app.py

import torch
from transformers import BigBirdForQuestionAnswering, BigBirdTokenizerFast
import gradio as gr

FLAX_MODEL_ID = "vasudevgupta/flax-bigbird-natural-questions"

if __name__ == "__main__":
  device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")
  
  model = BigBirdForQuestionAnswering.from_pretrained(FLAX_MODEL_ID, block_size=64, num_random_blocks=3, from_flax=True)
  model.to(device)
  tokenizer = BigBirdTokenizerFast.from_pretrained(FLAX_MODEL_ID)
  
  def get_answer(question, context):
    encoding = tokenizer(question, context, return_tensors="pt", max_length=4096, padding="max_length", truncation=True)
    input_ids = encoding.input_ids.to(device)
    attention_mask = encoding.attention_mask.to(device)

    with torch.no_grad():
        start_scores, end_scores = model(input_ids=input_ids, attention_mask=attention_mask).to_tuple()

    # Let's take the most likely token using `argmax` and retrieve the answer
    all_tokens = tokenizer.convert_ids_to_tokens(encoding["input_ids"].squeeze().tolist())

    answer_tokens = all_tokens[torch.argmax(start_scores): torch.argmax(end_scores)+1]
    answer = tokenizer.decode(tokenizer.convert_tokens_to_ids(answer_tokens))

    return answer

    default_context = "BigBird Pegasus just landed! Thanks to Vasudev Gupta, BigBird Pegasus from Google AI is merged into HuggingFace Transformers. Check it out today!!!"
    question = gr.inputs.TextBox(lines=2, default="Who added BigBird to HuggingFace Transformers?", label="Question")
    context = gr.inputs.TextBox(lines=10, default=default_context, label="Context")

    gr.Interface(fn=get_answer, inputs=[question, context], outputs="text").launch()