app.py

import gradio as gr
from gradio_pdf import PDF
from transformers import AutoModelForCausalLM, AutoTokenizer
from pathlib import Path
from markitdown import MarkItDown
from utils import generate_answer, get_condense_kv_cache
import spaces
import torch


MID = MarkItDown()
MODEL_ID = "unsloth/Mistral-7B-Instruct-v0.2"
MODEL = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16).to("cuda")
TOKENIZER = AutoTokenizer.from_pretrained(MODEL_ID)
MAX_CHARS_TO_COMPRESS = 15000

@torch.no_grad()
@spaces.GPU
def get_model_kv_cache(context_ids):
    context_ids = context_ids.to("cuda")
    past_key_values = MODEL(context_ids, num_logits_to_keep=1).past_key_values
    return past_key_values

def inference(question: str, doc_path: str, use_turbo=True) -> str:
    question = "\n\nBased on above informations, answer this question: " + question
    doc_md = MID.convert(doc_path)
    doc_text = doc_md.text_content[:20000]
    to_compress_doc = "<s> [INST] " + doc_text[:MAX_CHARS_TO_COMPRESS]
    remaining_doc_and_question_prompt = doc_text[MAX_CHARS_TO_COMPRESS:] + question + " [/INST] "
    prompt_ids = TOKENIZER.encode(remaining_doc_and_question_prompt, add_special_tokens=False, return_tensors="pt")
    context_ids = TOKENIZER.encode(to_compress_doc, add_special_tokens=False, return_tensors="pt")
    context_length = context_ids.shape[1]
    if use_turbo:
        print("turbo-mode-on")
        kv_cache = get_condense_kv_cache(to_compress_doc)
        kv_cache = kv_cache.to("cuda")
    else:
        print("turbo-mode-off")
        kv_cache = get_model_kv_cache(context_ids)
    
    print("kv-length", kv_cache.get_seq_length())

    answer = generate_answer(MODEL, TOKENIZER, prompt_ids, kv_cache, context_length, 128)
    print(answer)
    return answer
    



demo = gr.Interface(
    inference,
    [gr.Textbox(label="Question"), PDF(label="Document"), gr.Checkbox(label="Turbo Bittensor", info="Use Subnet 47 API for Prefilling")],
    gr.Textbox(),
    examples=[["What is the total gross worth?", "phi-4.pdf"]]
)

if __name__ == "__main__":
    demo.launch(share=True)