app.py

import gradio as gr
import logging
import os
import torch
import transformers
from transformers import AutoTokenizer

logging.basicConfig(level=logging.INFO)
if torch.cuda.is_available():
    logging.info("Running on GPU")
else:
    logging.info("Running on CPU")

# Language model
if "googleflan" == os.environ.get("MODEL"):
    ### Fast/small model used to debug UI on local machine
    model = "google/flan-t5-small"
    pipeline = transformers.pipeline("text2text-generation", model=model)
    def model_func(input_text, request: gr.Request):
        return pipeline(input_text)
elif "summary_bart" == os.environ.get("MODEL"):
    model="facebook/bart-large-cnn"
    summarizer = transformers.pipeline("summarization", model=model)
    def model_func(input_text, request: gr.Request):
        return summarizer(input_text, max_length=130, min_length=30, do_sample=False)[0]["summary_text"]
elif "llama" == os.environ.get("MODEL"):
    ### Works on CPU but runtime is > 4 minutes
    model = "meta-llama/Llama-2-7b-chat-hf"
    tokenizer = AutoTokenizer.from_pretrained(
        model,
        token=os.environ.get("HF_TOKEN"),
    )
    pipeline = transformers.pipeline(
        "text-generation",
        model=model,
        torch_dtype=torch.float32,
        device_map="auto",
        token=os.environ.get("HF_TOKEN"),
    )

    def model_func(input_text, request: gr.Request):
        sequences = pipeline(
            input_text,
            do_sample=True,
            top_k=10,
            num_return_sequences=1,
            eos_token_id=tokenizer.eos_token_id,
            max_length=200,
        )
        if "name" in list(request.query_params):
            output_text = f"{request.query_params['name']}:\n"
        else:
            output_text = ""
        for seq in sequences:
            output_text += seq["generated_text"] + "\n"
        return output_text

# UI: Gradio
input_label = "How can I help?"
if "summary" in os.environ.get("MODEL"):
    input_label = "Enter text to summarize"
demo = gr.Interface(
    fn=model_func, 
    inputs=gr.Textbox(
            label=input_label,
            lines=3,
            value="",
    ),
    outputs=gr.Textbox(
            label=f"Model: {model}",
            lines=5,
            value="",
    ),
    allow_flagging=False, 
    theme=gr.themes.Default(primary_hue="blue", secondary_hue="pink")
)
    
demo.launch(server_name="0.0.0.0", server_port=7860)