app.py

import gradio as gr
from transformers import AutoModelForCausalLM, LlamaTokenizer
import torch
from huggingface_hub import login
import re
import os

login(token=os.getenv("HF_TOKEN"))


# Load the model and tokenizer
model_name = "ranggafermata/Fermata-v1.2-lightcoder"
tokenizer = LlamaTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)
model.eval()

def generate_code(prompt, max_tokens, temperature, top_p):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

iface = gr.Interface(
    fn=generate_code,
    inputs=[
        gr.Textbox(lines=5, label="Prompt"),
        gr.Slider(10, 512, value=128, label="Max Tokens"),
        gr.Slider(0.1, 1.5, value=0.8, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, label="Top-p")
    ],
    outputs=gr.Textbox(lines=20, label="Generated Code"),
    title="Fermata v1.2 LightCoder",
    description="A fine-tuned code model based on TinyLlama."
)

iface.launch(mcp_server=True)