app.py

import transformers
import streamlit as st

from transformers import AutoModelForCausalLM, AutoTokenizer

import spaces

checkpoint = "."
  
tokenizer = AutoTokenizer.from_pretrained(checkpoint)

@st.cache_resource
def load_model(model_name):
    model = AutoModelForCausalLM.from_pretrained(model_name)
    return model

model = load_model(checkpoint)

@spaces.GPU
def infer(input_ids, bad_words_ids, max_tokens, temperature, top_k, top_p):

    output_sequences = model.generate(
        input_ids=input_ids,
        bad_words_ids = bad_words_ids,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        do_sample=True,
        no_repeat_ngram_size=2,
        early_stopping=True,
        num_beams=4,
        pad_token_id=tokenizer.eos_token_id,
        num_return_sequences=1
    )

    return output_sequences

default_value = "We are building the first ever"
    
#prompts
st.title("Write with vcGPT 🦄")
st.write("This is a LLM that was fine-tuned on a dataset of investment memos to help you generate your next pitch.")

sent = st.text_area("Text", default_value)

max_tokens = st.sidebar.slider("Max Tokens", min_value = 16, max_value=64)
temperature = st.sidebar.slider("Temperature", value = 0.8, min_value = 0.05, max_value=1.0, step=0.05)
top_k = st.sidebar.slider("Top-k", min_value = 0, max_value=5, value = 4)
top_p = st.sidebar.slider("Top-p", min_value = 0.0, max_value=1.0, step = 0.05, value = 0.9)

# print(model.config.max_position_embeddings)

encoded_prompt = tokenizer.encode(tokenizer.eos_token+sent, max_length=1024, return_tensors="pt", truncation=True)

# get tokens of words that should not be generated
bad_words_ids = tokenizer(["confidential", "angel.co", "angellist.com", "angellist"], add_special_tokens=False).input_ids

if encoded_prompt.size()[-1] == 0:
    input_ids = None
else:
    input_ids = encoded_prompt

output_sequences = infer(input_ids, bad_words_ids, max_tokens, temperature, top_k, top_p)

for generated_sequence_idx, generated_sequence in enumerate(output_sequences):
    print(f"=== GENERATED SEQUENCE {generated_sequence_idx + 1} ===")
    generated_sequences = generated_sequence.tolist()

    # Decode text
    text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True, skip_special_tokens=True)

    # Remove all text after the stop token
    #text = text[: text.find(args.stop_token) if args.stop_token else None]

    # Add the prompt at the beginning of the sequence. Remove the excess text that was used for pre-processing
    total_sequence = (
        sent + text[len(tokenizer.decode(encoded_prompt[0], clean_up_tokenization_spaces=True, skip_special_tokens=True)) :]
    )

    generated_sequences.append(total_sequence)
    print(total_sequence)

st.markdown(generated_sequences[-1])