import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = f"telmo000/bloom-positive-reframing"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    return_dict=True,
    load_in_8bit=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id)


def make_inference(original_text):
    str_strategy_prompt = f"### Negative sentence:\n{original_text}\n\n### Reframing strategy:\n"
    batch_1 = tokenizer(
        str_strategy_prompt,
        return_tensors="pt",
    )

    with torch.cuda.amp.autocast():
        output_tokens_1 = model.generate(**batch_1, max_new_tokens=50)

    output_1 = tokenizer.decode(output_tokens_1[0], skip_special_tokens=True)
    reframing_strategy = output_1[len(str_strategy_prompt):].partition('\n')[0]

    str_reframing_prompt = f"### Negative sentence:\n{original_text}\n\n### Reframing strategy:\n{reframing_strategy}\n\n### Reframing sentence:\n"
    batch_2 = tokenizer(
        str_reframing_prompt,
        return_tensors="pt",
    )

    with torch.cuda.amp.autocast():
        output_tokens_2 = model.generate(**batch_2, max_new_tokens=100)

    output_2 = tokenizer.decode(output_tokens_2[0], skip_special_tokens=True)
    reframing_sentence = output_2[len(str_reframing_prompt):]

    return reframing_sentence


if __name__ == "__main__":
    # make a gradio interface
    import gradio as gr

    gr.Interface(
        make_inference,
        [
            gr.inputs.Textbox(lines=3, label="Original Text"),
        ],
        gr.outputs.Textbox(label="Reframed Text"),
        title="Bloom positive reframing",
        description="Bloom positive reframing is a BLOOM-base generative model adjusted to the sentiment transfer task, where the objective is to reverse the sentiment polarity of a text without contradicting the original meaning. ",
    ).launch()