app.py

import os
from threading import Thread
from typing import Iterator
from PIL import Image

import gradio as gr
import spaces
import torch
from transformers import (
    AutoProcessor,
    Gemma3ForConditionalGeneration,
    TextIteratorStreamer,
)
import subprocess

subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)


DESCRIPTION = """\
# MamayLM-Gemma-3-12B-IT-v1.0 demo

[🪪 **Model card**](https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0)
"""

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 2048
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

model_id = "INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0"

processor = AutoProcessor.from_pretrained(model_id)
attn_impl = "flash_attention_2" if torch.cuda.is_available() else "eager"

model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation=attn_impl,
)
model.eval()

eos_token_ids = [1, 106]

print("Model loaded successfully.")


@spaces.GPU(duration=120)  # Increased duration for the larger model
def generate(
    message: dict,
    chat_history: list[list],
    system_message: str,
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    top_p: float = 0.95,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:
    """
    Generates a response from the model based on the user's message and chat history.
    This function is designed to work with Gradio's multimodal ChatInterface.
    """
    conversation = []
    all_images = []

    # Add system message if provided. This guides the model's behavior.
    if system_message:
        conversation.append(
            {"role": "system", "content": [{"type": "text", "text": system_message}]}
        )

    # Process past turns from Gradio's chat_history
    for user_turn, bot_turn in chat_history:
        # Reconstruct the user's turn, which might include an image
        user_content = []
        print("---")
        print(user_turn)
        print(bot_turn)
        print("---")
        if isinstance(user_turn, tuple):  # User turn with an image
            if len(user_turn) == 1: # TODO: IDK now how to fix this bug
                continue

            img_path, txt = user_turn

            pil_img = Image.open(img_path).convert("RGB")
            all_images.append(pil_img)
            user_content.append({"type": "image"})
            if txt:
                user_content.append({"type": "text", "text": txt})
        elif user_turn:  # Text-only user turn
            user_content.append({"type": "text", "text": user_turn})

        if user_content:
            conversation.append({"role": "user", "content": user_content})

        # Reconstruct the assistant's turn
        if bot_turn:
            conversation.append(
                {"role": "assistant", "content": [{"type": "text", "text": bot_turn}]}
            )

    # Process the current user message, which can include new images
    current_user_content = []
    text = message["text"]
    for img_path in message["files"]:
        pil_img = Image.open(img_path).convert("RGB")
        all_images.append(pil_img)
        current_user_content.append({"type": "image"})

    if text:
        current_user_content.append({"type": "text", "text": text})

    if current_user_content:
        conversation.append({"role": "user", "content": current_user_content})

    print("### DEBUG:")
    print(conversation)
    print("####")

    # Use the processor to create the prompt and preprocess images
    prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)

    if len(all_images) > 0:
        inputs = processor(text=prompt, images=all_images, return_tensors="pt").to(
            model.device
        )
    else:
        inputs = processor(text=prompt, return_tensors="pt").to(model.device)

    # Set up the streamer for text generation
    streamer = TextIteratorStreamer(
        processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True
    )

    generate_kwargs = dict(
        inputs,
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        eos_token_id=eos_token_ids,
        temperature=temperature
        if temperature > 0
        else 0.001,  # Temperature must be > 0
        top_p=top_p,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
    )
    # Run generation in a separate thread
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    # Yield generated text chunks
    outputs = []
    for text_chunk in streamer:
        outputs.append(text_chunk)
        yield "".join(outputs)


chat_interface = gr.ChatInterface(
    multimodal=True,
    fn=generate,
    additional_inputs=[
        gr.Textbox(
            value="",
            label="System message",
            render=False,
        ),
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0,
            maximum=4.0,
            step=0.1,
            value=0.1,  # default from https://huggingface.co/docs/transformers/en/main_classes/text_generation
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=1,  # from https://huggingface.co/google/gemma-3-270m-it/blob/main/generation_config.json
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=25,  # from https://huggingface.co/google/gemma-3-270m-it/blob/main/generation_config.json
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.1,  # default from https://huggingface.co/docs/transformers/en/main_classes/text_generation
        ),
    ],
    stop_btn="Stop Generation",
    examples=[
        ["Привіт! Як справи?"],
        [
            "Плюси та мінуси довгострокових стосунків. Маркований список із максимум 3 перевагами та 3 недоліками, стисло."
        ],
        ["Скільки годин потрібно людині, щоб з'їсти гелікоптер?"],
        ["Як відкрити файл JSON у Python?"],
        [
            "Створіть маркований список переваг і недоліків життя в Сан-Франциско. Максимум 2 переваги та 2 недоліки."
        ],
        ["Придумай коротке оповідання з тваринами про цінність дружби."],
        ["Чи можеш ти коротко пояснити, що таке мова програмування Python?"],
        [
            "Напишіть статтю на 100 слів на тему 'Переваги відкритого коду в дослідженнях ШІ'."
        ],
    ],
    cache_examples=False,
)

with gr.Blocks(css="style.css", fill_height=True, theme="soft") as demo:
    gr.Markdown(DESCRIPTION)
    chat_interface.render()

if __name__ == "__main__":
    demo.queue(max_size=20).launch()