app.py

import spaces
import subprocess
# Install flash attention, skipping CUDA build if necessary
subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)
import time
import logging
import gradio as gr
import cv2
import os
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch
from PIL import Image
import numpy as np
from pathlib import Path

# Cache for loaded model and processor
default_cache = {'model_id': None, 'processor': None, 'model': None, 'device': None}
model_cache = default_cache.copy()

# Check for XPU availability
has_xpu = hasattr(torch, 'xpu') and torch.xpu.is_available()

def update_model(model_id, device):
    if model_cache['model_id'] != model_id or model_cache['device'] != device:
        logging.info(f'Loading model {model_id} on {device}')
        try:
            processor = AutoProcessor.from_pretrained(model_id)
            # Load model with appropriate precision for each device
            if device == 'cuda':
                # Use bfloat16 for CUDA for performance
                model = AutoModelForImageTextToText.from_pretrained(
                    model_id,
                    torch_dtype=torch.bfloat16,
                    _attn_implementation='flash_attention_2'
                ).to('cuda')
            elif device == 'xpu' and has_xpu:
                # Use float32 on XPU to avoid bfloat16 layernorm issues
                model = AutoModelForImageTextToText.from_pretrained(
                    model_id,
                    torch_dtype=torch.float32
                ).to('xpu')
            else:
                # Default to float32 on CPU
                model = AutoModelForImageTextToText.from_pretrained(model_id).to('cpu')
            model.eval()
            model_cache.update({'model_id': model_id, 'processor': processor, 'model': model, 'device': device})
        except Exception as e:
            logging.error(f'Error loading model: {e}')
            raise e

def extract_frames_from_video(video_path, max_frames=10):
    """Extract frames from video file for processing"""
    if not os.path.exists(video_path):
        raise FileNotFoundError(f"Video file not found: {video_path}")
    
    # Validate video file
    if not video_path.lower().endswith(('.mp4', '.avi', '.mov', '.mkv', '.webm')):
        raise ValueError("Unsupported video format. Please use MP4, AVI, MOV, MKV, or WEBM.")
    
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Cannot open video file: {video_path}")
    
    frames = []
    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    if frame_count == 0:
        cap.release()
        raise ValueError("Video file appears to be empty or corrupted")
    
    # Calculate step size to extract evenly distributed frames
    step = max(1, frame_count // max_frames)
    
    frame_idx = 0
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        
        if frame_idx % step == 0:
            frames.append(frame)
            if len(frames) >= max_frames:
                break
        
        frame_idx += 1
    
    cap.release()
    return frames

@spaces.GPU
def caption_frame(frame, model_id, interval_ms, sys_prompt, usr_prompt, device):
    """Caption a single frame (used for webcam streaming)"""
    debug_msgs = []
    try:
        update_model(model_id, device)
        processor = model_cache['processor']
        model = model_cache['model']

        # Control capture interval
        time.sleep(interval_ms / 1000)

        # Preprocess frame
        t0 = time.time()
        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        pil_img = Image.fromarray(rgb)
        temp_path = 'frame.jpg'
        pil_img.save(temp_path, format='JPEG', quality=50)
        debug_msgs.append(f'Preprocess: {int((time.time()-t0)*1000)} ms')

        # Prepare multimodal chat messages
        messages = [
            {'role': 'system', 'content': [{'type': 'text', 'text': sys_prompt}]},
            {'role': 'user', 'content': [
                {'type': 'image', 'url': temp_path},
                {'type': 'text', 'text': usr_prompt}
            ]}
        ]

        # Tokenize and encode
        t1 = time.time()
        inputs = processor.apply_chat_template(
            messages,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors='pt'
        )
        # Move inputs to correct device and dtype (matching model parameters)
        param_dtype = next(model.parameters()).dtype
        cast_inputs = {}
        for k, v in inputs.items():
            if isinstance(v, torch.Tensor):
                if v.dtype.is_floating_point:
                    # cast floating-point tensors to model's parameter dtype
                    cast_inputs[k] = v.to(device=model.device, dtype=param_dtype)
                else:
                    # move integer/mask tensors without changing dtype
                    cast_inputs[k] = v.to(device=model.device)
            else:
                cast_inputs[k] = v
        inputs = cast_inputs
        debug_msgs.append(f'Tokenize: {int((time.time()-t1)*1000)} ms')

        # Inference
        t2 = time.time()
        outputs = model.generate(**inputs, do_sample=False, max_new_tokens=128)
        debug_msgs.append(f'Inference: {int((time.time()-t2)*1000)} ms')

        # Decode and strip history
        t3 = time.time()
        raw = processor.batch_decode(outputs, skip_special_tokens=True)[0]
        debug_msgs.append(f'Decode: {int((time.time()-t3)*1000)} ms')
        if "Assistant:" in raw:
            caption = raw.split("Assistant:")[-1].strip()
        else:
            lines = raw.splitlines()
            caption = lines[-1].strip() if len(lines) > 1 else raw.strip()

        # Clean up temp file
        if os.path.exists(temp_path):
            os.remove(temp_path)

        return caption, '\n'.join(debug_msgs)
    except Exception as e:
        return f"Error: {str(e)}", '\n'.join(debug_msgs)

@spaces.GPU
def process_video_file(video_file, model_id, sys_prompt, usr_prompt, device, max_frames):
    """Process uploaded video file and return captions for multiple frames"""
    if video_file is None:
        return "No video file uploaded", ""
    
    debug_msgs = []
    temp_files = []  # Track temporary files for cleanup
    
    try:
        update_model(model_id, device)
        processor = model_cache['processor']
        model = model_cache['model']
        
        # Extract frames from video
        t0 = time.time()
        frames = extract_frames_from_video(video_file, max_frames)
        debug_msgs.append(f'Extracted {len(frames)} frames in {int((time.time()-t0)*1000)} ms')
        
        if not frames:
            return "No frames could be extracted from the video", '\n'.join(debug_msgs)
        
        captions = []
        
        for i, frame in enumerate(frames):
            # Preprocess frame
            t1 = time.time()
            rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            pil_img = Image.fromarray(rgb)
            temp_path = f'frame_{i}.jpg'
            temp_files.append(temp_path)  # Track for cleanup
            pil_img.save(temp_path, format='JPEG', quality=50)
            
            # Prepare multimodal chat messages
            messages = [
                {'role': 'system', 'content': [{'type': 'text', 'text': sys_prompt}]},
                {'role': 'user', 'content': [
                    {'type': 'image', 'url': temp_path},
                    {'type': 'text', 'text': usr_prompt}
                ]}
            ]
            
            # Tokenize and encode
            inputs = processor.apply_chat_template(
                messages,
                add_generation_prompt=True,
                tokenize=True,
                return_dict=True,
                return_tensors='pt'
            )
            
            # Move inputs to correct device and dtype
            param_dtype = next(model.parameters()).dtype
            cast_inputs = {}
            for k, v in inputs.items():
                if isinstance(v, torch.Tensor):
                    if v.dtype.is_floating_point:
                        cast_inputs[k] = v.to(device=model.device, dtype=param_dtype)
                    else:
                        cast_inputs[k] = v.to(device=model.device)
                else:
                    cast_inputs[k] = v
            inputs = cast_inputs
            
            # Inference
            outputs = model.generate(**inputs, do_sample=False, max_new_tokens=128)
            
            # Decode and strip history
            raw = processor.batch_decode(outputs, skip_special_tokens=True)[0]
            if "Assistant:" in raw:
                caption = raw.split("Assistant:")[-1].strip()
            else:
                lines = raw.splitlines()
                caption = lines[-1].strip() if len(lines) > 1 else raw.strip()
            
            captions.append(f"Frame {i+1}: {caption}")
            debug_msgs.append(f'Frame {i+1} processed in {int((time.time()-t1)*1000)} ms')
        
        return '\n\n'.join(captions), '\n'.join(debug_msgs)
        
    except Exception as e:
        return f"Error processing video: {str(e)}", '\n'.join(debug_msgs)
    finally:
        # Clean up all temporary files
        for temp_file in temp_files:
            if os.path.exists(temp_file):
                try:
                    os.remove(temp_file)
                except Exception as cleanup_error:
                    logging.warning(f"Failed to cleanup {temp_file}: {cleanup_error}")

def toggle_input_mode(input_mode):
    """Toggle between webcam and video file input"""
    if input_mode == "Webcam":
        return gr.update(visible=True), gr.update(visible=False), gr.update(visible=False)
    else:  # Video File
        return gr.update(visible=False), gr.update(visible=True), gr.update(visible=True)

def main():
    logging.basicConfig(level=logging.INFO)
    model_choices = [
        'HuggingFaceTB/SmolVLM2-256M-Video-Instruct',
        'HuggingFaceTB/SmolVLM2-500M-Video-Instruct',
        'HuggingFaceTB/SmolVLM2-2.2B-Instruct'
    ]
    
    # Determine available devices
    device_options = ['cpu']
    if torch.cuda.is_available():
        device_options.append('cuda')
    if has_xpu:
        device_options.append('xpu')

    default_device = 'cuda' if torch.cuda.is_available() else ('xpu' if has_xpu else 'cpu')

    with gr.Blocks() as demo:
        gr.Markdown('## 🎥 Real-Time Webcam & Video File Captioning with SmolVLM2 (Transformers)')
        
        with gr.Row():
            input_mode = gr.Radio(
                choices=["Webcam", "Video File"], 
                value="Webcam", 
                label="Input Mode"
            )

        with gr.Row():
            model_dd = gr.Dropdown(model_choices, value=model_choices[0], label='Model ID')
            device_dd = gr.Dropdown(device_options, value=default_device, label='Device')

        # Webcam-specific controls
        with gr.Row() as webcam_controls:
            interval = gr.Slider(100, 20000, step=100, value=3000, label='Interval (ms)')
        
        # Video file-specific controls
        with gr.Row(visible=False) as video_controls:
            max_frames = gr.Slider(1, 20, step=1, value=5, label='Max Frames to Process')

        sys_p = gr.Textbox(lines=2, value='Describe the key action', label='System Prompt')
        usr_p = gr.Textbox(lines=1, value='What is happening in this image?', label='User Prompt')

        # Input components
        cam = gr.Image(sources=['webcam'], streaming=True, label='Webcam Feed')
        video_file = gr.File(
            label="Upload Video File",
            file_types=[".mp4", ".avi", ".mov", ".mkv", ".webm"],
            visible=False
        )
        
        # Process button for video files
        process_btn = gr.Button("Process Video", visible=False)

        # Output components
        caption_tb = gr.Textbox(interactive=False, label='Caption')
        log_tb = gr.Textbox(lines=4, interactive=False, label='Debug Log')

        # Toggle input mode
        input_mode.change(
            fn=toggle_input_mode,
            inputs=[input_mode],
            outputs=[cam, video_file, process_btn]
        )
        
        # Also toggle the control panels
        input_mode.change(
            fn=lambda mode: (gr.update(visible=mode=="Webcam"), gr.update(visible=mode=="Video File")),
            inputs=[input_mode],
            outputs=[webcam_controls, video_controls]
        )

        # Webcam streaming
        cam.stream(
            fn=caption_frame,
            inputs=[cam, model_dd, interval, sys_p, usr_p, device_dd],
            outputs=[caption_tb, log_tb],
            time_limit=600
        )
        
        # Video file processing
        process_btn.click(
            fn=process_video_file,
            inputs=[video_file, model_dd, sys_p, usr_p, device_dd, max_frames],
            outputs=[caption_tb, log_tb]
        )

    # Enable Gradio's async event queue
    demo.queue()

    # Launch the app
    demo.launch()


if __name__ == '__main__':
    main()