app.py

import spaces
import gradio as gr

from unsloth import FastLanguageModel

max_seq_length = 2048
dtype = (
    None
)
load_in_4bit = True

model = None

@spaces.GPU
def load_model():
    if model is None:
        model, tokenizer = FastLanguageModel.from_pretrained(
            model_name="ua-l/gemma-2-9b-legal-uk",
            max_seq_length=max_seq_length,
            dtype=dtype,
            load_in_4bit=load_in_4bit,
        )

load_model()

FastLanguageModel.for_inference(model)


@spaces.GPU
def predict(question):
    inputs = tokenizer(
    [f'''### Question:
    {question}
    
    ### Answer:
    '''], return_tensors = "pt").to("cuda")

    outputs = model.generate(**inputs, max_new_tokens = 128, use_cache = True)
    
    results = tokenizer.batch_decode(outputs, skip_special_tokens=True)

    return results[0]

inputs = gr.Textbox(lines=2, label="Enter a question", value="Як отримати виплати ВПО?")

outputs = gr.JSON(label="Answer")

demo = gr.Interface(fn=predict, inputs=inputs, outputs=outputs)
demo.launch()