI is not installed. Skipping CodeArtifact login."
    else
        aws codeartifact login --tool pip --domain {domain} --domain-owner {owner} --repository {repository} --region {region}
    fi
    pip install -r requirements.txt
fi

Suggested fix: Port codeartifact_repo_arn and _get_codeartifact_command from v2's PR #4145 into v3's FrameworkProcessor.

---

Bug 4: ModelTrainer bypasses sagemaker-training-toolkit, losing CodeArtifact support for requirements.txt

Affects: ModelTrainer.train() with SourceCode(requirements="requirements.txt")

ModelTrainer overrides the container's ENTRYPOINT with its own sm_train.sh driver script. This bypasses the sagemaker-training-toolkit installed in the container, which handled requirements.txt installation with CodeArtifact support (via the CA_REPOSITORY_ARN environment variable, added in sagemaker-training-toolkit#187).

The v3 sm_train.sh template for requirements installation is a bare pip install:

# from sagemaker.train.templates.INSTALL_REQUIREMENTS
echo "Installing requirements"
$SM_PIP_CMD install -r {requirements_file}

It does not check for CA_REPOSITORY_ARN or configure pip to use CodeArtifact. The env var is passed to the container but nothing reads it.

To reproduce:

from sagemaker.train import ModelTrainer
from sagemaker.core.training.configs import Compute, SourceCode

trainer = ModelTrainer(
    training_image="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-training:2.10-cpu-py313",
    role="arn:aws:iam::123456789:role/MyRole",
    source_code=SourceCode(
        entry_script="train.py",
        source_dir="src",
        requirements="requirements.txt",  # ← installed without CodeArtifact
    ),
    compute=Compute(instance_type="ml.m5.xlarge", instance_count=1),
    environment={"CA_REPOSITORY_ARN": "arn:aws:codeartifact:us-west-2:123:repository/domain/repo"},
)
trainer.train(wait=False)
# Container runs: pip install -r requirements.txt (using public PyPI, not CodeArtifact)
# Fails in VPC-isolated environments where PyPI is unreachable

v2 behaviour — the PyTorch estimator used the container's native entrypoint, which invoked sagemaker-training-toolkit. The toolkit checked CA_REPOSITORY_ARN, ran aws codeartifact login --tool pip, then installed requirements. This was added in sagemaker-training-toolkit v4.7.0 and tracked in deep-learning-containers#2509.

Suggested fix: The INSTALL_REQUIREMENTS template in sagemaker.train.templates should check for CA_REPOSITORY_ARN and configure pip accordingly before installing, matching the behaviour of sagemaker-training-toolkit.