app.py

import gradio as gr
import torch
import random
import transformers
from transformers import T5Tokenizer
from optimum.onnxruntime import ORTModelForSeq2SeqLM

# --- CUDA / Provider Setup ---
if torch.cuda.is_available():
    device = "cuda"
    provider = "CUDAExecutionProvider"
    print(f"Using GPU with {provider}")
else:
    device = "cpu"
    provider = "CPUExecutionProvider"
    print("Using CPU")

# Load Model with ONNX Runtime for Execution Provider support
# Note: This requires optimum installed: pip install optimum[onnxruntime-gpu]
try:
    model = ORTModelForSeq2SeqLM.from_pretrained(
        "roborovski/superprompt-v1",
        provider=provider,
        export=False # Set True if you want to force generate ONNX files from pytorch
    )
    print(f"Model loaded successfully using {provider}")
except Exception as e:
    print(f"Failed to load ONNX model: {e}")
    print("Falling back to standard PyTorch model...")
    from transformers import T5ForConditionalGeneration
    model = T5ForConditionalGeneration.from_pretrained(
        "roborovski/superprompt-v1", legacy=False,
        device_map="auto", 
        torch_dtype="auto"
    )
    # Standard torch model doesn't use ExecutionProvider string, but we keep the logic intact
    if device == "cuda":
        model.to(device)

tokenizer = T5Tokenizer.from_pretrained("roborovski/superprompt-v1")

def generate(your_prompt, task_prefix, max_new_tokens, repetition_penalty, temperature, model_precision_type, top_p, top_k, seed):
    
    if seed == 0:
        seed = random.randint(1, 2**32-1)
    transformers.set_seed(seed)
    
    # ONNX Runtime models usually manage their own precision/quantization via the file loaded,
    # but we can leave the UI option for users to switch logic if they were swapping models.
    # For this specific implementation, the precision is largely determined by the loaded provider/weights.
    
    repetition_penalty = float(repetition_penalty)

    input_text = f"{task_prefix}: {your_prompt}"
    # ONNX models generally handle input tensors on the device they were initialized with automatically
    input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to(device)
        
    # ONNX Runtime generate function might differ slightly in arguments, but standard transformers args usually map over.
    # We ensure we pass the device properly for PyTorch fallback.
    if hasattr(model, 'device'):
        input_ids = input_ids.to(model.device)
        
    outputs = model.generate(
        input_ids,
        max_new_tokens=max_new_tokens,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
    )
        
    better_prompt = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return better_prompt


your_prompt = gr.Textbox(label="Your Prompt", info="Your Prompt that you wanna make better")

task_prefix = gr.Textbox(label="Task Prefix", info="The prompt prefix for how the AI should make yours better",value="Expand the following prompt to add more detail")

max_new_tokens = gr.Slider(value=512, minimum=25, maximum=512, step=1, label="Max New Tokens", info="The maximum numbers of new tokens, controls how long is the output")
    
repetition_penalty = gr.Slider(value=1.2, minimum=0, maximum=2.0, step=0.05, label="Repetition Penalty", info="Penalize repeated tokens, making the AI repeat less itself")

temperature = gr.Slider(value=0.7, minimum=0, maximum=1, step=0.05, label="Temperature", info="Higher values produce more diverse outputs")

model_precision_type = gr.Dropdown(["fp16", "fp32"], value="fp16", label="Model Precision Type", info="The precision type to load the model, like fp16 which is faster, or fp32 which is more precise but more resource consuming")

top_p = gr.Slider(value=1, minimum=0, maximum=2, step=0.05, label="Top P", info="Higher values sample more low-probability tokens")

top_k = gr.Slider(value=50, minimum=1, maximum=100, step=1, label="Top K", info="Higher k means more diverse outputs by considering a range of tokens")

seed = gr.Slider(value=42, minimum=0, maximum=2**32-1, step=1, label="Seed", info="A starting point to initiate the generation process, put 0 for a random one")

examples = [
    ["A storefront with 'Text to Image' written on it.", "Expand the following prompt to add more detail", 512, 1.2, 0.5, "fp16", 1, 50, 42]
]

gr.Interface(
    fn=generate,
    inputs=[your_prompt, task_prefix, max_new_tokens, repetition_penalty, temperature, model_precision_type, top_p, top_k, seed],
    outputs=gr.Textbox(label="Better Prompt"),
    title="SuperPrompt-v1",
    description='Make your prompts more detailed! <br> <br> Hugging Face Space made by Nick088 improved bu NeoPy/BF667',
    examples=examples,
    theme="NeoPy/Soft"
).launch(share=True)