Build Generative AI Apps with Docker And Hugging Face's Docker Spaces

A practical path to ship a text-generation generative AI app is laid out end-to-end: build a FastAPI service that wraps a Hugging Face Transformers text-generation pipeline, dockerize it with a reproducible Python environment, then deploy it to Hugging Face Spaces so it runs automatically on commit. The payoff is immediate—once the Space finishes building, users get both a working API endpoint and FastAPI’s Swagger UI to test generation requests in the browser.

The workflow starts with creating a local Python environment (Python 3.9) and installing the core dependencies needed for an LLM-backed web service. The project uses FastAPI to expose HTTP routes, Transformers to access model inference, and PyTorch (torch) as the underlying compute library. A requirements.txt file is created to pin the needed packages, including FastAPI, requests, sentencepiece, torch, and Transformers. After installing dependencies in the virtual environment, the next step is writing the application code.

The FastAPI app defines a simple homepage route that returns a “hello world” message to confirm the service is alive. It also defines a /generate endpoint that accepts input text and feeds it into a Transformers pipeline configured for text-to-text generation. For the model choice, the walkthrough uses flan-t5-small (a smaller variant intended to fit free-tier resource limits on Hugging Face Spaces). The pipeline output is returned as JSON, with the generated text extracted from the pipeline’s response structure (noted as generator_text in the model’s output). FastAPI’s built-in Swagger documentation becomes the interactive control panel once deployed.

Dockerization is treated as the central engineering step. A Dockerfile is created from the official Python 3.9 image, sets a working directory, copies requirements.txt into the container, and installs dependencies using pip install with caching disabled. The Dockerfile also creates and switches to a non-root user for safer execution, then copies the application code into the container. Finally, it starts the FastAPI app via uvicorn, binding to 0.0.0.0 on a specified port (the walkthrough uses 7860).

Deployment happens through Hugging Face Spaces. A new Space is created using a Docker template, then the app.py, Dockerfile, and requirements.txt are uploaded to the Space repository. Each commit triggers an automated Docker build and container run. If the build succeeds, the homepage endpoint returns “hello world,” confirming the container is executing correctly. From there, users can open a direct URL or navigate to /docs to access Swagger UI and test /generate.

In the Swagger interface, sample prompts like “tell me about machine learning” or “how to be happy” produce completed sentences, demonstrating text-to-text generation powered by Transformers and the flan-t5-small model. The overall message is less about model novelty and more about operational reliability: containerizing the FastAPI + Transformers stack makes the generative AI app portable and deployable with minimal friction.