app.py

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

MODEL_ID = "EssentialAI/rnj-1-instruct"

# Load tokenizer & model once (ZeroGPU will put it on GPU when @spaces.GPU fn runs)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",   # ZeroGPU will map to H200 slice
)

SYSTEM_PROMPT = (
    "You are RNJ-1, a precise, math-friendly assistant. "
    "Solve problems step by step, avoid rambling, and give the final answer clearly."
)


def build_prompt(user_text: str) -> str:
    # Keep it simple – RNJ handles plain text well
    return f"{SYSTEM_PROMPT}\n\nUser: {user_text}\nAssistant:"


@spaces.GPU  # <-- THIS is what ZeroGPU wants
def generate_reply(
    user_text: str,
    max_new_tokens: int,
    temperature: float,
) -> str:
    prompt = build_prompt(user_text)

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=int(max_new_tokens),
            do_sample=temperature > 0.0,
            temperature=float(temperature),
            top_p=0.9,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    full = tokenizer.decode(output[0], skip_special_tokens=True)

    # Strip everything before the last "Assistant:" so prompt text doesn't echo back
    if "Assistant:" in full:
        reply = full.split("Assistant:")[-1].strip()
    else:
        reply = full.strip()

    return reply


with gr.Blocks() as demo:
    gr.Markdown("# RNJ-1 Instruct on ZeroGPU (H200)")

    with gr.Row():
        prompt_box = gr.Textbox(
            label="Input",
            placeholder="Ask RNJ-1 to solve an equation, prove something, etc.",
            lines=6,
        )

    with gr.Row():
        max_tokens_slider = gr.Slider(
            minimum=64,
            maximum=4096,   # long answers so it stops cutting off
            value=1024,
            step=64,
            label="Max new tokens",
        )
        temperature_slider = gr.Slider(
            minimum=0.0,
            maximum=1.0,
            value=0.2,
            step=0.05,
            label="Temperature",
        )

    output_box = gr.Textbox(
        label="RNJ-1 Output",
        lines=20,
    )

    run_btn = gr.Button("Generate")

    run_btn.click(
        fn=generate_reply,
        inputs=[prompt_box, max_tokens_slider, temperature_slider],
        outputs=[output_box],
    )

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