app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

access_token = st.secrets["HF_ACCESS_TOKEN"]

model_name = "meta-llama/Meta-Llama-3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name, token=access_token)
model = AutoModelForCausalLM.from_pretrained(model_name, token=access_token)


text_input = st.text_area("Enter text:")

if text_input:
    inputs = tokenizer(text_input, return_tensors="pt")
    input_ids = inputs.input_ids

    # Use the model to get the output logits
    with torch.no_grad():
        output = model(input_ids)

    # Extract logits (remove the batch dimension as there's only one input example)
    logits = output.logits.squeeze(0)

    # Pair tokens with their corresponding logits
    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
    token_logit_pairs = [(token, logits[idx].tolist()) for idx, token in enumerate(tokens)]
    st.json(token_logit_pairs)