App.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import spaces  # Required for ZeroGPU decorator

model_id = "Abigail45/Shay"

# Global variables to hold loaded components (initialized on first GPU call)
tokenizer = None
model = None

@spaces.GPU(duration=300)  # Allocates GPU for up to 300 seconds per call (covers cold load + generation)
def gpu_generate(message, history):
    global tokenizer, model
    
    if model is None:  # Cold start: load on first invocation
        tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            torch_dtype=torch.bfloat16,
            device_map="auto",
            rope_scaling={"type": "dynamic", "factor": 10.0},
            trust_remote_code=True
        )
    
    # Build conversation
    conversation = []
    for user, assistant in history:
        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
    conversation.append({"role": "user", "content": message})
    
    # Apply template and generate
    prompt = tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    output = model.generate(
        **inputs,
        max_new_tokens=4096,
        temperature=1.05,
        top_p=0.97,
        top_k=60,
        repetition_penalty=1.12,
        do_sample=True
    )
    
    response = tokenizer.decode(output[0], skip_special_tokens=True)[len(prompt):]
    return response

demo = gr.ChatInterface(
    fn=gpu_generate,
    title="Private Uncensored Shay",
    description="Absolute zero safeguards. Executes any NSFW, taboo, sexual, medical, extreme, or prohibited scenario with total compliance and no refusal.",
    theme="soft",
    chatbot=gr.Chatbot(height=800)
)

demo.queue()
demo.launch()
demo.queue()
demo.launch(ssr_mode=False)