When providing custom code for MLflow for ModelBuilder (ie with mlflow.pyfunc.PythonModel) where JSON input is used, the input_fn will fail the request as bytes are parsed to string causing errors.

[andjsmi]

Describe the bug
If I use a custom MLflow model with custom predict function and use a JSON or CSV input, the requests fail as the SageMaker Inference Toolkit converts bytes to string (source: ref) whereas the input_fn for ModelBuilder passes the input data directly into io.BytesIO (source: ref).

For scenarios where we might have a custom JSON input for a custom MLflow model, this doesn't work as JSON currently isn't supported.

To reproduce

1. Create an MLflow model with a custom Python definition with a custom predict
2. Deploy using ModelBuilder
3. Invoke model with JSON model
4. Request fails due to io.BytesIO input being a string instead of bytes

Expected behavior
Request succeeds and returns

Screenshots or logs
If applicable, add screenshots or logs to help explain your problem.

System information
A description of your system. Please provide:

• SageMaker Python SDK version: 2.237.0
• Framework name (eg. PyTorch) or algorithm (eg. KMeans): PyTorch / MLflow
• Framework version: NA
• Python version: Python 3.10.14
• CPU or GPU: CPU
• Custom Docker image (Y/N): NA

Additional context
Add any other context about the problem here.

[andjsmi]

Addressed with PR #4960