app.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
hf_token = os.environ["hf_token"]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 


b_tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-1b1")#using small parameter version of model for faster inference on hf
b_model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b1",device_map = "auto")

g_tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b",token = hf_token)#using small paramerter version of model for faster inference on hf
g_model = AutoModelForCausalLM.from_pretrained("google/gemma-2-2b",token = hf_token,device_map="auto")

def Sentence_Commpletion(model_name, input):
    
    
    if model_name == "Bloom":
        tokenizer, model = b_tokenizer, b_model
        inputss = tokenizer(input, return_tensors="pt")
        outputs = model.generate(inputss.input_ids, max_new_tokens=31, num_return_sequences=1)
    elif model_name == "Gemma":
        tokenizer, model = g_tokenizer, g_model
        inputs= tokenizer(input, return_tensors="pt")
        outputs = model.generate(inputs.input_ids, max_new_tokens=32)
    return tokenizer.decode(outputs[0])


interface = gr.Interface(
fn=Sentence_Commpletion,
inputs=[gr.Radio(["Bloom", "Gemma"], label="Choose model"),
        
        gr.Textbox(placeholder="Enter sentece"),],
outputs="text",
title="Bloom vs Gemma Sentence completion",)

interface.launch(share = True, debug = True)