app.py

import os
import random
import uuid
import json
import time
import asyncio
from threading import Thread

import gradio as gr
import spaces
import torch
import numpy as np
from PIL import Image
import cv2
import requests

from transformers import (
    Qwen3VLMoeForConditionalGeneration,
    AutoProcessor,
    TextIteratorStreamer,
)
from transformers.image_utils import load_image

# Constants for text generation
MAX_MAX_NEW_TOKENS = 4096
DEFAULT_MAX_NEW_TOKENS = 2048
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

# Let the environment (e.g., Hugging Face Spaces) determine the device.
# This avoids conflicts with the CUDA environment setup by the platform.
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print("CUDA_VISIBLE_DEVICES=", os.environ.get("CUDA_VISIBLE_DEVICES"))
print("torch.__version__ =", torch.__version__)
print("torch.version.cuda =", torch.version.cuda)
print("cuda available:", torch.cuda.is_available())
print("cuda device count:", torch.cuda.device_count())
if torch.cuda.is_available():
    print("current device:", torch.cuda.current_device())
    print("device name:", torch.cuda.get_device_name(torch.cuda.current_device()))

print("Using device:", device)
# --- Model Loading ---

# To address the warnings, we add `use_fast=False` to ensure we use the
# processor version the model was originally saved with.

# Load Qwen3VL
MODEL_ID_Q3VL = "Qwen/Qwen3-VL-30B-A3B-Instruct"
processor_q3vl = AutoProcessor.from_pretrained(MODEL_ID_Q3VL, trust_remote_code=True, use_fast=False)
model_q3vl = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    MODEL_ID_Q3VL,
    trust_remote_code=True,
    torch_dtype=torch.float16
).to(device).eval()


def downsample_video(video_path):
    """
    Downsamples the video to evenly spaced frames.
    Each frame is returned as a PIL image along with its timestamp.
    """
    vidcap = cv2.VideoCapture(video_path)
    total_frames = int(vidcap.get(cv2.CAP_PROP_FRAME_COUNT))
    fps = vidcap.get(cv2.CAP_PROP_FPS)
    frames = []
    # Use a maximum of 10 frames to avoid excessive memory usage
    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)
            pil_image = Image.fromarray(image)
            timestamp = round(i / fps, 2)
            frames.append((pil_image, timestamp))
    vidcap.release()
    return frames

@spaces.GPU
def generate_image(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 Qwen3-VL model for image input.
    """
    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_q3vl.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor_q3vl(
        text=[prompt_full], images=[image], return_tensors="pt", padding=True,
        truncation=True, max_length=MAX_INPUT_TOKEN_LENGTH
    ).to(device)
    streamer = TextIteratorStreamer(processor_q3vl, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
    thread = Thread(target=model_q3vl.generate, kwargs=generation_kwargs)
    thread.start()
    buffer = ""
    for new_text in streamer:
        buffer += new_text
        time.sleep(0.01)
        yield buffer, buffer

@spaces.GPU
def generate_video(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 Qwen3-VL model for video input.
    """
    if video_path is None:
        yield "Please upload a video.", "Please upload a video."
        return

    frames_with_ts = downsample_video(video_path)
    if not frames_with_ts:
        yield "Could not process video.", "Could not process video."
        return

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

    prompt_full = processor_q3vl.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor_q3vl(
        text=[prompt_full], images=images_for_processor, return_tensors="pt", padding=True,
        truncation=True, max_length=MAX_INPUT_TOKEN_LENGTH
    ).to(device)
    streamer = TextIteratorStreamer(processor_q3vl, 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_q3vl.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


# Define examples for image and video inference
image_examples = [
    ["Describe the safety measures in the image. Conclude (Safe / Unsafe)..", "images/5.jpg"],
    ["Convert this page to doc [markdown] precisely.", "images/3.png"],
    ["Convert this page to doc [markdown] precisely.", "images/4.png"],
    ["Explain the creativity in the image.", "images/6.jpg"],
    ["Convert this page to doc [markdown] precisely.", "images/1.png"],
    ["Convert chart to OTSL.", "images/2.png"]
]

video_examples = [
    ["Explain the video in detail.", "videos/2.mp4"],
    ["Explain the ad in detail.", "videos/1.mp4"]
]

css = """
.submit-btn { background-color: #2980b9 !important; color: white !important; }
.submit-btn:hover { background-color: #3498db !important; }
.canvas-output { border: 2px solid #4682B4; border-radius: 10px; padding: 20px; }
"""

# Create the Gradio Interface
with gr.Blocks(css=css, theme="bethecloud/storj_theme") as demo:
    gr.Markdown("# **[Multimodal VLM Thinking](https://huggingface.co/collections/prithivMLmods/multimodal-implementations-67c9982ea04b39f0608badb0)**")
    with gr.Row():
        with gr.Column():
            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="Image", height=290)
                    image_submit = gr.Button("Submit", elem_classes="submit-btn")
                    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="Video", height=290)
                    video_submit = gr.Button("Submit", elem_classes="submit-btn")
                    gr.Examples(examples=video_examples, inputs=[video_query, video_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():
            with gr.Column(elem_classes="canvas-output"):
                gr.Markdown("## Output")
                output = gr.Textbox(label="Raw Output Stream", interactive=False, lines=5, show_copy_button=True)
                with gr.Accordion("(Result.md)", open=False):
                    markdown_output = gr.Markdown(label="(Result.Md)")
            gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/Multimodal-VLM-Thinking/discussions)")
            gr.Markdown("> [Qwen/Qwen3-VL-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct) is a powerful, versatile vision-language model. It excels at understanding and processing both text and visual information, making it suitable for a wide range of multimodal tasks. The model demonstrates strong performance in areas like visual question answering, image captioning, and video analysis.")
            gr.Markdown("> ⚠️ Note: Video inference performance can vary depending on the complexity and length of the video.")

    image_submit.click(
        fn=generate_image,
        inputs=[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=[video_query, video_upload, max_new_tokens, temperature, top_p, top_k, repetition_penalty],
        outputs=[output, markdown_output]
    )

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