bug: ModelBuilder overwrites user-provided HF_MODEL_ID for DJL Serving, preventing S3 model loading

[michaelraskansky]

Description

When using ModelBuilder with ModelServer.DJL_SERVING, the SDK unconditionally overwrites HF_MODEL_ID with the model parameter value. This prevents users from loading models from S3, even when explicitly setting HF_MODEL_ID in env_vars.

Expected Behavior

Users should be able to specify an S3 URI for model loading via env_vars:

env_vars={"HF_MODEL_ID": "s3://bucket/model/"}

Per DJL documentation, option.model_id accepts both HuggingFace Hub IDs and S3 URIs.

Actual Behavior

The SDK ignores user-provided values and always sets HF_MODEL_ID to the HuggingFace model ID from the model parameter.

Root Cause

In sagemaker-serve/src/sagemaker/serve/model_builder_servers.py, multiple build methods use:

self.env_vars.update({"HF_MODEL_ID": self.model})

This unconditionally overwrites any user-provided HF_MODEL_ID. Affected lines: 139, 215, 323, 429, 535.

Suggested Fix

Use setdefault() to preserve user-provided values:

self.env_vars.setdefault("HF_MODEL_ID", self.model)

Reproduction

from sagemaker.serve.model_builder import ModelBuilder
from sagemaker.serve.utils.types import ModelServer
from sagemaker.serve.builder.schema_builder import SchemaBuilder
from sagemaker.core.image_uris import retrieve as get_image_uri

S3_PATH = 's3://my-bucket/models/Qwen/'

mb = ModelBuilder(
    model='Qwen/Qwen3-VL-4B-Instruct',
    image_uri=get_image_uri(framework='djl-lmi', version='0.30.0', region='us-west-2'),
    model_server=ModelServer.DJL_SERVING,
    schema_builder=SchemaBuilder({'inputs': 'Hello', 'parameters': {}}, [{'generated_text': 'Hi'}]),
    role_arn='arn:aws:iam::123456789012:role/SageMakerRole',
    env_vars={'HF_MODEL_ID': S3_PATH},
    instance_type='ml.g5.2xlarge',
)

print('Before build:', mb.env_vars.get('HF_MODEL_ID'))  # s3://my-bucket/models/Qwen/
mb.build()
print('After build:', mb.env_vars.get('HF_MODEL_ID'))   # Qwen/Qwen3-VL-4B-Instruct (WRONG)

Use Case

Loading models from S3 is essential for:

• Pre-synced models to avoid runtime download latency
• Models with architectures not supported by TGI (e.g., qwen3_vl) that require vLLM backend

Environment

• SageMaker Python SDK: 3.4.0 (also affects 3.3.x)
• Python: 3.12
• DJL LMI container: 0.30.0

I'm happy to submit a PR with the fix.