app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# 1. Setup Model IDs
base_model_id = "unsloth/Qwen2.5-3B-Instruct"
lora_model_id = "10Aizen01/qwen-2.5-3b-engine-simulator-beta"

# 2. Load Tokenizer and Base Model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
# We use float32 and force CPU for the free Hugging Face tier
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.float32, 
    device_map={"": "cpu"},
    low_cpu_mem_usage=True
)

# 3. Load your LoRA adapters
model = PeftModel.from_pretrained(base_model, lora_model_id)

def generate_engine_code(prompt):
    # Removed .to("cuda") here
    inputs = tokenizer(f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n", return_tensors="pt")
    
    # Generate on CPU
    outputs = model.generate(
        **inputs, 
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# 4. Create the Web UI
demo = gr.Interface(
    fn=generate_engine_code,
    inputs=gr.Textbox(label="Describe your engine (e.g., V8, 4.0L, 9000 RPM)"),
    outputs=gr.Code(label="Generated .mr Script", language="cpp"),
    title="Engine Simulator AI Assistant"
)

demo.launch()