app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
import os
import spaces
import torch
from datasets import load_dataset
from huggingface_hub import CommitScheduler
from pathlib import Path
import uuid


device = "cuda:0" if torch.cuda.is_available() else "cpu"
print(f'[INFO] Using device: {device}')

# token
token = os.environ['TOKEN']

# Load the pretrained model and tokenizer
MODEL_NAME = "atlasia/Al-Atlas-0.5B" # "atlasia/Al-Atlas-LLM-mid-training" # "BounharAbdelaziz/Al-Atlas-LLM-0.5B" #"atlasia/Al-Atlas-LLM"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME,token=token) # , token=token
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME,token=token).to(device)

# Fix tokenizer padding
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token  # Set pad token

# Predefined examples
examples = [
    ["الذكاء الاصطناعي هو فرع من علوم الكمبيوتر اللي كيركز"
     , 256, 0.7, 0.9, 150, 8, 1.5],
    ["المستقبل ديال الذكاء الصناعي فالمغرب"
     , 256, 0.7, 0.9, 150, 8, 1.5],
    [" المطبخ المغربي"
     , 256, 0.7, 0.9, 150, 8, 1.5],
    ["الماكلة المغربية كتعتبر من أحسن الماكلات فالعالم"
     , 256, 0.7, 0.9, 150, 8, 1.5],
]

#inf_dataset=load_dataset("atlasia/atlaset_inference_ds",token=token,split="test",name="llm")
submit_file = Path("user_submit/") / f"data_{uuid.uuid4()}.json"
scheduler = CommitScheduler(
            repo_id="atlasia/atlaset_inference_ds",
            repo_type="dataset",
            folder_path=submit_file,
            every=5,
            token=token
        )
@spaces.GPU
def generate_text(prompt, max_length=256, temperature=0.7, top_p=0.9, top_k=150, num_beams=8, repetition_penalty=1.5):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    output = model.generate(
        **inputs, 
        max_length=max_length, 
        temperature=temperature, 
        top_p=top_p, 
        do_sample=True,
        repetition_penalty=repetition_penalty,
        num_beams=num_beams,
        top_k= top_k,
        early_stopping = True,
        pad_token_id=tokenizer.pad_token_id,  # Explicit pad token
        eos_token_id=tokenizer.eos_token_id,  # Explicit eos token
    )
    result=tokenizer.decode(output[0], skip_special_tokens=True)
    #inf_dataset.add_item({"inputs":prompt,"outputs":result,"params":f"{max_length},{temperature},{top_p},{top_k},{num_beams},{repetition_penalty}"})
    save_feedback(prompt,result,f"{max_length},{temperature},{top_p},{top_k},{num_beams},{repetition_penalty}")
    return result

def save_feedback(input,output,params) -> None:
    with scheduler.lock:
        with submit_file.open("a") as f:
            f.write(json.dumps({"input": input, "output": output, "params": params}))
            f.write("\n")
    
if __name__ == "__main__":
    # Create the Gradio interface
    with gr.Blocks() as app:
        gr.Interface(
            fn=generate_text,
            inputs=[
                gr.Textbox(label="Prompt: دخل النص بالدارجة"),
                gr.Slider(8, 4096, value=256, label="Max Length"),
                gr.Slider(0.0, 2, value=0.7, label="Temperature"),
                gr.Slider(0.0, 1.0, value=0.9, label="Top-p"),
                gr.Slider(1, 10000, value=150, label="Top-k"),
                gr.Slider(1, 20, value=8, label="Number of Beams"),
                gr.Slider(0.0, 100.0, value=1.5, label="Repetition Penalty"),
            ],
            outputs=gr.Textbox(label="Generated Text in Moroccan Darija"),
            title="Moroccan Darija LLM",
            description="Enter a prompt and get AI-generated text using our pretrained LLM on Moroccan Darija.",
            examples=examples,
        )
    app.launch()