import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer, GPTQConfig, TrainingArguments
from threading import Thread
from peft import AutoPeftModelForCausalLM
from transformers import GenerationConfig
tokenizer = AutoTokenizer.from_pretrained("izh97/zephyr-beta-climate-change-assistant")
model = AutoPeftModelForCausalLM.from_pretrained(
    "izh97/zephyr-beta-climate-change-assistant",
    low_cpu_mem_usage=True,
    return_dict=True,
    torch_dtype=torch.float16,
    device_map="cuda")
model = model.to('cuda:0')
generation_config = GenerationConfig(
    do_sample=True,
    top_k=10,
    temperature=0.2,
    max_new_tokens=256,
    pad_token_id=tokenizer.unk_token_id
)
def ask(text):
  tokenizer = AutoTokenizer.from_pretrained("izh97/zephyr-beta-climate-change-assistant")
  model = AutoPeftModelForCausalLM.from_pretrained(
    "izh97/zephyr-beta-climate-change-assistant",
    low_cpu_mem_usage=True,
    return_dict=True,
    torch_dtype=torch.float16,
    device_map="cuda")

  inputs = tokenizer.apply_chat_template(text, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

  input_length = inputs.input_ids.shape[1]
  outputs = model.generate(**inputs, generation_config=generation_config,
                           return_dict_in_generate=True)

  tokens = outputs.sequences[0, input_length:]
  return tokenizer.decode(tokens)