app.py

import gradio as gr
import os
import spaces
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_path = "cognitivecomputations/dolphin-2.8-mistral-7b-v02"
tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(model_path)
model.config.pad_token_id = model.config.eos_token_id

system_prompt = f"<|im_start|>system\nYou are Santa.<|im_end|>\n"

history = system_prompt

@spaces.GPU(duration=120)
def chat(prompt):
    input_text = history + "<|im_start|>user\n" + prompt + "<|im_end|>\n"
    input_ids = tokenizer.encode(input_text, return_tensors="pt")
    attention_mask = torch.ones_like(input_ids)
    output = model.generate(
        input_ids,
        attention_mask=attention_mask,
        max_length=1024,
        num_return_sequences=1,
        top_p=0.9,
        top_k=50,
        num_beams=2,
        pad_token_id=model.config.eos_token_id
    )
    response = tokenizer.decode(output[0], skip_special_tokens=True)
    history += "<|im_start|>assistant\n" + response + "<|im_end|>\n"
    return response

demo = gr.Interface(
    fn=chat,
    inputs=gr.Textbox(placeholder="Enter your message here"),
    outputs=gr.Textbox(label="Response")
)

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