app.py

import gradio as gr

num_sequences=4

demo_mode = False
if not demo_mode:
    from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

    model = AutoModelForCausalLM.from_pretrained("d2weber/german-gpt2-finetuned-coldmirror-hpodcast1")
    tokenizer = AutoTokenizer.from_pretrained("dbmdz/german-gpt2", use_fast=True)
    lm = pipeline("text-generation", model=model, tokenizer=tokenizer)

    def generate(*args, **kwargs):
        return [o["generated_text"] for o in lm(*args, **kwargs, pad_token_id=tokenizer.eos_token_id)]

with gr.Blocks() as app:
    prompt = gr.TextArea(value="Hallo und herzlich willkommen", label="Input")
    sequences = []
    for _ in range(num_sequences):
        seq = gr.Textbox("", visible=False)
        box = gr.CheckboxGroup(choices=[], label="", interactive=True)
        sequences.append(seq)

        @seq.change(inputs=seq, outputs=box)
        def split(seq):
            return gr.CheckboxGroup(seq.split(), value=[])

        @box.select(inputs=[prompt, seq], outputs=prompt)
        def handle(prompt, sequence, selected: gr.SelectData):
            to_append = " ".join(sequence.split()[:selected.index+1])
            delimiter = " " if to_append[:1].isalnum() else ""
            return prompt.rstrip() + delimiter + to_append
    
    max_new_tokens = gr.Slider(1, 100, value=18, step=1, label="How long should the generated sequences be:")

    gr.Examples([
        ["Hallo und herzlich willkommen"],
    ], prompt)

    @prompt.change(inputs=[prompt, max_new_tokens], outputs=sequences)
    def handle(prompt, max_new_tokens):
        prompt = prompt.rstrip()
        texts = ["some new words"]*num_sequences if demo_mode else generate(
            prompt,
            return_full_text=False,
            num_return_sequences=num_sequences,
            max_new_tokens=int(max_new_tokens),
        )
        return texts

app.launch()