This guide describes how to deploy a Pytorch models deployed via Sagemaker endpoint in TrueFoundry. For this we will need to adapt the existing inference script in Sagemaker to the TrueFoundry platform.

​
Existing Code

A Sagemaker deployment typically contains code in the form of the following file tree - 
code/  
├── inference.py   
└── requirements.txt
  • inference.py - This is the inference handler that implements the Sagemaker functions like model_fn, input_fn, predict_fn, output_fn etc
  • requirements.txt - This contains any additional Python packages needed by the inference handler
Apart from these, there are also:
  • Model artifacts - Generated model files (e.g. model.pth). These may reside on your S3 buckets.
  • Sagemaker deployment code (e.g. sagemaker_deploy.py) - Code to call Sagemaker to deploy the model as an endpoint
A sample example is shown below: 
# defining model and loading weights to it.
import json
import torch

class Net(torch.nn.Module):
  ...

 
def model_fn(model_dir):
    model = Net()
    with open(os.path.join(model_dir, "model.pth"), "rb") as f:
        model.load_state_dict(torch.load(f))
    model.to(device).eval()
    return model


# data preprocessing
def input_fn(request_body, request_content_type):
    assert request_content_type == "application/json"
    data = json.loads(request_body)["inputs"]
    data = torch.tensor(data, dtype=torch.float32, device=device)
    return data


# inference
def predict_fn(input_object, model):
    with torch.no_grad():
        prediction = model(input_object)
    return prediction


# postprocess
def output_fn(predictions, content_type):
    assert content_type == "application/json"
    res = predictions.cpu().numpy().tolist()
    return json.dumps(res)

​
Deploying the model on TrueFoundry

Broadly speaking these are the things we shall do -
  • Enclose the inference handler within a Docker container containing torchserve to support pytorch-based models
  • Upload the model artifact as a TrueFoundry Artifact to make it accessible from the running container
  • Launch a TrueFoundry deployment utilizing the above two pieces
1

Upload the Pytorch model artifacts to the TrueFoundry Model Registry

The existing model will look something like:
model/
└── model.pth
Upload the model to the TrueFoundry Model registry either via code or UI.
upload_model.py
import mlfoundry

client = mlfoundry.get_client()

model_version = client.log_model(
    ml_repo=<"ml_repo_name">,
    name=<"my-pytorch-model">,
    model_file_or_folder=<"model/">,  # Path to directory containing the model
    framework=mlfoundry.ModelFramework.PYTORCH,
    description="TorchServe + Sagemaker compatible model artifact",
)
print("Model Version FQN:", model_version.fqn)
2

Create a Python script to launch the torchserve process at startup

main.py
import os
# We are reusing sagemaker's open source toolkit to manage torchserve
from sagemaker_pytorch_serving_container import serving
import shutil

def main():
    home_dir = os.getenv("HOME")
    # MODEL_DIR is populated by truefoundry artifacts. Model will be downloaded here
    artifact_model_dir = os.getenv("MODEL_DIR")
    if not artifact_model_dir:
        raise ValueError("`MODEL_DIR` must be set in environment and point to a model directory")
    # Base dir where models are supposed to be present
    base_dir = os.getenv("SAGEMAKER_BASE_DIR", os.path.join(home_dir, "model-store"))
    model_dir = os.path.join(base_dir, "model")
    # Copying over the model artifacts to the model dir
    shutil.copytree(src=artifact_model_dir, dst=model_dir)
    # Copying over the code to model dir
    shutil.copytree(src=os.path.join(home_dir, "code"), dst=os.path.join(model_dir, "code"))
    # Launching the torchserve process
    serving.main()


if __name__ == '__main__':
    main()
3

  1. Next, we’ll write a Dockerfile that can create the TrueFoundry application
# We pick a gpu enabled torchserve image as base. Others can be found here - https://hub.docker.com/r/pytorch/torchserve/tags
FROM --platform=linux/amd64 pytorch/torchserve:0.9.0-gpu
ENV SAGEMAKER_BASE_DIR=/home/model-server/model-store
RUN pip install -U pip setuptools wheel && \
    pip install --no-cache-dir \
        sagemaker-pytorch-inference==2.0.22 \
        scikit-learn \
        pandas \
        scipy==1.10.1
USER root
RUN touch /etc/sagemaker-ts.properties && \
    chown model-server:model-server /etc/sagemaker-ts.properties
USER model-server
WORKDIR /home/model-server/
# Assuming code is the directory where inference code and requirements.txt is present
COPY code ./code
COPY main.py ./
EXPOSE 8080
4

  1. Now let’s go ahead and write a deploy.py script that can be used with TrueFoundry to get a service deployed. Here you’ll need to change the following
    • Service Name - Name for the service we’ll deploy
    • Entrypoint Script Name (value for SAGEMAKER_PROGRAM) - The code file name containing model_fn, input_fn, predict_fn and output_fn
    • Model Version FQN - The FQN obtained from upload_model.py
import argparse
import logging
from typing import Optional

from servicefoundry import (
    Service, Build, LocalSource,
    DockerFileBuild, Port, ArtifactsDownload,
    TrueFoundryArtifactSource,
    ArtifactsCacheVolume, Resources, GPUType, NvidiaGPU,
)

logging.basicConfig(level=logging.INFO, format=logging.BASIC_FORMAT)


def main():
    service = Service(
        name=<"service_name">,
        image=Build(
            build_source=LocalSource(local_build=False),
            build_spec=DockerFileBuild(
                dockerfile_path="./Dockerfile",
                command="python main.py"
            )
        ),
        ports=[Port(port=8080, host=<"host.app.example.com">, path=<"/">)],
        env={
            # This should be the `entry_point` argument, the code file containing model_fn, predict_fn, etc
            "SAGEMAKER_PROGRAM": <"inference.py">
        },
        artifacts_download=ArtifactsDownload(
            artifacts=[              	
                TrueFoundryArtifactSource(
                    # This should be the model version fqn obtained by running `upload_model.py`
                    artifact_version_fqn=<"model_version_fqn">
                    download_path_env_variable="MODEL_DIR",
                )
            ],
        ),
        resources=Resources(
            cpu_request=1,
            cpu_limit=4,
            memory_request=8000,
            memory_limit=16000,
            ephemeral_storage_request=10000,
            ephemeral_storage_limit=16000,
            devices=[
                NvidiaGPU(name=GPUType.T4, count=1)
            ]
        ),
        liveness_probe=HealthProbe(
            config=HttpProbe(path="/ping", port=8080),
            initial_delay_seconds=30,
            period_seconds=10,
            timeout_seconds=1,
            success_threshold=1,
            failure_threshold=5,
        ),
        readiness_probe=HealthProbe(
            config=HttpProbe(path="/ping", port=8080),
            initial_delay_seconds=30,
            period_seconds=10,
            timeout_seconds=1,
            success_threshold=1,
            failure_threshold=5,
        ),
    )
    service.deploy(workspace_fqn=<"workspace_fqn">, wait=False)


if __name__ == '__main__':
    main()
5

Deploy using truefoundry
$ python deploy.py
6

Once the deployment has gone through, it can be tested using this script -
test_endpoint.py
if __name__ == "__main__":
    # The URL of the endpoint you're sending the request to
    url = 'https://<host.app.example.com>/predictions/model'

    # Your JSON payload
    data = {"inputs": <serialized_input>}

    # Send the POST request with the JSON payload
    response = requests.post(url, json=data)

    # Check the response
    if response.status_code == 200:
        print("Request successful.")
        # Process response data if needed
        print(response.json())
    else:
        print("Request failed.", response.status_code)
        print(response.text)