app.py

import os
import tarfile
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr

# 1. Download the tar.gz archive
archive_path = hf_hub_download(
    repo_id="SamOrion/Llama_3.2_3b_Hindi_Pruned",
    filename="llama-3.2-3b-hindi-pruned.tar.gz",
    repo_type="model",
)

# 2. Extract into './model', stripping the top-level folder
extract_dir = "./model"
os.makedirs(extract_dir, exist_ok=True)

with tarfile.open(archive_path, "r:gz") as tar:
    for member in tar.getmembers():
        # Skip the first path component (the folder name)
        parts = member.name.split("/", 1)
        if len(parts) == 2:
            member.name = parts[1]
            tar.extract(member, path=extract_dir)

# 3. Verify that config.json is at ./model/config.json
config_path = os.path.join(extract_dir, "config.json")
if not os.path.isfile(config_path):
    raise FileNotFoundError(f"config.json not found in {extract_dir}")

# 4. Load tokenizer and model straight from './model'
tokenizer = AutoTokenizer.from_pretrained(extract_dir)
model = AutoModelForCausalLM.from_pretrained(
    extract_dir,
    torch_dtype="auto",
    device_map="auto",
    low_cpu_mem_usage=True,
)

# 5. Define chat function using OpenAI-style messages
def chat_fn(prompt, history):
    history = history or []
    history.append({"role": "user", "content": prompt})

    # Tokenize and send to device
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    input_length = inputs["input_ids"].shape[1]

    # Generate full-sequence tokens (prompt + continuation)
    outputs = model.generate(**inputs, max_new_tokens=100)

    # Remove the prompt tokens from the start
    generated_ids = outputs[0][input_length:]
    response = tokenizer.decode(generated_ids, skip_special_tokens=True)

    history.append({"role": "assistant", "content": response})
    return history, ""


# 6. Build Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("## 🌐 Indus 3.0 Hindi LLM Demo")
    chat = gr.Chatbot(type="messages")
    msg = gr.Textbox(placeholder="Type here…")
    clear = gr.Button("Clear")
    msg.submit(chat_fn, [msg, chat], [chat, msg])
    clear.click(lambda: ([], ""), None, [chat, msg])

# 7. Launch without `share=True` on Spaces
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860, show_api=False)