Update app.py

app.py

@@ -1,313 +1,329 @@
 import os
-import gradio as gr
+import random
+import uuid
 import json
+import time
+import asyncio
+import re
+from threading import Thread
+
+import gradio as gr
 import spaces
-import logging
 import torch
+import numpy as np
 from PIL import Image
-import random
-import time
-from hi_diffusers import HiDreamImagePipeline
-from hi_diffusers import HiDreamImageTransformer2DModel
-from hi_diffusers.schedulers.fm_solvers_unipc import FlowUniPCMultistepScheduler
-from hi_diffusers.schedulers.flash_flow_match import FlashFlowMatchEulerDiscreteScheduler
-from transformers import LlamaForCausalLM, PreTrainedTokenizerFast
-from huggingface_hub import ModelCard
-
-# Constants
-MODEL_PREFIX = "HiDream-ai"
-LLAMA_MODEL_NAME = "meta-llama/Meta-Llama-3.1-8B-Instruct"
-
-FAST_MODEL_CONFIG = {
-        "path": f"{MODEL_PREFIX}/HiDream-I1-Full",
-        "guidance_scale": 5.0,
-        "num_inference_steps": 50,
-        "shift": 3.0,
-        "scheduler": FlowUniPCMultistepScheduler
-}
-
-RESOLUTION_OPTIONS = [
-    "1024 × 1024 (Square)",
-    "768 × 1360 (Portrait)",
-    "1360 × 768 (Landscape)",
-    "880 × 1168 (Portrait)",
-    "1168 × 880 (Landscape)",
-    "1248 × 832 (Landscape)",
-    "832 × 1248 (Portrait)"
-]
-
-# Load LoRAs from JSON file (assumed to be compatible with Hi-Dream)
-with open('loras.json', 'r') as f:
-    loras = json.load(f)
+import cv2
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-MAX_SEED = 2**32 - 1
+from transformers import (
+    AutoProcessor,
+    Gemma3ForConditionalGeneration,
+    Qwen2VLForConditionalGeneration,
+    TextIteratorStreamer,
+)
+from transformers.image_utils import load_image
 
-# Parse resolution string to height and width
-def parse_resolution(res_str):
-    mapping = {
-        "1024 × 1024": (1024, 1024),
-        "768 × 1360": (768, 1360),
-        "1360 × 768": (1360, 768),
-        "880 × 1168": (880, 1168),
-        "1168 × 880": (1168, 880),
-        "1248 × 832": (1248, 832),
-        "832 × 1248": (832, 1248)
-    }
-    for key, (h, w) in mapping.items():
-        if key in res_str:
-            return h, w
-    return 1024, 1024  # fallback
+# Constants
+MAX_MAX_NEW_TOKENS = 2048
+DEFAULT_MAX_NEW_TOKENS = 1024
+MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
+MAX_SEED = np.iinfo(np.int32).max
 
-# Load the Hi-Dream Fast Model pipeline
-pipe, MODEL_CONFIG = None, None
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
-def load_fast_model():
-    global pipe, MODEL_CONFIG
-    config = FAST_MODEL_CONFIG
-    scheduler = config["scheduler"](
-        num_train_timesteps=1000,
-        shift=config["shift"],
-        use_dynamic_shifting=False
-    )
+# Helper function to return a progress bar HTML snippet.
+def progress_bar_html(label: str) -> str:
+    return f'''
+<div style="display: flex; align-items: center;">
+    <span style="margin-right: 10px; font-size: 14px;">{label}</span>
+    <div style="width: 110px; height: 5px; background-color: #F0FFF0; border-radius: 2px; overflow: hidden;">
+        <div style="width: 100%; height: 100%; background-color: #00FF00; animation: loading 1.5s linear infinite;"></div>
+    </div>
+</div>
+<style>
+@keyframes loading {{
+    0% {{ transform: translateX(-100%); }}
+    100% {{ transform: translateX(100%); }}
+}}
+</style>
+    '''
 
-    tokenizer = PreTrainedTokenizerFast.from_pretrained(
-        LLAMA_MODEL_NAME,
-        use_fast=False
-    )
-    text_encoder = LlamaForCausalLM.from_pretrained(
-        LLAMA_MODEL_NAME,
-        output_hidden_states=True,
-        output_attentions=True,
-        torch_dtype=torch.bfloat16
-    ).to(device)
+# Qwen2-VL (for optional image inference)
 
-    transformer = HiDreamImageTransformer2DModel.from_pretrained(
-        config["path"],
-        subfolder="transformer",
-        torch_dtype=torch.bfloat16
-    ).to(device)
+MODEL_ID_VL = "prithivMLmods/Qwen2-VL-OCR-2B-Instruct" 
+processor = AutoProcessor.from_pretrained(MODEL_ID_VL, trust_remote_code=True)
+model_m = Qwen2VLForConditionalGeneration.from_pretrained(
+    MODEL_ID_VL,
+    trust_remote_code=True,
+    torch_dtype=torch.float16
+).to("cuda").eval()
 
-    pipe = HiDreamImagePipeline.from_pretrained(
-        config["path"],
-        scheduler=scheduler,
-        tokenizer_4=tokenizer,
-        text_encoder_4=text_encoder,
-        torch_dtype=torch.bfloat16
-    ).to(device, torch.bfloat16)
+def clean_chat_history(chat_history):
+    cleaned = []
+    for msg in chat_history:
+        if isinstance(msg, dict) and isinstance(msg.get("content"), str):
+            cleaned.append(msg)
+    return cleaned
 
-    pipe.transformer = transformer
-    MODEL_CONFIG = config
-    return pipe, config
+bad_words = json.loads(os.getenv('BAD_WORDS', "[]"))
+bad_words_negative = json.loads(os.getenv('BAD_WORDS_NEGATIVE', "[]"))
+default_negative = os.getenv("default_negative", "")
 
-# Generate image
-@spaces.GPU
-def generate_image(prompt, resolution, seed, guidance_scale, num_inference_steps):
-    global pipe, MODEL_CONFIG
-    if pipe is None:
-        pipe, MODEL_CONFIG = load_fast_model()
+def check_text(prompt, negative=""):
+    for i in bad_words:
+        if i in prompt:
+            return True
+    for i in bad_words_negative:
+        if i in negative:
+            return True
+    return False
 
-    height, width = parse_resolution(resolution)
-    if seed == -1 or seed is None:
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
         seed = random.randint(0, MAX_SEED)
-    generator = torch.Generator(device=device).manual_seed(int(seed))
+    return seed
 
-    result = pipe(
-        prompt=prompt,
-        height=height,
-        width=width,
-        guidance_scale=guidance_scale,
-        num_inference_steps=num_inference_steps,
-        num_images_per_prompt=1,
-        generator=generator
-    )
+CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES", "0") == "1"
+MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "2048"))
+USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
+ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
 
-    return result.images[0], seed
-
-class calculateDuration:
-    def __init__(self, activity_name=""):
-        self.activity_name = activity_name
-
-    def __enter__(self):
-        self.start_time = time.time()
-        return self
-    
-    def __exit__(self, exc_type, exc_value, traceback):
-        self.end_time = time.time()
-        self.elapsed_time = self.end_time - self.start_time
-        if self.activity_name:
-            print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
-        else:
-            print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
+dtype = torch.float16 if device.type == "cuda" else torch.float32
 
-def update_selection(evt: gr.SelectData, resolution):
-    selected_lora = loras[evt.index]
-    new_placeholder = f"Type a prompt for {selected_lora['title']}"
-    lora_repo = selected_lora["repo"]
-    updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
-    if "aspect" in selected_lora:
-        if selected_lora["aspect"] == "portrait":
-            resolution = "768 × 1360 (Portrait)"
-        elif selected_lora["aspect"] == "landscape":
-            resolution = "1360 × 768 (Landscape)"
-        else:
-            resolution = "1024 × 1024 (Square)"
-    return (
-        gr.update(placeholder=new_placeholder),
-        updated_text,
-        evt.index,
-        resolution,
-    )
 
-def run_lora(prompt, resolution, cfg_scale, steps, selected_index, randomize_seed, seed):
-    global pipe
-    if pipe is None:
-        pipe, _ = load_fast_model()
+# Gemma3 Model (default for text, image, & video inference)
 
-    if selected_index is not None:
-        selected_lora = loras[selected_index]
-        lora_path = selected_lora["repo"]
-        weight_name = selected_lora.get("weights", None)
-        with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
-            pipe.load_lora_weights(lora_path, weight_name=weight_name, low_cpu_mem_usage=True)
-        trigger_word = selected_lora.get("trigger_word", "")
-        if trigger_word:
-            if "trigger_position" in selected_lora and selected_lora["trigger_position"] == "prepend":
-                prompt = f"{trigger_word} {prompt}"
-            else:
-                prompt = f"{prompt} {trigger_word}"
+gemma3_model_id = "google/gemma-3-4b-it" #[or] Duplicate the space to use 12b
+gemma3_model = Gemma3ForConditionalGeneration.from_pretrained(
+    gemma3_model_id, device_map="auto"
+).eval()
+gemma3_processor = AutoProcessor.from_pretrained(gemma3_model_id)
 
-    if randomize_seed:
-        seed = random.randint(0, MAX_SEED)
+# VIDEO PROCESSING HELPER
 
-    with calculateDuration("Generating image"):
-        final_image, used_seed = generate_image(prompt, resolution, seed, cfg_scale, steps)
-    return final_image, used_seed
+def downsample_video(video_path):
+    vidcap = cv2.VideoCapture(video_path)
+    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
+    fps = vidcap.get(cv2.CAP_PROP_FPS)
+    frames = []
+    # Sample 10 evenly spaced frames.
+    frame_indices = np.linspace(0, total_frames - 1, 10, dtype=int)
+    for i in frame_indices:
+        vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
+        success, image = vidcap.read()
+        if success:
+            # Convert from BGR to RGB and then to PIL Image.
+            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+            pil_image = Image.fromarray(image)
+            timestamp = round(i / fps, 2)
+            frames.append((pil_image, timestamp))
+    vidcap.release()
+    return frames
 
-def check_custom_model(link):
-    split_link = link.split("/")
-    if len(split_link) != 2:
-        raise Exception("Invalid Hugging Face repository link format.")
-    model_card = ModelCard.load(link)
-    image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
-    trigger_word = model_card.data.get("instance_prompt", "")
-    image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
-    safetensors_name = None  # Simplified; assumes a safetensors file exists
-    return split_link[1], link, safetensors_name, trigger_word, image_url
+# MAIN GENERATION FUNCTION
 
-def add_custom_lora(custom_lora):
-    global loras
-    if custom_lora:
-        try:
-            title, repo, path, trigger_word, image = check_custom_model(custom_lora)
-            card = f'''
-            <div class="custom_lora_card">
-              <span>Loaded custom LoRA:</span>
-              <div class="card_internal">
-                <img src="{image}" />
-                <div>
-                    <h3>{title}</h3>
-                    <small>{"Using: <code><b>"+trigger_word+"</code></b> as the trigger word" if trigger_word else "No trigger word found."}</small>
-                </div>
-              </div>
-            </div>
-            '''
-            existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
-            if not existing_item_index:
-                new_item = {
-                    "image": image,
-                    "title": title,
-                    "repo": repo,
-                    "weights": path,
-                    "trigger_word": trigger_word
-                }
-                existing_item_index = len(loras)
-                loras.append(new_item)
-            return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
-        except Exception as e:
-            gr.Warning(f"Invalid LoRA: {str(e)}")
-            return gr.update(visible=True, value=f"Invalid LoRA: {str(e)}"), gr.update(visible=True), gr.update(), "", None, ""
-    return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
+@spaces.GPU
+def generate(
+    input_dict: dict,
+    chat_history: list[dict],
+    max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
+    temperature: float = 0.6,
+    top_p: float = 0.9,
+    top_k: int = 50,
+    repetition_penalty: float = 1.2,
+):
+    text = input_dict["text"]
+    files = input_dict.get("files", [])
+    lower_text = text.lower().strip()
 
-def remove_custom_lora():
-    return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
+    # ----- Qwen2-VL branch (triggered with @qwen2-vl) -----
+    if lower_text.startswith("@qwen2-vl"):
+        prompt_clean = re.sub(r"@qwen2-vl", "", text, flags=re.IGNORECASE).strip().strip('"')
+        if files:
+            images = [load_image(f) for f in files]
+            messages = [{
+                "role": "user",
+                "content": [
+                    *[{"type": "image", "image": image} for image in images],
+                    {"type": "text", "text": prompt_clean},
+                ]
+            }]
+            prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+            inputs = processor(text=[prompt], images=images, return_tensors="pt", padding=True).to("cuda")
+        else:
+            messages = [
+                {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+                {"role": "user", "content": [{"type": "text", "text": prompt_clean}]}
+            ]
+            inputs = processor.apply_chat_template(
+                messages, add_generation_prompt=True, tokenize=True,
+                return_dict=True, return_tensors="pt"
+            ).to("cuda", dtype=torch.float16)
+        streamer = TextIteratorStreamer(processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
+        generation_kwargs = {
+            **inputs,
+            "streamer": streamer,
+            "max_new_tokens": max_new_tokens,
+            "do_sample": True,
+            "temperature": temperature,
+            "top_p": top_p,
+            "top_k": top_k,
+            "repetition_penalty": repetition_penalty,
+        }
+        thread = Thread(target=model_m.generate, kwargs=generation_kwargs)
+        thread.start()
+        buffer = ""
+        yield progress_bar_html("Processing with Qwen2VL")
+        for new_text in streamer:
+            buffer += new_text
+            buffer = buffer.replace("<|im_end|>", "")
+            time.sleep(0.01)
+            yield buffer
+        return
 
-css = '''
-#gen_btn{height: 100%}
-#gen_column{align-self: stretch}
-#title{text-align: center}
-#title h1{font-size: 3em; display:inline-flex; align-items:center}
-#title img{width: 100px; margin-right: 0.5em}
-#gallery .grid-wrap{height: 10vh}
-#lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%}
-.card_internal{display: flex;height: 100px;margin-top: .5em}
-.card_internal img{margin-right: 1em}
-.styler{--form-gap-width: 0px !important}
-'''
+    # ----- Default branch: Gemma3 (for text, image, & video inference) -----
+    if files:
+        # Check if any provided file is a video based on extension.
+        video_extensions = (".mp4", ".mov", ".avi", ".mkv", ".webm")
+        if any(str(f).lower().endswith(video_extensions) for f in files):
+            # Video inference branch.
+            prompt_clean = re.sub(r"@video-infer", "", text, flags=re.IGNORECASE).strip().strip('"')
+            video_path = files[0]
+            frames = downsample_video(video_path)
+            messages = [
+                {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+                {"role": "user", "content": [{"type": "text", "text": prompt_clean}]}
+            ]
+            # Append each frame (with its timestamp) to the conversation.
+            for frame in frames:
+                image, timestamp = frame
+                image_path = f"video_frame_{uuid.uuid4().hex}.png"
+                image.save(image_path)
+                messages[1]["content"].append({"type": "text", "text": f"Frame {timestamp}:"})
+                messages[1]["content"].append({"type": "image", "url": image_path})
+            inputs = gemma3_processor.apply_chat_template(
+                messages, add_generation_prompt=True, tokenize=True,
+                return_dict=True, return_tensors="pt"
+            ).to(gemma3_model.device, dtype=torch.bfloat16)
+            streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
+            generation_kwargs = {
+                **inputs,
+                "streamer": streamer,
+                "max_new_tokens": max_new_tokens,
+                "do_sample": True,
+                "temperature": temperature,
+                "top_p": top_p,
+                "top_k": top_k,
+                "repetition_penalty": repetition_penalty,
+            }
+            thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)
+            thread.start()
+            buffer = ""
+            yield progress_bar_html("Processing video with Gemma3")
+            for new_text in streamer:
+                buffer += new_text
+                time.sleep(0.01)
+                yield buffer
+            return
+        else:
+            # Image inference branch.
+            prompt_clean = re.sub(r"@gemma3", "", text, flags=re.IGNORECASE).strip().strip('"')
+            images = [load_image(f) for f in files]
+            messages = [{
+                "role": "user",
+                "content": [
+                    *[{"type": "image", "image": image} for image in images],
+                    {"type": "text", "text": prompt_clean},
+                ]
+            }]
+            inputs = gemma3_processor.apply_chat_template(
+                messages, tokenize=True, add_generation_prompt=True,
+                return_dict=True, return_tensors="pt"
+            ).to(gemma3_model.device, dtype=torch.bfloat16)
+            streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
+            generation_kwargs = {
+                **inputs,
+                "streamer": streamer,
+                "max_new_tokens": max_new_tokens,
+                "do_sample": True,
+                "temperature": temperature,
+                "top_p": top_p,
+                "top_k": top_k,
+                "repetition_penalty": repetition_penalty,
+            }
+            thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)
+            thread.start()
+            buffer = ""
+            yield progress_bar_html("Processing with Gemma3")
+            for new_text in streamer:
+                buffer += new_text
+                time.sleep(0.01)
+                yield buffer
+            return
+    else:
+        # Text-only inference branch.
+        messages = [
+            {"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
+            {"role": "user", "content": [{"type": "text", "text": text}]}
+        ]
+        inputs = gemma3_processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True,
+            return_dict=True, return_tensors="pt"
+        ).to(gemma3_model.device, dtype=torch.bfloat16)
+        streamer = TextIteratorStreamer(gemma3_processor.tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
+        generation_kwargs = {
+            **inputs,
+            "streamer": streamer,
+            "max_new_tokens": max_new_tokens,
+            "do_sample": True,
+            "temperature": temperature,
+            "top_p": top_p,
+            "top_k": top_k,
+            "repetition_penalty": repetition_penalty,
+        }
+        thread = Thread(target=gemma3_model.generate, kwargs=generation_kwargs)
+        thread.start()
+        outputs = []
+        for new_text in streamer:
+            outputs.append(new_text)
+            yield "".join(outputs)
+        final_response = "".join(outputs)
+        yield final_response
 
-font = [gr.themes.GoogleFont("Source Sans Pro"), "Arial", "sans-serif"]
-with gr.Blocks(theme=gr.themes.Soft(font=font), css=css, delete_cache=(60, 60)) as app:
-    title = gr.HTML(
-        """<h1>Hi-Dream Full LoRA DLC 🤩</h1>""",
-        elem_id="title",
-    )
-    selected_index = gr.State(None)
-    with gr.Row():
-        with gr.Column(scale=3):
-            prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Type a prompt after selecting a LoRA")
-        with gr.Column(scale=1, elem_id="gen_column"):
-            generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
-    with gr.Row():
-        with gr.Column():
-            selected_info = gr.Markdown("")
-            gallery = gr.Gallery(
-                [(item["image"], item["title"]) for item in loras],
-                label="LoRA Gallery",
-                allow_preview=False,
-                columns=3,
-                elem_id="gallery",
-                show_share_button=False
-            )
-            with gr.Group():
-                custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="linoyts/HiDream-yarn-art-LoRA")
-                gr.Markdown("[Check the list of Hi-Dream LoRAs]", elem_id="lora_list")
-            custom_lora_info = gr.HTML(visible=False)
-            custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
-        with gr.Column():
-            result = gr.Image(label="Generated Image")
 
-    with gr.Row():
-        with gr.Accordion("Advanced Settings", open=False):
-            cfg_scale = gr.Slider(label="Guidance Scale", minimum=0, maximum=20, step=0.1, value=FAST_MODEL_CONFIG["guidance_scale"])
-            steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=FAST_MODEL_CONFIG["num_inference_steps"])
-            resolution = gr.Radio(
-                choices=RESOLUTION_OPTIONS,
-                value=RESOLUTION_OPTIONS[0],
-                label="Resolution"
-            )
-            randomize_seed = gr.Checkbox(True, label="Randomize seed")
-            seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
+# Gradio Interface
 
-    gallery.select(
-        update_selection,
-        inputs=[resolution],
-        outputs=[prompt, selected_info, selected_index, resolution]
-    )
-    custom_lora.input(
-        add_custom_lora,
-        inputs=[custom_lora],
-        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt]
-    )
-    custom_lora_button.click(
-        remove_custom_lora,
-        outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
-    )
-    gr.on(
-        triggers=[generate_button.click, prompt.submit],
-        fn=run_lora,
-        inputs=[prompt, resolution, cfg_scale, steps, selected_index, randomize_seed, seed],
-        outputs=[result, seed]
-    )
+demo = gr.ChatInterface(
+    fn=generate,
+    additional_inputs=[
+        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=0.1, maximum=4.0, step=0.1, value=0.6),
+        gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9),
+        gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50),
+        gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2),
+    ],
+    examples=[
+        [{"text": "Create a short story based on the image.","files": ["examples/1111.jpg"]}],
+        [{"text": "Explain the Image", "files": ["examples/3.jpg"]}],
+        [{"text": "Explain the content of the Advertisement", "files": ["examples/videoplayback.mp4"]}],
+        [{"text": "Which movie character is this?", "files": ["examples/9999.jpg"]}],
+        ["Explain Critical Temperature of Substance"],
+        [{"text": "@qwen2-vl Transcription of the letter", "files": ["examples/222.png"]}],
+        [{"text": "Explain the content of the video in detail", "files": ["examples/breakfast.mp4"]}],
+        [{"text": "Describe the video", "files": ["examples/Missing.mp4"]}],
+        [{"text": "Explain what is happening in this video ?", "files": ["examples/oreo.mp4"]}],
+        [{"text": "Summarize the events in this video", "files": ["examples/sky.mp4"]}],
+        [{"text": "What is in the video ?", "files": ["examples/redlight.mp4"]}],
+        ["Python Program for Array Rotation"],
+        ["Explain Critical Temperature of Substance"]
+    ],
+    cache_examples=False,
+    type="messages",
+    description="# **Gemma 3 Multimodal** \n`Use @qwen2-vl to switch to Qwen2-VL OCR for image inference and @video-infer for video input`",
+    fill_height=True,
+    textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image", "video"], file_count="multiple", placeholder="Tag with @qwen2-vl for Qwen2-VL inference if needed."),
+    stop_btn="Stop Generation",
+    multimodal=True,
+)
 
-app.queue()
-app.launch()
+if __name__ == "__main__":
+    demo.queue(max_size=20).launch(share=True)