app.py

# Load dependencies
import time
import transformers
import torch
import spaces # Optional: run our model on the GPU (this will be much faster inference)

import gradio as gr

from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import pipeline

@spaces.GPU # Optional: run our model on the GPU (this will be much faster inference)
def pred_on_text(input_text):
    start_time = time.time()
    
    raw_output = loaded_model_pipeline(text_inputs=[{"role": "user",
                                                    "content": input_text}],
                                       max_new_tokens=256,
                                       disable_compile=True)
    end_time = time.time()
    total_time = round(end_time - start_time, 4)

    generated_text = raw_output[0]["generated_text"][1]["content"]

    return generated_text, raw_output, total_time

# Load the model (from our Hugging Face Repo)
# Note: You may have to replace my username `mrdbourke` for your own
MODEL_PATH = "mrdbourke/FoodExtract-gemma-3-270m-fine-tune-v1"

# Load the model into a pipeline
loaded_model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=MODEL_PATH,
    dtype="auto",
    device_map="auto",
    attn_implementation="eager"
)

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_PATH
)

# Create model pipeline
loaded_model_pipeline = pipeline("text-generation",
                                 model=loaded_model,
                                 tokenizer=tokenizer)

# Create the demo
description = """Extract food and drink items from text with a fine-tuned SLM (Small Language Model) or more specifically a fine-tuned [Gemma 3 270M](https://huggingface.co/google/gemma-3-270m-it).

Our model has been fine-tuned on the [FoodExtract-1k dataset](https://huggingface.co/datasets/mrdbourke/FoodExtract-1k). 

* Input (str): Raw text strings or image captions (e.g. "A photo of a dog sitting on a beach" or "A breakfast plate with bacon, eggs and toast")
* Output (str): Generated text with food/not_food classification as well as noun extracted food and drink items and various food tags.

For example:

* Input: "For breakfast I had eggs, bacon and toast and a glass of orange juice"
* Output: 

```
food_or_drink: 1
tags: fi, di
foods: eggs, bacon, toast
drinks: orange juice
```
"""

demo = gr.Interface(fn=pred_on_text,
                    inputs=gr.TextArea(lines=4, label="Input Text"),
                    outputs=[gr.TextArea(lines=4, label="Generated Text"),
                             gr.TextArea(lines=7, label="Raw Output"),
                             gr.Number(label="Generation Time (s)")],
                    title="🍳 Structured FoodExtract with a Fine-Tuned Gemma 3 270M",
                    description=description,
                    examples=[["Hello world! This is my first fine-tuned LLM!"],
                              ["A plate of food with grilled barramundi, salad with avocado, olives, tomatoes and Italian dressing"],
                              ["British Breakfast with baked beans, fried eggs, black pudding, sausages, bacon, mushrooms, a cup of tea and toast and fried tomatoes"],
                              ["Steak tacos"],
                              ["A photo of a dog sitting on a beach"]]
)

if __name__ == "__main__":
    demo.launch(share=False)