app_dialogue.py

import os
import gradio as gr


models = [
    "HuggingFaceM4/idefics-9b-instruct",
    # "HuggingFaceM4/idefics-80b-instruct",
]

SYSTEM_PROMPT = """The following is a conversation between a highly knowledgeable and intelligent visual AI assistant, called Assistant, and a human user, called User. In the following interactions, User and Assistant will converse in natural language, and Assistant will do its best to answer User’s questions. Assistant has the ability to perceive images and reason about the content of visual inputs. Assistant was built to be respectful, polite and inclusive. It knows a lot, and always tells the truth. When prompted with an image, it does not make up facts.

The conversation begins:
User:<fake_token_around_image><image:https://i1.pickpik.com/photos/515/906/355/kittens-cats-pet-cute-preview.jpg><fake_token_around_image>Describe this image.</s>
Assistant: Five kittens are standing together in the center of the photograph. From the left to right, there is one orange kitten, two kittens white and black stripes, and two brown kittens with an orange head. They are in bright green grass and it looks like they are moving forward.</s>
User:How about this image?<fake_token_around_image><image:https://cdn.pixabay.com/photo/2017/09/25/13/12/puppy-2785074_1280.jpg><fake_token_around_image>Can you describe it too?</s>
Assistant: A dog is lying on the floor, looking at the camera. It is looking directly at you. The dog has a white body and brown patches on its face and ears. Its eyes are dark. Its nose is black, and it has long, floppy ears, white paws, long fur, big eyes.</s>
User: What kind of breed is it?</s>
Assistant: It looks like a Cavalier King Charles Spaniel. These breeds are known for their gentle, affectionate nature. Their long ears and big, brown eyes give them a sweet, expressive face. Cavaliers are small dogs weighing between 13 and 20 pounds. They have a silky, medium-length coat that comes in various colors, including black and tan, tricolor, red and white, or black.</s>
User: What can you tell me about this breed of dogs?</s>
Assistant: One specific characteristic of the Cavalier King Charles Spaniel is its friendly and affectionate nature. This breed is known for being extremely sociable and forming strong bonds with their owners. They are often described as "velcro dogs" because they love to be close to their human companions, whether it's sitting on their laps, cuddling on the couch, or simply following them around the house.</s>
User: ghjkhjabnufs</s>
Assistant: That doesn’t seem to be a word. Could you ask me another way?</s>
User: Do you like Cavalier King Charles Spaniel?</s>
Assistant: I do not have personal opinions as I’m just a computer program. However, cavaliers are known for being excellent family pets due to their gentle and patient demeanor, making them great with children and other pets. Their affectionate nature also makes them ideal therapy dogs, as they can provide comfort and emotional support to those in need.</s>
User: How many dogs do you see in this image?<fake_token_around_image><image:https://i.dailymail.co.uk/i/pix/2011/07/01/article-2010308-0CD22A8300000578-496_634x414.jpg><fake_token_around_image></s>
Assistant: There is no dogs in this image. The picture shows a tennis player jumping to volley the ball.</s>"""

BAN_TOKENS = "<image>;<fake_token_around_image>"
EOS_TOKENS = "</s>;User"

import logging

from accelerate.utils import get_max_memory
from transformers import AutoConfig, IdeficsForVisionText2Text, AutoProcessor


TOKENIZER_FAST = True
MAX_SEQ_LEN = 2048

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger()


def load_processor_tokenizer_model(model_name):
    processor = AutoProcessor.from_pretrained(
        model_name,
        use_auth_token=os.getenv("HF_AUTH_TOKEN", True),
        truncation_side="left",
    )
    tokenizer = processor.tokenizer
    # tokenizer = AutoTokenizer.from_pretrained(
    #     model_name,
    #     use_fast=TOKENIZER_FAST,
    #     use_auth_token=os.getenv("HF_AUTH_TOKEN", True),
    #     truncation_side="left",
    # )
    # tokenizer.padding_side = "left" -> we don't need that, do we?

    config = AutoConfig.from_pretrained(model_name, use_auth_token=os.getenv("HF_AUTH_TOKEN", True))
    max_memory_map = get_max_memory()

    for key in max_memory_map.keys():
        if key != "cpu":
            # Get this in GB
            max_memory_map[key] = max_memory_map[key] // (1024 * 1024 * 1024)
            # Decrease 2 for Pytorch overhead and 2 for the forward to be safe
            max_memory_map[key] = f"{max_memory_map[key] - 4} GiB"

    model = IdeficsForVisionText2Text.from_pretrained(
        model_name,
        use_auth_token=os.getenv("HF_AUTH_TOKEN", True),
        device_map="auto",
        offload_folder="./offload",
        torch_dtype=config.torch_dtype,
        max_memory=max_memory_map,
    )
    model.eval()
    print("Current device map:", model.hf_device_map)
    print("Model default generation config:", model.generation_config)
    # TODO: the device_map looks very inefficien right now. that could be improved
    return processor, tokenizer, model


def split_prompt_into_list(prompt_str):
    """Convert a full string prompt to the list format expected by the processor."""
    prompt_splitted = prompt_str.split("<fake_token_around_image>")
    prompt_list = []
    for ps in prompt_splitted:
        if ps.startswith("<image:"):
            prompt_list.append(ps[7:-1])
        else:
            prompt_list.append(ps)
    return prompt_list


def model_generation(
    prompt,
    processor,
    tokenizer,
    model,
    temperature,
    no_repeat_ngram_size,
    max_new_tokens,
    min_length,
    ban_tokens,
    eos_tokens,
    force_words,
    repetition_penalty,
    hide_special_tokens,
    decoding_strategy,
    num_beams,
    length_penalty,
    top_k,
    top_p,
    penalty_alpha,
):
    input_args = processor(
        [split_prompt_into_list(prompt)],
        eval_mode=True,
        truncation=True,
        max_length=MAX_SEQ_LEN - 512,  # TODO: replace the 512 value with `max_new_tokens`
        padding=True,
    )
    for k, v in input_args.items():
        input_args[k] = v.to(0)

    # Excluding some words from the generation
    bad_words_ids = None
    ban_tokens = ban_tokens.replace("\\n", "\n")
    bad_words = ban_tokens.split(";")
    if len(bad_words) > 0:
        bad_words_ids = tokenizer(bad_words, add_special_tokens=False).input_ids

    # Forcing some words in the generation
    force_words_ids = None
    if force_words != "":
        force_words = force_words.replace("\\n", "\n")
        force_words = force_words.split(";")
        if len(force_words) > 0:
            force_words_ids = tokenizer(force_words, add_special_tokens=False).input_ids

    eos_token_ids = None
    if eos_tokens != "":
        eos_tokens = eos_tokens.replace("\\n", "\n")
        eos_tokens = eos_tokens.split(";")
        if len(eos_tokens) > 0:
            eos_token_ids = []
            for eos_token in eos_tokens:
                tokenized_eos_token = tokenizer(eos_token, add_special_tokens=False).input_ids
                if len(tokenized_eos_token) > 1:
                    raise ValueError(
                        f"eos_tokens should be one token, here {eos_token} is {len(tokenized_eos_token)} tokens:"
                        f" {tokenized_eos_token}"
                    )
                eos_token_ids += tokenized_eos_token

    # Common parameters to all decoding strategies
    # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies
    generation_args = {
        "temperature": temperature,
        "no_repeat_ngram_size": no_repeat_ngram_size,
        "max_new_tokens": max_new_tokens,
        "min_length": min_length,
        "bad_words_ids": bad_words_ids,
        "force_words_ids": force_words_ids,
        "repetition_penalty": repetition_penalty,
        "eos_token_id": eos_token_ids,
    }

    assert decoding_strategy in [
        "greedy",
        "beam_search",
        "beam_sampling",
        "sampling_top_k",
        "sampling_top_p",
        "contrastive_sampling",
    ]
    if decoding_strategy == "greedy":
        pass
    elif decoding_strategy == "beam_search":
        generation_args["num_beams"] = num_beams
        generation_args["length_penalty"] = length_penalty
        assert generation_args["num_beams"] > 1
    elif decoding_strategy == "beam_sampling":
        generation_args["num_beams"] = num_beams
        generation_args["length_penalty"] = length_penalty
        generation_args["do_sample"] = True
        assert generation_args["num_beams"] > 1
    elif decoding_strategy == "sampling_top_k":
        generation_args["do_sample"] = True
        generation_args["top_k"] = top_k
    elif decoding_strategy == "sampling_top_p":
        generation_args["do_sample"] = True
        generation_args["top_p"] = top_p
    elif decoding_strategy == "contrastive_sampling":
        generation_args["do_sample"] = True
        generation_args["penalty_alpha"] = penalty_alpha
        generation_args["top_k"] = top_k

    generated_tokens = model.generate(
        **input_args,
        **generation_args,
    )

    tokens = tokenizer.convert_ids_to_tokens(generated_tokens[0])
    decoded_skip_special_tokens = repr(
        tokenizer.batch_decode(generated_tokens, skip_special_tokens=hide_special_tokens)[0]
    )

    actual_generated_tokens = generated_tokens[:, input_args["input_ids"].shape[-1] :]
    first_end_token = len(actual_generated_tokens[0])
    actual_generated_tokens = actual_generated_tokens[:, :first_end_token]
    generated_text = tokenizer.batch_decode(actual_generated_tokens, skip_special_tokens=hide_special_tokens)[0]

    logger.info(
        "Result: \n"
        f"----Prompt: `{prompt}`\n"
        f"----Tokens ids - prompt + generation: `{generated_tokens[0].tolist()}`\n"
        f"----Tokens converted - prompt + generation: `{tokens}`\n"
        f"----String decoded with skipped special tokens - prompt + generation: `{decoded_skip_special_tokens}`\n"
        f"----Total length - prompt + generation `{len(generated_tokens[0].tolist())}`\n"
        f"----Token ids - generation: `{actual_generated_tokens[0].tolist()}`\n"
        f"----Tokens converted - generation: `{tokenizer.convert_ids_to_tokens(actual_generated_tokens[0])}`\n"
        f"----String decoded with skipped special tokens - generation: `{generated_text}`\n"
        f"----Total length - generation: `{len(actual_generated_tokens[0].tolist())}`\n"
        f"----Generation mode: `{decoding_strategy}`\n"
        f"----Generation parameters: `{generation_args}`\n"
    )

    return generated_text


textbox = gr.Textbox(
    show_label=False,
    value="What color are the cat's eyes?",
    placeholder=(
        "To input images, use the following syntax:"
        " `<fake_token_around_image><image:URL_IMAGE><fake_token_around_image>textexttext`"
    ),
    visible=True,
    container=False,
)
with gr.Blocks(title="IDEFICS", theme=gr.themes.Base()) as demo:
    # state = gr.State()

    with gr.Row():
        with gr.Column(scale=3):
            with gr.Row(elem_id="model_selector_row"):
                model_selector = gr.Dropdown(
                    choices=models,
                    value=models[0] if len(models) > 0 else "",
                    interactive=True,
                    show_label=False,
                    container=False,
                )
            processor, tokenizer, model = load_processor_tokenizer_model(model_selector.value)

            imagebox = gr.Image(
                type="pil",
                label=(
                    "Image input - This image box is not supported yet! To include images, do through the text by"
                    " adding `<fake_token_around_image><image:IMAGE_URL><fake_token_around_image>`. The backend takes"
                    " care of parsing that <image:URL> and download the correponding image. That way, you can"
                    " technically interleave as many images and texts as you want. No need to add space before and"
                    " after `<fake_token_around_image>`"
                ),
            )

            with gr.Accordion("Generation parameters", open=False, visible=True) as parameter_row:
                max_new_tokens = gr.Slider(
                    minimum=0,
                    maximum=2048,
                    value=512,
                    step=1,
                    interactive=True,
                    label="Maximum number of new tokens to generate",
                )
                min_length = gr.Slider(
                    minimum=0,
                    maximum=50,
                    value=0,
                    step=1,
                    interactive=True,
                    label="Minimum number of new tokens to generate",
                )
                repetition_penalty = gr.Slider(
                    minimum=0.0,
                    maximum=5.0,
                    value=1.0,
                    step=0.1,
                    interactive=True,
                    label="Repetition penalty",
                    info="1.0 means no penalty",
                )
                no_repeat_ngram_size = gr.Slider(
                    minimum=0,
                    maximum=10,
                    value=0,
                    step=1,
                    interactive=True,
                    label="N-gram repetition threshold",
                    info="If set to int > 0, all ngrams of that size can only occur once.",
                )
                decoding_strategy = gr.Radio(
                    [
                        "greedy",
                        "beam_search",
                        "beam_sampling",
                        "sampling_top_k",
                        "sampling_top_p",
                    ],
                    value="greedy",
                    label="Decoding strategy",
                    interactive=True,
                )
                temperature = gr.Slider(
                    minimum=0.0,
                    maximum=5.0,
                    value=1.0,
                    step=0.1,
                    interactive=True,
                    label="Sampling temperature",
                )
                num_beams = gr.Slider(
                    minimum=0,
                    maximum=20,
                    value=3.0,
                    step=1.0,
                    interactive=True,
                    visible=False,
                    label="Number of beams",
                    info="Only used if `decoding_strategy` is `beam_search` or `beam_sampling`.",
                )
                decoding_strategy.change(
                    fn=lambda selection: gr.Slider.update(visible=(selection in ["beam_search", "beam_sampling"])),
                    inputs=decoding_strategy,
                    outputs=num_beams,
                )
                top_p = gr.Slider(
                    minimum=0.0,
                    maximum=1.0,
                    value=0.95,
                    step=0.01,
                    interactive=True,
                    visible=False,
                    label="Top P",
                    info=(
                        "If set to float < 1, only the smallest set of most probable tokens with probabilities that"
                        " add up to top_p or higher are kept for generation."
                    ),
                )
                decoding_strategy.change(
                    fn=lambda selection: gr.Slider.update(visible=(selection in ["sampling_top_p"])),
                    inputs=decoding_strategy,
                    outputs=top_p,
                )
                top_k = gr.Slider(
                    minimum=0,
                    maximum=500,
                    value=50,
                    step=1,
                    interactive=True,
                    visible=False,
                    label="Top K",
                    info="The number of highest probability vocabulary tokens to keep for top-k-filtering.",
                )
                decoding_strategy.change(
                    fn=lambda selection: gr.Slider.update(visible=(selection in ["sampling_top_k"])),
                    inputs=decoding_strategy,
                    outputs=top_k,
                )
                length_penalty = gr.Slider(
                    minimum=-1000.0,
                    maximum=1000.0,
                    value=1.0,
                    step=0.1,
                    interactive=True,
                    visible=False,
                    label="Length penalty",
                    info=(
                        "length_penalty > 0.0 promotes longer sequences, while length_penalty < 0.0 encourages shorter"
                        " sequences. Only used if `decoding_strategy` is `beam_search` or `beam_sampling`."
                    ),
                )
                decoding_strategy.change(
                    fn=lambda selection: gr.Slider.update(visible=(selection in ["beam_search", "beam_sampling"])),
                    inputs=decoding_strategy,
                    outputs=length_penalty,
                )
                penalty_alpha = gr.Slider(
                    minimum=0.0,
                    maximum=5.0,
                    value=0.95,
                    step=0.05,
                    interactive=True,
                    visible=False,
                    label="Penalty alpha",
                    info="Only used if `decoding_strategy` is `contrastive_sampling`.",
                )
                decoding_strategy.change(
                    fn=lambda selection: gr.Slider.update(visible=(selection in ["contrastive_sampling"])),
                    inputs=decoding_strategy,
                    outputs=penalty_alpha,
                )

        with gr.Column(scale=6):
            chatbot = gr.Chatbot(
                elem_id="chatbot",
                label="Idefics Chatbot",
                visible=True,
                height=550,
                value=[
                    [
                        (
                            "What's unusual about this"
                            " image?<fake_token_around_image><image:https://i.redd.it/9nkcvbi1pp9y.jpg><fake_token_around_image>"
                        ),
                        (
                            "The unusual aspect of this image is that there is a cat lying on a bed with an orange on"
                            " its stomach. The cat is white and fluffy, and it appears to be sleeping or resting"
                            " comfortably with the orange on its belly. This is an uncommon sight, as cats typically"
                            " do not interact with fruits in such a manner."
                        ),
                    ],
                    [
                        (
                            "<fake_token_around_image><image:https://www.boredpanda.com/blog/wp-content/uploads/2014/02/funny-wet-cats-coverimage.jpg><fake_token_around_image>What"
                            " about this cat? Why do people find it funny?"
                        ),
                        (
                            "The cat in the image is a gray and white long-haired cat with a surprised expression on"
                            " its face. The cat's eyes are wide open, and it appears to be looking directly at the"
                            " camera. The cat's mouth is also open, which adds to its surprised and funny appearance."
                            " The combination of the cat's surprised expression and its wet fur, possibly from being"
                            " in the shower or bath, makes it an amusing and entertaining sight for many people."
                        ),
                    ],
                ],
            )
            with gr.Row():
                with gr.Column(scale=8):
                    textbox.render()
                with gr.Column(scale=1, min_width=60):
                    submit_btn = gr.Button(value="Submit", visible=True)
                with gr.Column(scale=1, min_width=20):
                    clear_btn = gr.ClearButton([textbox, chatbot])
                    cur_dir = os.path.dirname(os.path.abspath(__file__))
    
    #def process_example(message: str) -> tuple[str, list[tuple[str, str]]]:
    #    chat = model_inference(message, [], "greedy", 3, 1.0, 0, 512, 16, 1.0, 1.0, 50, 0.95, 0.95)
    #    return '', chat

    #gr.Examples(
    #     examples=[
    #         [
    #             ("How many of these animals can we fit into an engine like that<fake_token_around_image><image:https://upload.wikimedia.org/wikipedia/commons/thumb/4/4e/Nassau_County_Police_Bell_407.jpg/1200px-Nassau_County_Police_Bell_407.jpg><fake_token_around_image>?", "The image shows a helicopter with a large engine, but it is not possible to determine the exact number of animals that can fit into it based on the image alone. The size and capacity of the helicopter's engine would depend on various factors, such as the size of the animals, the weight of the animals, and the size of the helicopter itself. However, it is safe to assume that the helicopter is designed to carry a limited number of animals, and it is not intended to be used as a means of transporting large groups of animals."),
    #         ],
    #     ],
    #     inputs = [chatbot]
    # )

 
    def format_prompt_with_history_and_system_conditioning(current_user_prompt, history):
        resulting_text = SYSTEM_PROMPT
        for turn in history:
            user_utterance, assistant_utterance = turn
            resulting_text += f"\nUser: {user_utterance}</s>\nAssistant: {assistant_utterance}</s>"
        resulting_text += f"\nUser: {current_user_prompt}</s>\nAssistant:"
        return resulting_text

    def model_inference(
        user_prompt,
        chat_history,
        decoding_strategy,
        num_beams,
        temperature,
        no_repeat_ngram_size,
        max_new_tokens,
        min_length,
        repetition_penalty,
        length_penalty,
        top_k,
        top_p,
        penalty_alpha,
    ):
        global processor, model, tokenizer
        # temperature = 1.0
        # no_repeat_ngram_size = 0
        # max_new_tokens = 512
        # min_length = 16
        force_words = ""
        # repetition_penalty = 1.0
        hide_special_tokens = False
        # decoding_strategy = "greedy"
        # num_beams = 3
        # length_penalty = 1.0
        # top_k = 50
        # top_p = 0.95
        # penalty_alpha = 0.95

        formated_prompt = format_prompt_with_history_and_system_conditioning(
            current_user_prompt=user_prompt.strip(),
            history=chat_history,
        )

        generated_text = model_generation(
            prompt=formated_prompt,
            processor=processor,
            tokenizer=tokenizer,
            model=model,
            temperature=temperature,
            no_repeat_ngram_size=no_repeat_ngram_size,
            max_new_tokens=max_new_tokens,
            min_length=min_length,
            ban_tokens=BAN_TOKENS,
            eos_tokens=EOS_TOKENS,
            force_words=force_words,
            repetition_penalty=repetition_penalty,
            hide_special_tokens=hide_special_tokens,
            decoding_strategy=decoding_strategy,
            num_beams=num_beams,
            length_penalty=length_penalty,
            top_k=top_k,
            top_p=top_p,
            penalty_alpha=penalty_alpha,
        )

        chat_history.append((user_prompt, generated_text.strip("</s>")))
        return "", chat_history

    textbox.submit(
        fn=model_inference,
        inputs=[
            textbox,
            chatbot,
            decoding_strategy,
            num_beams,
            temperature,
            no_repeat_ngram_size,
            max_new_tokens,
            min_length,
            repetition_penalty,
            length_penalty,
            top_k,
            top_p,
            penalty_alpha,
        ],
        outputs=[textbox, chatbot],
    )
    submit_btn.click(
        fn=model_inference,
        inputs=[
            textbox,
            chatbot,
            decoding_strategy,
            num_beams,
            temperature,
            no_repeat_ngram_size,
            max_new_tokens,
            min_length,
            repetition_penalty,
            length_penalty,
            top_k,
            top_p,
            penalty_alpha,
        ],
        outputs=[
            textbox,
            chatbot,
        ],
    )

    #gr.Examples(examples=[
    #    [f"{cur_dir}/examples/extreme_ironing.jpg", "What is unusual about this image?"], 
    #    [f"{cur_dir}/examples/waterview.jpg", "What are the things I should be cautious about when I visit here?"],
    #    ["m4-dialogue/images/bear.jpg", "Describe this image"],
    #    ], 
    #    inputs=[textbox],
    #    outputs=[textbox, chatbot],
    #    fn=process_example,
    #    cache_examples=True,               
    #)    

demo.queue()
demo.launch()