app.py

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

# =========================
# CONFIG
# =========================

# Base NLLB model
BASE_MODEL = "facebook/nllb-200-distilled-600M"

# Your LoRA repo on HF Hub
# 👉 CHANGE THIS to your actual repo if different
LORA_REPO = "flt7007/nllb-mizo-bible-lora"
# e.g. "frankiethiak/nllb-mizo-bible-lora"

# NLLB language codes
SRC_LANG = "eng_Latn"  # English
TGT_LANG = "lus_Latn"  # Mizo (Lushai / Mizo)


# =========================
# LOAD TOKENIZER + MODEL
# =========================

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

print("Using device:", device)

# 🔴 IMPORTANT:
# Load tokenizer from the LoRA repo, not the base model
# This fixes the “ mojibake issue.
tokenizer = AutoTokenizer.from_pretrained(
    LORA_REPO,
    src_lang=SRC_LANG
)

# Load base NLLB model
base_model = AutoModelForSeq2SeqLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=dtype
)

# Attach LoRA
model = PeftModel.from_pretrained(
    base_model,
    LORA_REPO
)

model.to(device)
model.eval()

# Try to set forced BOS for Mizo if available
forced_bos_token_id = None
if hasattr(tokenizer, "lang_code_to_id"):
    forced_bos_token_id = tokenizer.lang_code_to_id.get(TGT_LANG, None)
    print("forced_bos_token_id:", forced_bos_token_id)
else:
    print("Tokenizer has no lang_code_to_id; continuing without forced BOS.")


# =========================
# TRANSLATION FUNCTION
# =========================

def translate_en_to_mizo(text, max_new_tokens, num_beams):
    text = text.strip()
    if not text:
        return ""

    inputs = tokenizer(
        text,
        return_tensors="pt"
    ).to(device)

    gen_kwargs = {
        "max_new_tokens": int(max_new_tokens),
        "num_beams": int(num_beams),
    }
    # Only pass forced_bos_token_id if we actually have it
    if forced_bos_token_id is not None:
        gen_kwargs["forced_bos_token_id"] = forced_bos_token_id

    with torch.no_grad():
        outputs = model.generate(**inputs, **gen_kwargs)

    decoded = tokenizer.batch_decode(
        outputs,
        skip_special_tokens=True
    )[0]

    return decoded.strip()


# =========================
# GRADIO UI
# =========================

TITLE = "English → Mizo (NLLB-200 + Bible+Dict LoRA)"
DESC = """
Low-resource MT demo for **English → Mizo** using:
- Base model: `facebook/nllb-200-distilled-600M`
- LoRA: fine-tuned on dictionary + Bible parallel data  
Model is more Bible/education style and still in-progress.
"""

with gr.Blocks() as demo:
    gr.Markdown(f"# {TITLE}")
    gr.Markdown(DESC)

    with gr.Row():
        with gr.Column():
            en_input = gr.Textbox(
                label="English input",
                lines=4,
                placeholder="Type an English sentence here…"
            )
            max_new_tokens = gr.Slider(
                minimum=10,
                maximum=200,
                value=80,
                step=5,
                label="Max new tokens"
            )
            num_beams = gr.Slider(
                minimum=1,
                maximum=8,
                value=4,
                step=1,
                label="Beam size"
            )
            translate_btn = gr.Button("Translate → Mizo")

        with gr.Column():
            mz_output = gr.Textbox(
                label="Mizo output",
                lines=6
            )

    translate_btn.click(
        fn=translate_en_to_mizo,
        inputs=[en_input, max_new_tokens, num_beams],
        outputs=mz_output
    )

demo.queue()
if __name__ == "__main__":
    demo.launch()