app.py

import os
import random
import uuid
import json
import time
import asyncio
from threading import Thread
from io import BytesIO
from typing import Optional, Tuple, Dict, Any, Iterable
import fitz
import gradio as gr
import spaces
import torch
import numpy as np
from PIL import Image
import cv2
from transformers import (
    Qwen2_5_VLForConditionalGeneration,
    Qwen3VLForConditionalGeneration,
    AutoTokenizer,
    AutoProcessor,
    TextIteratorStreamer,
)
from transformers.image_utils import load_image
from gradio.themes import Soft
from gradio.themes.utils import colors, fonts, sizes

import shlex
import subprocess

subprocess.run(shlex.split("pip install flash-attn  --no-build-isolation"), env=os.environ | {"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"}, check=True)

MAX_MAX_NEW_TOKENS = 4096
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load Qwen3-VL-4B-Instruct
MODEL_ID_Q = "Qwen/Qwen3-VL-4B-Instruct"
processor_q = AutoProcessor.from_pretrained(MODEL_ID_Q, trust_remote_code=True)
model_q = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID_Q,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16).to(device).eval()

# Load Qwen3-VL-8B-Instruct
MODEL_ID_Y = "Qwen/Qwen3-VL-8B-Instruct"
processor_y = AutoProcessor.from_pretrained(MODEL_ID_Y, trust_remote_code=True)
model_y = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID_Y,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16).to(device).eval()

# Load Qwen3-VL-2B-Instruct
MODEL_ID_L = "Qwen/Qwen3-VL-2B-Instruct"
processor_l = AutoProcessor.from_pretrained(MODEL_ID_L, trust_remote_code=True)
model_l = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID_L,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16).to(device).eval()

# Load Qwen3-VL-2B-Thinking
MODEL_ID_J = "Qwen/Qwen3-VL-2B-Thinking"
processor_j = AutoProcessor.from_pretrained(MODEL_ID_J, trust_remote_code=True)
model_j = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID_J,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16).to(device).eval()

# Load Qwen3-VL-4B-Thinking
MODEL_ID_T = "Qwen/Qwen3-VL-4B-Thinking"
processor_t = AutoProcessor.from_pretrained(MODEL_ID_T, trust_remote_code=True)
model_t = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL_ID_T,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16).to(device).eval()

def convert_pdf_to_images(file_path: str, dpi: int = 128):
    if not file_path:
        return []
    images = []
    pdf_document = fitz.open(file_path)
    zoom = dpi / 72.0
    mat = fitz.Matrix(zoom, zoom)
    for page_num in range(len(pdf_document)):
        page = pdf_document.load_page(page_num)
        pix = page.get_pixmap(matrix=mat)
        img_data = pix.tobytes("png")
        images.append(Image.open(BytesIO(img_data)))
    pdf_document.close()
    return images

def get_initial_pdf_state() -> Dict[str, Any]:
    return {"pages": [], "total_pages": 0, "current_page_index": 0}

def load_and_preview_pdf(file_path: Optional[str]) -> Tuple[Optional[Image.Image], Dict[str, Any], str]:
    state = get_initial_pdf_state()
    if not file_path:
        return None, state, '<div style="text-align:center;">No file loaded</div>'
    try:
        pages = convert_pdf_to_images(file_path)
        if not pages:
            return None, state, '<div style="text-align:center;">Could not load file</div>'
        state["pages"] = pages
        state["total_pages"] = len(pages)
        page_info_html = f'<div style="text-align:center;">Page 1 / {state["total_pages"]}</div>'
        return pages[0], state, page_info_html
    except Exception as e:
        return None, state, f'<div style="text-align:center;">Failed to load preview: {e}</div>'

def navigate_pdf_page(direction: str, state: Dict[str, Any]):
    if not state or not state["pages"]:
        return None, state, '<div style="text-align:center;">No file loaded</div>'
    current_index = state["current_page_index"]
    total_pages = state["total_pages"]
    if direction == "prev":
        new_index = max(0, current_index - 1)
    elif direction == "next":
        new_index = min(total_pages - 1, current_index + 1)
    else:
        new_index = current_index
    state["current_page_index"] = new_index
    image_preview = state["pages"][new_index]
    page_info_html = f'<div style="text-align:center;">Page {new_index + 1} / {total_pages}</div>'
    return image_preview, state, page_info_html

def downsample_video(video_path, max_dim=720):
    vidcap = cv2.VideoCapture(video_path)
    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
    frames = []
    frame_indices = np.linspace(0, total_frames - 1, min(total_frames, 10), dtype=int)

    for i in frame_indices:
        vidcap.set(cv2.CAP_PROP_POS_FRAMES, i)
        success, image = vidcap.read()
        if success:
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

            h, w = image.shape[:2]
            scale = max_dim / max(h, w)
            if scale < 1:
                image = cv2.resize(image, (int(w*scale), int(h*scale)), interpolation=cv2.INTER_AREA)

            pil_image = Image.fromarray(image)
            frames.append(pil_image)

    vidcap.release()
    return frames


@spaces.GPU
def generate_image(model_name: str, text: str, image: Image.Image,
                   max_new_tokens: int = 1024,
                   temperature: float = 0.6,
                   top_p: float = 0.9,
                   top_k: int = 50,
                   repetition_penalty: float = 1.2):
    """
    Generates responses using the selected model for image input.
    """
    if model_name == "Qwen3-VL-4B-Instruct":
        processor, model = processor_q, model_q
    elif model_name == "Qwen3-VL-8B-Instruct":
        processor, model = processor_y, model_y
    elif model_name == "Qwen3-VL-4B-Thinking":
        processor, model = processor_t, model_t
    elif model_name == "Qwen3-VL-2B-Instruct":
        processor, model = processor_l, model_l
    elif model_name == "Qwen3-VL-2B-Thinking":
        processor, model = processor_j, model_j
    else:
        yield "Invalid model selected.", "Invalid model selected."
        return
    if image is None:
        yield "Please upload an image.", "Please upload an image."
        return
    messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": text}]}]
    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(
        text=[prompt_full], images=[image], return_tensors="pt", padding=True).to(device)
    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        time.sleep(0.01)
        yield buffer, buffer

@spaces.GPU(duration=180)
def generate_video(model_name: str, text: str, video_path: str,
                   max_new_tokens: int = 1024,
                   temperature: float = 0.6,
                   top_p: float = 0.9,
                   top_k: int = 50,
                   repetition_penalty: float = 1.2):
    """
    Generates responses using the selected model for video input.
    """
    if model_name == "Qwen3-VL-4B-Instruct":
        processor, model = processor_q, model_q
    elif model_name == "Qwen3-VL-8B-Instruct":
        processor, model = processor_y, model_y
    elif model_name == "Qwen3-VL-4B-Thinking":
        processor, model = processor_t, model_t
    elif model_name == "Qwen3-VL-2B-Instruct":
        processor, model = processor_l, model_l
    elif model_name == "Qwen3-VL-2B-Thinking":
        processor, model = processor_j, model_j
    else:
        yield "Invalid model selected.", "Invalid model selected."
        return
    if video_path is None:
        yield "Please upload a video.", "Please upload a video."
        return
    frames = downsample_video(video_path)
    if not frames:
        yield "Could not process video.", "Could not process video."
        return
    messages = [{"role": "user", "content": [{"type": "text", "text": text}]}]
    images_for_processor = []
    for frame in frames:
        messages[0]["content"].append({"type": "image"})
        images_for_processor.append(frame)
    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(
        text=[prompt_full], images=images_for_processor, return_tensors="pt", padding=True).to(device)
    streamer = TextIteratorStreamer(processor, 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.generate, kwargs=generation_kwargs)
    thread.start()
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        buffer = buffer.replace("<|im_end|>", "")
        time.sleep(0.01)
        yield buffer, buffer

@spaces.GPU(duration=180)
def generate_pdf(model_name: str, text: str, state: Dict[str, Any], 
                 max_new_tokens: int = 2048, 
                 temperature: float = 0.6, 
                 top_p: float = 0.9, 
                 top_k: int = 50, 
                 repetition_penalty: float = 1.2):
    
    if model_name == "Qwen3-VL-4B-Instruct":
        processor, model = processor_q, model_q
    elif model_name == "Qwen3-VL-8B-Instruct":
        processor, model = processor_y, model_y
    elif model_name == "Qwen3-VL-4B-Thinking":
        processor, model = processor_t, model_t
    elif model_name == "Qwen3-VL-2B-Instruct":
        processor, model = processor_l, model_l
    elif model_name == "Qwen3-VL-2B-Thinking":
        processor, model = processor_j, model_j
    else:
        yield "Invalid model selected.", "Invalid model selected."
        return

    if not state or not state["pages"]:
        yield "Please upload a PDF file first.", "Please upload a PDF file first."
        return
    
    page_images = state["pages"]

    messages = [{"role": "user", "content": [{"type": "text", "text": text}]}]
    images_for_processor = []
    for frame in page_images:
        messages[0]["content"].append({"type": "image"})
        images_for_processor.append(frame)

    prompt_full = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    
    inputs = processor(
        text=[prompt_full], 
        images=images_for_processor,  # Truyền cả list ảnh
        return_tensors="pt", 
        padding=True
    ).to(device)
    
    streamer = TextIteratorStreamer(processor, 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.generate, kwargs=generation_kwargs)
    thread.start()
    
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        buffer = buffer.replace("<|im_end|>", "") # Thêm dòng này giống video
        yield buffer, buffer
        time.sleep(0.01)

image_examples = [
    ["Explain the content in detail.", "images/force.jpg"],
    ["Explain the content (ocr).", "images/ocr.jpg"],
    ["Extract the content in the json format", "images/bill.jpg"],
    ["Choose the right answer .", "images/math.jpg"],
]
video_examples = [
    ["Explain the ad in detail", "videos/1.mp4"],
    ["Identify the main actions in the video", "videos/2.mp4"],
]
pdf_examples = [
    ["Extract the content precisely.", "pdfs/doc1.pdf"],
    ["Nội dung của văn bản trong ảnh là gì?.", "pdfs/doc2.pdf"]
]

css = """
#main-title h1 {
    font-size: 2.3em !important;
}
#output-title h2 {
    font-size: 2.1em !important;
}
"""

with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
    
    pdf_state = gr.State(value=get_initial_pdf_state())
    
    gr.Markdown("# 🎉**Qwen3-VL-Demo**🎉", elem_id="main-title")
    with gr.Row():
        with gr.Column(scale=2):
            with gr.Tabs():
                with gr.TabItem("Image Inference"):
                    image_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                    image_upload = gr.Image(type="pil", label="Upload Image", height=290)
                    image_submit = gr.Button("Submit", variant="primary")
                    gr.Examples(examples=image_examples, inputs=[image_query, image_upload])
                
                with gr.TabItem("Video Inference"):
                    video_query = gr.Textbox(label="Query Input", placeholder="Enter your query here...")
                    video_upload = gr.Video(label="Upload Video", height=290)
                    video_submit = gr.Button("Submit", variant="primary")
                    gr.Examples(examples=video_examples, inputs=[video_query, video_upload])
                
                with gr.TabItem("PDF Inference"):
                    with gr.Row():
                        with gr.Column(scale=1):
                            pdf_query = gr.Textbox(label="Query Input", placeholder="e.g., 'Summarize this document'")
                            pdf_upload = gr.File(label="Upload PDF", file_types=[".pdf"])
                            pdf_submit = gr.Button("Submit", variant="primary")
                        with gr.Column(scale=1):
                            pdf_preview_img = gr.Image(label="PDF Preview", height=290)
                            with gr.Row():
                                prev_page_btn = gr.Button("◀ Previous")
                                page_info = gr.HTML('<div style="text-align:center;">No file loaded</div>')
                                next_page_btn = gr.Button("Next ▶")
                    gr.Examples(examples=pdf_examples, inputs=[pdf_query, pdf_upload])

            with gr.Accordion("Advanced options", open=False):
                max_new_tokens = gr.Slider(label="Max new tokens", minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1, value=DEFAULT_MAX_NEW_TOKENS)
                temperature = gr.Slider(label="Temperature", minimum=0.1, maximum=4.0, step=0.1, value=0.6)
                top_p = gr.Slider(label="Top-p (nucleus sampling)", minimum=0.05, maximum=1.0, step=0.05, value=0.9)
                top_k = gr.Slider(label="Top-k", minimum=1, maximum=1000, step=1, value=50)
                repetition_penalty = gr.Slider(label="Repetition penalty", minimum=1.0, maximum=2.0, step=0.05, value=1.2)
        
        with gr.Column(scale=3):
            gr.Markdown("## Output", elem_id="output-title")
            output = gr.Textbox(label="Raw Output Stream", interactive=False, lines=14, show_copy_button=True)
            with gr.Accordion("(Result.md)", open=False):
                markdown_output = gr.Markdown(latex_delimiters=[
                                {"left": "$$", "right": "$$", "display": True},
                                {"left": "$", "right": "$", "display": False}
                            ])
            
            model_choice = gr.Radio(
                choices=["Qwen3-VL-4B-Instruct", "Qwen3-VL-8B-Instruct", "Qwen3-VL-2B-Instruct", "Qwen3-VL-2B-Thinking", "Qwen3-VL-4B-Thinking"],
                label="Select Model",
                value="Qwen3-VL-4B-Instruct"
            )

    image_submit.click(
        fn=generate_image,
        inputs=[model_choice, image_query, image_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=[output, markdown_output]
    )
    
    video_submit.click(
        fn=generate_video,
        inputs=[model_choice, video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=[output, markdown_output]
    )
    
    pdf_submit.click(
        fn=generate_pdf,
        inputs=[model_choice, pdf_query, pdf_state, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=[output, markdown_output]
    )
    
    pdf_upload.change(
        fn=load_and_preview_pdf, 
        inputs=[pdf_upload], 
        outputs=[pdf_preview_img, pdf_state, page_info]
    )
    
    prev_page_btn.click(
        fn=lambda s: navigate_pdf_page("prev", s), 
        inputs=[pdf_state], 
        outputs=[pdf_preview_img, pdf_state, page_info]
    )
    
    next_page_btn.click(
        fn=lambda s: navigate_pdf_page("next", s), 
        inputs=[pdf_state], 
        outputs=[pdf_preview_img, pdf_state, page_info]
    )


if __name__ == "__main__":
    demo.queue(max_size=50).launch(mcp_server=True, ssr_mode=False, show_error=True)