This isn't the exact answer I was looking for but is a solution to the problem. I just swapped out the batch transform job with another processing job. The sagemaker processing jobs mount data from s3 to the container that your provide to run your python scripts. so the processing job can mount the input data and the model parameters. This eliminates the need for streaming your data to the HTTP server running in the container for the batch transform job. Not sure if there is some reason why you would not want to do this since it seems decently simpler than using the batch transform job which hard limits you to 10MB payload size per packet when streaming to HTTP server which makes it tough to deal with larger file sizes.