app.py

from transformers import BloomTokenizerFast, BloomModel
import torch

mname = "bigscience/bloom-1b3"
tokenizer = BloomTokenizerFast.from_pretrained(mname, use_cache=True)
model = BloomModel.from_pretrained(mname, use_cache=True)

def take_last_tokens(inputs, note_history, history):
    """Filter the last 256 tokens"""
    if inputs['input_ids'].shape[1] > 256:
        inputs['input_ids'] = torch.tensor([inputs['input_ids'][0][-256:].tolist()])
        inputs['attention_mask'] = torch.tensor([inputs['attention_mask'][0][-256:].tolist()])
        note_history = ['</s> <s>'.join(note_history[0].split('</s> <s>')[2:])]
        history = history[1:]
    return inputs, note_history, history

def add_note_to_history(note, note_history):
    """Add a note to the historical information"""
    note_history.append(note)
    note_history = '</s> <s>'.join(note_history)
    return [note_history]

def chat(message, history):
    history = history or []
    if history: 
        history_useful = ['</s> <s>'.join([str(a[0])+'</s> <s>'+str(a[1]) for a in history])]
    else:
        history_useful = []
    history_useful = add_note_to_history(message, history_useful)
    inputs = tokenizer(history_useful, return_tensors="pt")
    inputs, history_useful, history = take_last_tokens(inputs, history_useful, history)
    reply_ids = model.generate(**inputs)
    response = tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0]
    history_useful = add_note_to_history(response, history_useful)
    list_history = history_useful[0].split('</s> <s>')
    history.append((list_history[-2], list_history[-1]))
    return history, history

gr.Interface(
    fn=chat,
    theme="huggingface",
    css=".footer {display:none !important}",
    inputs=["text", "state"],
    outputs=["message", "state"],
    title="Bloom 1b3 chat",
    allow_flagging="never",
    ).launch()