End-to-end ML pipelines — from training to serving across multiple frameworks

This project aims to demonstrate progressive MLOps maturity, inspired by Google’s MLOps maturity model.

• GitHub as the source of truth: All code, configuration, and promotion logic lives in GitHub.
• MLflow is used to log experiments (see notebooks) and track metadata such as:
  • Git commit hash
  • Model configuration
  • Artifacts for reproducible deployment
• GitHub Actions are explored to automate the promotion of models across environments (e.g. from development to staging/production) triggered via PR merges.

• 🔁 CI/CD is not yet implemented — promotion pipelines are triggered on PR merge.
• 🧪 Not full GitOps — no state reconciliation between deployed models and registry.
• 🕓 Continuous Training (CT) is planned — will enable pipeline re-training on fresh data periodically (e.g. on a schedule).
• 🧩 Feature Store support is not yet implemented (optional for this use case).

📝 A future enhancement will show how to use KitOps to manage model promotion without relying on a centralized model registry.

Explore fully functional ML deployments of the Iris dataset via different tooling and frameworks:

• app/ — FastAPI + GPU inference with PyTorch or ONNX
• data/ — raw Iris dataset / data ingestion
• deployment/ — Dockerfiles, Kubernetes-ready configs
• notebooks/ — Jupyter notebooks for training and tuning
• src/ — core model definition and training scripts
• tests/ — unit/integration tests
• torch_serve/ — TorchServe .mar archives & handler configs
• triton_model_repository/ — Triton Server model repo structure
• model.* — pre-trained model artifacts (.pt, .onnx, .compiled.pt)
• pyproject.toml, uv.lock — dependency management
• LICENSE, .gitignore, etc.

1. Install dependencies

   uv sync --frozen --no-editable --no-cache --extra torch --extra torchserve

2. Train & export models

   uv run src/train/train_iris.py

   Outputs:

   • model.pt — PyTorch weights
   • model.onnx — export via ONNX

3. Build Docker inference containers

   🚀 FastAPI + PyTorch

   docker build -f Dockerfile.torch -t fastapi-torch .
   docker run --gpus all -p 8000:8000 fastapi-torch

   Test:

   curl -X POST http://localhost:8000/predict \
     -H "Content-Type: application/json" \
     -d '{"features":[[5.1,3.5,1.4,0.2]]}'

   🚀 FastAPI + ONNX

   docker build -f Dockerfile.onnx -t fastapi-onnx .
   docker run --gpus all -p 8000:8000 fastapi-onnx

4. TorchServe Deployment

   uv run torch-model-archiver \
     --model-name iris \
     --version 1.0 \
     --serialized-file model.pt \
     --handler src/models/torch_handler.py \
     --export-path torch_serve \
     --extra-files "src/models/iris.py,src/models/torch_model_config.py,torch_serve/config.properties"
   
   docker build -f deployment/Dockerfile.torchserve -t iris-serve .
   docker run --gpus all -p 8080:8080 iris-serve
   curl -X POST http://localhost:8080/predictions/iris \
     -H "Content-Type: application/json" \
     -d '[[5.1,3.5,1.4,0.2],[6.2,3.4,5.4,2.3]]'

5. Triton Inference Server

   docker run --rm --gpus all \
     -p 8000:8000 -p 8001:8001 -p 8002:8002 \
     -v $PWD/triton_model_repository:/models \
     nvcr.io/nvidia/tritonserver:25.06-py3 \
     tritonserver --model-repository=/models

   Test:

   curl -X POST http://localhost:8000/v2/models/iris/infer \
     -H "Content-Type: application/json" \
     --data-binary @- <<EOF
   {
     "inputs":[{"name":"input","shape":[1,4],"datatype":"FP32","data":[[5.1,3.5,1.4,0.2]]}],
     "outputs":[{"name":"output"}]
   }
   EOF

• Model training and tuning with Optuna + PyTorch
• Export to TorchScript and ONNX
• GPU-powered inference via FastAPI, TorchServe, and Triton
• Containerization workflows and best practices
• How to register and serve models in diverse environments

• Run unit tests:

  uv run pytest -q

• CI/CD integrations can easily be added via GitHub Actions or other platforms, using uv sync, pytest, and Docker builds.

We welcome:

• Corrections or bug fixes
• New deployment workflows (e.g., KServe, TFKeras)
• Improvements in deployment docs or examples

Please open an issue or submit a pull request!

Apache 2.0 — see LICENSE for details.