app.py

"""MOSS-VL-Instruct Space — multi-turn media chat demo.

Full-width chat log with a chat/raw toggle; the input bar sits directly under
it: attach button INSIDE the input area (right side), pending media thumbnails
in a row just above the text (click a thumbnail to open the full-resolution
file), round send button beside the bar. Each turn sends whatever is staged:
media + text, media alone (default prompt), or text alone (no images sent).
TRUE MULTI-TURN: the model sees the whole conversation — earlier turns and
their media — via its offline_batch_generate session-state API (the state is
a plain message list held in gr.State and re-encoded on every stateless
ZeroGPU call). The 🗑 button resets the conversation. analyze_image /
analyze_video / analyze_text remain exposed as single-turn API/MCP tools.

MOCK mode (MOSS_DEMO_MOCK=1): no torch / spaces imports; canned responses for
local UI work.
"""

import inspect
import os
import time

os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")

MOCK = os.getenv("MOSS_DEMO_MOCK") == "1"

import gradio as gr

MODEL_ID = "OpenMOSS-Team/MOSS-VL-Instruct-0708"

if not MOCK:
    import spaces
    import torch
    from transformers import AutoModelForCausalLM, AutoProcessor

    processor = AutoProcessor.from_pretrained(
        MODEL_ID,
        trust_remote_code=True,
        frame_extract_num_threads=1,
    )

    # transformers 4.5x fast image processors pass `interpolation=` into
    # _preprocess, but this model's remote-code MossVLImageProcessorFast still
    # names the parameter `resample` (same torchvision InterpolationMode type,
    # only the name is stale) -> TypeError on every image request. Bridge the
    # rename here; this no-ops once the model repo updates the signature.
    _ip_cls = type(processor.image_processor)
    if "resample" in inspect.signature(_ip_cls._preprocess).parameters:
        _orig_preprocess = _ip_cls._preprocess

        def _compat_preprocess(self, images, *args, **kwargs):
            if "interpolation" in kwargs and "resample" not in kwargs:
                kwargs["resample"] = kwargs.pop("interpolation")
            return _orig_preprocess(self, images, *args, **kwargs)

        _ip_cls._preprocess = _compat_preprocess

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_ID,
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
        attn_implementation="sdpa",
    ).to("cuda")
else:
    class _MockSpaces:
        """Effect-free stand-in for the spaces module in MOCK mode."""

        @staticmethod
        def GPU(*args, **kwargs):
            if args and callable(args[0]):
                return args[0]
            return lambda fn: fn

    spaces = _MockSpaces()


@spaces.GPU(duration=120)
def analyze_image(
    image: str,
    prompt: str,
    max_new_tokens: int,
    temperature: float,
    top_p: float,
    do_sample: bool,
) -> str:
    """Analyze an image with a text prompt.

    Args:
        image: Path of the input image to analyze.
        prompt: Text instruction or question about the image.
        max_new_tokens: Maximum number of tokens to generate.
        temperature: Sampling temperature (1.0 = greedy when do_sample=False).
        top_p: Nucleus sampling probability threshold.
        do_sample: Whether to use sampling instead of greedy decoding.
    """
    if image is None:
        return "Please upload an image."
    if not prompt.strip():
        prompt = "Describe this image in detail."
    if MOCK:
        time.sleep(1.2)
        return (
            f"[MOCK] Scripted image analysis for {os.path.basename(image)} — prompt: {prompt!r}.\n\n"
            + "This is a long mock paragraph to exercise scrolling in the response view. " * 12
        )

    return model.offline_image_generate(
        processor,
        prompt=prompt,
        image=image,
        max_new_tokens=int(max_new_tokens),
        temperature=float(temperature),
        top_p=float(top_p),
        do_sample=do_sample,
    )


@spaces.GPU(duration=180)
def analyze_video(
    video: str,
    prompt: str,
    max_new_tokens: int,
    temperature: float,
    top_p: float,
    video_fps: float,
    max_frames: int,
    do_sample: bool,
) -> str:
    """Analyze a video with a text prompt.

    Args:
        video: Path of the input video file to analyze.
        prompt: Text instruction or question about the video.
        max_new_tokens: Maximum number of tokens to generate.
        temperature: Sampling temperature (1.0 = greedy when do_sample=False).
        top_p: Nucleus sampling probability threshold.
        video_fps: Frames per second to sample from the video.
        max_frames: Maximum number of frames to extract.
        do_sample: Whether to use sampling instead of greedy decoding.
    """
    if video is None:
        return "Please upload a video."
    if not prompt.strip():
        prompt = "Describe this video in detail."
    if MOCK:
        time.sleep(1.5)
        return (
            f"[MOCK] Scripted video analysis for {os.path.basename(video)} — prompt: {prompt!r}.\n\n"
            + "This is a long mock paragraph to exercise scrolling in the response view. " * 12
        )

    return model.offline_video_generate(
        processor,
        prompt=prompt,
        video=video,
        max_new_tokens=int(max_new_tokens),
        temperature=float(temperature),
        top_p=float(top_p),
        video_fps=float(video_fps),
        max_frames=int(max_frames),
        do_sample=do_sample,
    )


@spaces.GPU(duration=90)
def analyze_text(
    prompt: str,
    max_new_tokens: int,
    temperature: float,
    top_p: float,
    do_sample: bool,
) -> str:
    """Answer a text-only prompt (no media attached).

    Args:
        prompt: The text instruction or question.
        max_new_tokens: Maximum number of tokens to generate.
        temperature: Sampling temperature (1.0 = greedy when do_sample=False).
        top_p: Nucleus sampling probability threshold.
        do_sample: Whether to use sampling instead of greedy decoding.
    """
    if not prompt.strip():
        return "Please enter a prompt."
    if MOCK:
        time.sleep(0.8)
        return f"[MOCK] Scripted text answer — prompt: {prompt!r}."

    query = {
        "prompt": prompt,
        "images": [],
        "videos": [],
        "generate_kwargs": {
            "max_new_tokens": int(max_new_tokens),
            "temperature": float(temperature),
            "top_p": float(top_p),
            "repetition_penalty": 1.0,
            "do_sample": bool(do_sample),
        },
    }
    return model._offline_generate_one_with_processor_overrides(processor, query)


@spaces.GPU(duration=180)
def chat_generate(session: list, prompt: str, images: list, videos: list,
                  max_new_tokens: int, temperature: float, top_p: float,
                  do_sample: bool, video_fps: float, max_frames: int) -> dict:
    """One multi-turn chat step: prior session messages + this turn -> reply.

    Built on the model's offline_batch_generate session API. `session` is a
    plain list of message dicts (media referenced by file path), so it
    round-trips through gr.State across stateless ZeroGPU calls; every turn
    re-encodes the full context including earlier media.

    Args:
        session: Message history from the previous call's "session" output ([] to start).
        prompt: This turn's text instruction (may be empty when media is attached).
        images: Image file paths attached to this turn (usually 0 or 1).
        videos: Video file paths attached to this turn (usually 0 or 1).
        max_new_tokens: Maximum number of tokens to generate.
        temperature: Sampling temperature (1.0 = greedy when do_sample=False).
        top_p: Nucleus sampling probability threshold.
        do_sample: Whether to use sampling instead of greedy decoding.
        video_fps: Frames per second to sample from attached videos.
        max_frames: Maximum frames to extract from attached videos.

    Returns:
        {"text": assistant reply, "session": updated message history to pass next turn}.
    """
    session = list(session or [])
    if MOCK:
        time.sleep(1.0)
        n_prev = sum(1 for m in session if isinstance(m, dict) and m.get("role") == "assistant")
        media = (list(images or []) + list(videos or []))
        media_note = f" with {os.path.basename(media[0])}" if media else ""
        text = (f"[MOCK] Turn {n_prev + 1} reply{media_note} — prompt: {prompt!r}. "
                f"History carried: {len(session)} messages.")
        content = [{"type": "image", "image": m} for m in (images or [])]
        content += [{"type": "video", "video": m} for m in (videos or [])]
        content.append({"type": "text", "text": prompt})
        return {"text": text, "session": session + [
            {"role": "user", "content": content},
            {"role": "assistant", "content": text},
        ]}

    query = {
        "prompt": prompt,
        "images": list(images or []),
        "videos": list(videos or []),
        "generate_kwargs": {
            "max_new_tokens": int(max_new_tokens),
            "temperature": float(temperature),
            "top_p": float(top_p),
            "repetition_penalty": 1.0,
            "do_sample": bool(do_sample),
        },
    }
    if videos:
        query["media_kwargs"] = {"video_fps": float(video_fps), "max_frames": int(max_frames)}
    out = model.offline_batch_generate(processor, [query], session_states=[session])
    return {"text": out["results"][0]["text"], "session": out["session_states"][0]}


# --- UI glue ---

IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".webp", ".bmp", ".gif"}

MOCK_BADGE = "\n\n`MOCK mode — scripted model responses`"

# UI language packs, same switcher pattern as the live demo: radio choices are
# (label, value) pairs so switching language never fires change events.
LANGS = {
    "en": {
        "btn": "中文",
        "title": "# MOSS-VL-Instruct-0708\n\n"
                 "An 11B multimodal vision-language model for image and video understanding. "
                 "Attach an image or video (or none) and ask anything, turn by turn.",
        "view_chat": "💬 Chat",
        "view_raw": "{ } Raw",
        "ph": "Ask anything — attach an image/video for this turn, or none…",
        "adv": "Advanced Settings",
        "max_tokens": "Max New Tokens",
        "temperature": "Temperature",
        "top_p": "Top-p",
        "do_sample": "Do Sample",
        "fps": "Video FPS (video only)",
        "max_frames": "Max Frames (video only)",
    },
    "zh": {
        "btn": "English",
        "title": "# MOSS-VL-Instruct-0708\n\n"
                 "110 亿参数的多模态视觉语言模型，支持图片与视频理解。"
                 "每一轮可附带图片/视频（或不带媒体），随意提问。",
        "view_chat": "💬 对话",
        "view_raw": "{ } 原始",
        "ph": "随意提问——本轮可附带图片/视频，也可不带…",
        "adv": "高级设置",
        "max_tokens": "最大生成长度",
        "temperature": "温度",
        "top_p": "Top-p",
        "do_sample": "启用采样",
        "fps": "视频采样帧率（仅视频）",
        "max_frames": "最大帧数（仅视频）",
    },
}


def switch_lang(lang):
    lang = "zh" if lang == "en" else "en"
    L = LANGS[lang]
    return (
        lang,
        gr.update(value=L["btn"]),
        L["title"] + (MOCK_BADGE if MOCK else ""),
        gr.update(choices=[(L["view_chat"], "chat"), (L["view_raw"], "raw")]),
        gr.update(placeholder=L["ph"]),
        gr.update(label=L["adv"]),
        gr.update(label=L["max_tokens"]),
        gr.update(label=L["temperature"]),
        gr.update(label=L["top_p"]),
        gr.update(label=L["do_sample"]),
        gr.update(label=L["fps"]),
        gr.update(label=L["max_frames"]),
    )


def classify_media(path):
    return "image" if os.path.splitext(path)[1].lower() in IMAGE_EXTS else "video"


def _file_path(value):
    """Normalize a file entry (str | dict | FileData) to a path or None."""
    if isinstance(value, dict):
        value = value.get("path") or value.get("name")
    path = getattr(value, "path", value)
    return path or None


def run_analyze(msg, history, events, session,
                max_new_tokens, temperature, top_p, do_sample, video_fps, max_frames):
    """One MULTI-TURN chat turn: send whatever is staged — media + text, media
    alone (default prompt), or text alone (no images sent). The model sees the
    whole session (earlier turns and their media) via the session-state API."""
    msg = msg or {}
    text = (msg.get("text") or "").strip()
    files = [p for p in (_file_path(f) for f in (msg.get("files") or [])) if p]
    if not text and not files:
        gr.Warning("Type a question or attach an image/video · 请输入问题或添加图片/视频")
        yield (gr.skip(),) * 4
        return

    media = files[0] if files else None
    kind = classify_media(media) if media else "text"
    prompt = text or f"Describe this {kind} in detail."
    session = list(session or [])

    history = list(history or [])
    events = list(events or [])
    if media:
        history.append({"role": "user", "content": {"path": media}})
    if text:
        history.append({"role": "user", "content": text})
    history.append({"role": "assistant", "content": "⏳ Analyzing · 分析中…"})
    params = {
        "max_new_tokens": int(max_new_tokens),
        "temperature": float(temperature),
        "top_p": float(top_p),
        "do_sample": bool(do_sample),
    }
    if kind == "video":
        params.update(video_fps=float(video_fps), max_frames=int(max_frames))
    events.append({"event": "request", "kind": kind,
                   "media": os.path.basename(media) if media else None,
                   "prompt": prompt, "session_msgs": len(session), "params": params})
    yield history, events, {"text": "", "files": []}, gr.skip()

    t0 = time.monotonic()
    try:
        out = chat_generate(
            session, prompt,
            [media] if kind == "image" else [],
            [media] if kind == "video" else [],
            max_new_tokens, temperature, top_p, do_sample, video_fps, max_frames,
        )
    except Exception as exc:
        history[-1] = {"role": "assistant", "content": f"⚠️ {type(exc).__name__}: {exc}"}
        events.append({"event": "error", "message": f"{type(exc).__name__}: {exc}"})
        yield history, events, gr.skip(), gr.skip()
        return
    history[-1] = {"role": "assistant", "content": out["text"]}
    events.append({"event": "response", "elapsed_s": round(time.monotonic() - t0, 2), "text": out["text"]})
    yield history, events, gr.skip(), out["session"]


def clear_chat():
    return [], [], {"text": "", "files": []}, []


def toggle_view(choice):
    chat = choice == "chat"
    return gr.update(visible=chat), gr.update(visible=not chat)


CSS = """
#col-container { max-width: 1250px; margin: 0 auto; --console-h: 46px; --stage-h: 640px; }
#links-row { align-items: center; }
#links-md { flex-grow: 1; }
#lang-btn { flex-grow: 0 !important; min-width: 84px; }
/* full-width chat stage with a fixed height — content scrolls inside */
#message-stage {
  flex-grow: 0 !important; height: var(--stage-h); min-height: var(--stage-h); max-height: var(--stage-h);
  display: flex; flex-direction: column;
}
#message-stage > *, #message-stage .styler {
  flex-grow: 1 !important; display: flex; flex-direction: column; min-height: 0;
}
#chatbot, #raw-json {
  flex-grow: 1 !important; min-height: 0 !important; height: auto !important; max-height: none !important;
}
/* segmented chat/raw toggle: two half-width buttons, centered text, no radio
   dot — black = off, accent yellow = on (same as the live demo) */
#view-toggle { flex-grow: 0 !important; flex-shrink: 0 !important; min-height: fit-content; }
#view-toggle .wrap { display: flex; flex-direction: row; gap: 0; width: 100%; }
#view-toggle label {
  flex: 1 1 50%; justify-content: center; text-align: center; margin: 0;
  padding: var(--spacing-sm) 0; cursor: pointer; border-radius: 0;
  background: #141414; color: #9aa0a6; border: 1px solid var(--border-color-primary);
  transition: background 0.15s, color 0.15s;
}
#view-toggle label:first-child { border-radius: var(--radius-md) 0 0 var(--radius-md); }
#view-toggle label:last-child { border-radius: 0 var(--radius-md) var(--radius-md) 0; }
#view-toggle label.selected { background: var(--color-accent); color: #111; font-weight: 600; }
#view-toggle input[type="radio"] { display: none; }
/* input line INSIDE the message block: one divider line separates it from
   the chat — no second box, no gap */
#session-console {
  flex-grow: 0 !important; align-items: center; gap: 2px; margin: 0;
  border-top: 1px solid var(--border-color-primary);
  padding: 4px 8px; min-height: var(--console-h);
}
#session-console .form {
  flex: 1 1 auto; border: none !important; background: transparent !important; box-shadow: none !important;
}
/* the multimodal input renders bare — the surrounding block provides the box */
#prompt-box { background: transparent !important; border: none !important; box-shadow: none !important; }
#prompt-box .full-container { padding: 0 !important; }
/* the floating pending strip sits above the input line — nothing may clip it */
#prompt-box, #prompt-box .full-container, #prompt-box .input-container,
#session-console, #session-console .form { overflow: visible !important; }
#prompt-box .input-wrapper {
  position: relative; min-height: 38px;
  border: none; border-radius: 0; background: transparent;
  padding: 0 var(--spacing-xs);
}
/* pending media live in a floating strip at the BOTTOM OF THE MESSAGE BLOCK
   (just above the input bar), not inside the bar */
#prompt-box .thumbnails {
  position: absolute; bottom: calc(100% + 8px); left: 0; margin: 0;
  padding: 5px 8px; z-index: 30; width: auto; max-width: 100%;
  background: rgba(20, 20, 20, 0.55); backdrop-filter: blur(3px);
  border-radius: var(--radius-md);
}
#prompt-box .input-row { align-items: center; }
#prompt-box .input-row textarea {
  order: 1; padding: 6px 8px; line-height: 20px; resize: none;
  max-height: 34px; overflow-y: auto; scrollbar-width: none;
}
#prompt-box .input-row textarea::-webkit-scrollbar { display: none; width: 0; }
#prompt-box .input-row button { order: 2; }
/* pending thumbnails: whole tile opens the full-res file; the delete control
   shrinks to a corner badge so it doesn't swallow the click */
#prompt-box .thumbnails img { cursor: zoom-in; }
#prompt-box .thumbnail-wrapper .delete-button {
  inset: auto; top: -4px; right: -4px; width: 18px; height: 18px;
  border-radius: 50%; opacity: 1; padding: 0;
}
#prompt-box .thumbnail-wrapper .delete-button svg { width: 10px; height: 10px; }
/* Citrus dark mode leaves example chips near-black on dark background —
   follow the theme's body text color (white on black / black on white) */
.dark .gallery-item { color: var(--body-text-color) !important; }
/* send + clear-chat: flat line-form icon buttons, same look as the attach
   icon inside the input line */
#send-btn, #clear-chat-btn {
  width: 36px; min-width: 36px !important;
  height: 36px !important; min-height: 36px !important;
  border-radius: var(--radius-md); padding: 0; font-size: 1rem; flex-grow: 0 !important;
  background: transparent !important; border: none !important; box-shadow: none !important;
  color: var(--body-text-color) !important;
}
#send-btn:hover, #clear-chat-btn:hover { background: var(--button-secondary-background-fill) !important; }
/* scrollbars: no rail is drawn; the thumb is invisible until the pointer
   enters the rail strip at the scrollable edge (.sb-hot, set by the page
   script) or grabs the thumb itself. scrollbar-color drives modern
   Chrome/Firefox; the ::-webkit rules cover older WebKit. */
* { scrollbar-width: thin; scrollbar-color: transparent transparent; }
.sb-hot { scrollbar-color: rgba(128, 128, 128, 0.55) transparent !important; }
::-webkit-scrollbar { width: 10px; height: 10px; background: transparent !important; }
::-webkit-scrollbar-track, ::-webkit-scrollbar-corner { background: transparent !important; }
::-webkit-scrollbar-button { display: none !important; width: 0 !important; height: 0 !important; }
::-webkit-scrollbar-thumb { background: transparent; border-radius: 5px; }
.sb-hot::-webkit-scrollbar-thumb,
::-webkit-scrollbar-thumb:hover, ::-webkit-scrollbar-thumb:active { background: rgba(128, 128, 128, 0.6) !important; }
"""

# Page JS: (1) keep the raw event view pinned to the bottom while it grows,
# unless the user has scrolled up to read; (2) clicking a pending thumbnail
# opens the full-resolution file (the corner ✕ still deletes).
PAGE_JS = """
() => {
  if (!window.__mossSbHot) {
    window.__mossSbHot = true;
    let hot = null;
    document.addEventListener('mousemove', (e) => {
      let el = e.target instanceof Element ? e.target : null;
      let found = null;
      while (el && el !== document.documentElement) {
        const vs = el.scrollHeight > el.clientHeight + 4;
        const hs = el.scrollWidth > el.clientWidth + 4;
        const cs = vs || hs ? getComputedStyle(el) : null;
        if (cs && /(auto|scroll)/.test(cs.overflowY + cs.overflowX)) {
          const r = el.getBoundingClientRect();
          if ((vs && r.right - e.clientX <= 18) || (hs && r.bottom - e.clientY <= 18)) found = el;
          break;
        }
        el = el.parentElement;
      }
      if (found !== hot) {
        if (hot) hot.classList.remove('sb-hot');
        if (found) found.classList.add('sb-hot');
        hot = found;
      }
    }, {passive: true});
  }
  if (window.__mossInstructInit) return;
  window.__mossInstructInit = true;
  document.addEventListener('click', (e) => {
    if (e.target.closest('button')) return;
    const img = e.target.closest('#prompt-box .thumbnails img');
    if (img && img.src) window.open(img.src, '_blank');
  }, true);
  const timer = setInterval(() => {
    const rawRoot = document.querySelector('#raw-json');
    if (!rawRoot) return;
    clearInterval(timer);
    const obs = new MutationObserver(() => {
      if (rawRoot.offsetParent === null) return;
      const sc = [...rawRoot.querySelectorAll('*')].find(e => e.scrollHeight > e.clientHeight + 8);
      if (!sc) return;
      if (sc.scrollHeight - sc.scrollTop - sc.clientHeight < 240) sc.scrollTop = sc.scrollHeight;
    });
    obs.observe(rawRoot, {childList: true, subtree: true, characterData: true});
  }, 500);
}
"""

with gr.Blocks(title="MOSS-VL-Instruct Demo") as demo:
    EN = LANGS["en"]
    with gr.Column(elem_id="col-container"):
        title_md = gr.Markdown(EN["title"] + (MOCK_BADGE if MOCK else ""))
        with gr.Row(elem_id="links-row"):
            gr.Markdown(
                "[Model Card](https://huggingface.co/OpenMOSS-Team/MOSS-VL-Instruct-0708) | "
                "[GitHub](https://github.com/OpenMOSS/MOSS-VL)",
                elem_id="links-md",
            )
            lang_btn = gr.Button(EN["btn"], size="sm", scale=0, elem_id="lang-btn")

        with gr.Group(elem_id="message-stage"):
            view_toggle = gr.Radio(
                [(EN["view_chat"], "chat"), (EN["view_raw"], "raw")],
                value="chat",
                show_label=False,
                container=False,
                elem_id="view-toggle",
            )
            chatbot = gr.Chatbot(
                height=560,
                show_label=False,
                buttons=["copy"],
                autoscroll=True,
                elem_id="chatbot",
            )
            raw_json = gr.JSON(
                value=[], show_label=False, visible=False, height=560, elem_id="raw-json"
            )
            # input line lives INSIDE the same block, separated from the chat
            # by a single divider line; send/clear are flat line-form icons
            # matching the attach icon
            with gr.Row(elem_id="session-console"):
                prompt_box = gr.MultimodalTextbox(
                    scale=1,
                    show_label=False,
                    container=False,
                    file_types=["image", "video"],
                    file_count="single",
                    submit_btn=False,
                    placeholder=EN["ph"],
                    elem_id="prompt-box",
                )
                send_btn = gr.Button("➤", variant="secondary", scale=0, elem_id="send-btn")
                clear_chat_btn = gr.Button("🗑︎", variant="secondary", scale=0, elem_id="clear-chat-btn")

        with gr.Accordion(EN["adv"], open=False) as adv_acc:
            with gr.Row():
                max_new_tokens = gr.Slider(64, 2048, value=512, step=64, label=EN["max_tokens"])
                temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.1, label=EN["temperature"])
                top_p = gr.Slider(0.1, 1.0, value=1.0, step=0.05, label=EN["top_p"])
                do_sample = gr.Checkbox(label=EN["do_sample"], value=False)
            with gr.Row():
                video_fps = gr.Slider(0.5, 4.0, value=1.0, step=0.5, label=EN["fps"])
                max_frames = gr.Slider(8, 256, value=64, step=8, label=EN["max_frames"])

        gr.Examples(
            examples=[
                [{"text": "Extract the store name, waiter name, bill number, number of people, items purchased with their quantities and amounts, total amount, and print time from this receipt. Output in JSON format.", "files": ["example_bill.png"]}],
                [{"text": "Describe this image in detail.", "files": ["astronaut.jpg"]}],
                [{"text": "What species of bird is this? Describe its appearance and habitat.", "files": ["bird_kingfisher.jpg"]}],
                [{"text": "Describe what happens in this video.", "files": ["example_video.mp4"]}],
            ],
            inputs=[prompt_box],
            label="Examples",
        )

    ui_lang = gr.State("en")
    chat_session = gr.State([])  # model-side multi-turn message history

    lang_btn.click(
        switch_lang,
        [ui_lang],
        [ui_lang, lang_btn, title_md, view_toggle, prompt_box,
         adv_acc, max_new_tokens, temperature, top_p, do_sample, video_fps, max_frames],
        api_visibility="private",
    )
    view_toggle.change(toggle_view, [view_toggle], [chatbot, raw_json], api_visibility="private")

    analyze_inputs = [prompt_box, chatbot, raw_json, chat_session,
                      max_new_tokens, temperature, top_p, do_sample, video_fps, max_frames]
    analyze_outputs = [chatbot, raw_json, prompt_box, chat_session]
    send_btn.click(
        run_analyze, analyze_inputs, analyze_outputs,
        api_visibility="private", show_progress="hidden",
    )
    prompt_box.submit(
        run_analyze, analyze_inputs, analyze_outputs,
        api_visibility="private", show_progress="hidden",
    )
    clear_chat_btn.click(
        clear_chat, None, [chatbot, raw_json, prompt_box, chat_session],
        api_visibility="private",
    )

    gr.api(analyze_image, api_name="analyze_image")
    gr.api(analyze_video, api_name="analyze_video")
    gr.api(analyze_text, api_name="analyze_text")
    gr.api(chat_generate, api_name="chat")
    demo.load(None, None, None, js=PAGE_JS, api_visibility="private")


if __name__ == "__main__":
    # gradio 6.x: theme/css/mcp_server are launch() parameters
    demo.queue(max_size=32)
    demo.launch(theme=gr.themes.Citrus(), css=CSS, mcp_server=True)