Iris Model Development and Deployment Guide

Local Development

Model Training

Train and save model on iris dataset:

python dev.model.py > ./local_artifacts/train_log.txt 2>&1

Model Inference

Run inference tests on iris dataset and custom inputs:

python dev.inference.py > ./local_artifacts/inference_log.txt 2>&1

Local Testing

FastAPI Server

Start the server:

uvicorn serve.inference_api:app --reload

Sample API Requests

Predict Setosa

curl -X POST "http://localhost:8000/predict" \
-H "Content-Type: application/json" \
-d '{
    "sepal_length": 5.1,
    "sepal_width": 3.5,
    "petal_length": 1.4,
    "petal_width": 0.2
}'

Docker Testing

sudo docker build -t iris .
sudo docker run -d -p 8000:8000 iris

Use the same curl requests as above for testing.

Fly.io Deployment

Deployment Commands

CPU Version:

flyctl deploy --remote-only --config cpu.fly.toml --dockerfile ./Dockerfile.cpu

GPU Version:

flyctl deploy --remote-only --config gpu.fly.toml --dockerfile ./Dockerfile.gpu

API Testing

Health Check:

curl -X GET "https://good-old-iris-model.fly.dev/env" \
-H "Content-Type: application/json"

Expected Response:

{ "env_hf_token": true, "env_hf_repo": true, "is_gpu_available": false }

Model Inference:

Predict Virginica

curl -X POST "https://good-old-iris-model.fly.dev/predict" \
-H "Content-Type: application/json" \
-d '{
    "sepal_length": 7.7,
    "sepal_width": 3.8,
    "petal_length": 6.7,
    "petal_width": 2.2
}'

Predict Setosa

curl -X POST "https://good-old-iris-model.fly.dev/predict" \
-H "Content-Type: application/json" \
-d '{
    "sepal_length": 5.1,
    "sepal_width": 3.5,
    "petal_length": 1.4,
    "petal_width": 0.2
}'

Expected Response example:

{
  "predicted_class": 2,
  "predicted_class_name": "virginica",
  "confidence": 0.54,
  "probabilities": {
    "setosa": 0.1,
    "versicolor": 0.36,
    "virginica": 0.54
  }
}

Known Issues

1. GPU deployment requires manual account review from Fly.io:

✖ Failed: error creating a new machine: failed to launch VM: Your organization is not allowed to use GPU machines. Please contact billing@fly.io
Please contact billing@fly.io (Request ID: 01JGHZ65QWG5FNV757MFYJTBQ7-iad) (Trace ID: 7c92684e1b16fd263ee75b2b5b34e7e9)

See forum discussion.

2. CI/CD is functional for both CPU and GPU versions, but GPU machine creation fails due to Fly.io restrictions.

References

1. Fly.io Resource Pricing
2. Fly.io Continuous Deployment with Github Actions
3. Fly.io GPU Quickstart