app.py

import gradio as gr
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

HUGGING_FACE_USER_NAME = "elalimy"
model_name = "my_awesome_peft_finetuned_helsinki_model"
peft_model_id = f"{HUGGING_FACE_USER_NAME}/{model_name}"

# Load model configuration (assuming it's saved locally)
config = PeftConfig.from_pretrained(peft_model_id)
# Load the base model from its local directory (replace with actual model type)
base_model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=False)

# Load the tokenizer from its local directory (replace with actual tokenizer type)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Peft model (assuming it's a custom class or adaptation)
AI_model = PeftModel.from_pretrained(base_model, peft_model_id)

def generate_translation(source_text, device="cpu"):
    # Encode the source text
    input_ids = tokenizer.encode(source_text, return_tensors='pt').to(device)

    # Move the model to the same device as input_ids
    model = base_model.to(device)

    # Generate the translation with adjusted decoding parameters
    generated_ids = model.generate(
        input_ids=input_ids,
        max_length=512,  # Adjust max_length if needed
        num_beams=4,
        length_penalty=5,  # Adjust length_penalty if needed
        no_repeat_ngram_size=4,
        early_stopping=True
    )

    # Decode the generated translation excluding special tokens
    generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

    return generated_text

def translate(text):
    return generate_translation(text)

# Define the Gradio interface
iface = gr.Interface(
    fn=translate,
    inputs="text",
    outputs="text",
    title="Translation App",
    description="Translate text using a fine-tuned Helsinki model."
)

# Launch the Gradio app
iface.launch()