app.py

import gradio as gr
from transformers import AutoProcessor, AutoModelForImageTextToText, TextIteratorStreamer
from threading import Thread
import re
import time
import torch
import spaces
import subprocess
from io import BytesIO

subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM2-500M-Video-Instruct")
model = AutoModelForImageTextToText.from_pretrained(
    "HuggingFaceTB/SmolVLM2-500M-Video-Instruct",
    _attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16
).to("cuda:0")


@spaces.GPU
def model_inference(
    input_dict, history, max_tokens
):

    text = input_dict["text"].strip()
    user_content = []
    media_queue = []

    for file_path in input_dict.get("files", []):
        if file_path.lower().endswith((".png", ".jpg", ".jpeg", ".gif", ".bmp")):
            media_queue.append({"type": "image", "path": file_path})
        elif file_path.lower().endswith((".mp4", ".mov", ".avi", ".mkv", ".flv")):
            media_queue.append({"type": "video", "path": file_path})


    if not text and not media_queue:
        gr.Warning("Please input a query and optionally image(s)/video(s).")
        return

    if not text and media_queue:
        gr.Warning("Please input a text query along with the image(s)/video(s).")
        return


    if "<image>" in text or "<video>" in text:
        parts = re.split(r'(<image>|<video>)', text)
        temp_media_queue = list(media_queue)
        for part in parts:
            if part == "<image>" and temp_media_queue:
                media_item = temp_media_queue.pop(0)
                if media_item["type"] == "image":
                    user_content.append(media_item)
                else:
                    gr.Warning(f"Placeholder <image> found, but next media is a video: {media_item['path']}. Skipping placeholder.")
                    user_content.append({"type": "text", "text": part})
                    temp_media_queue.insert(0, media_item)
            elif part == "<video>" and temp_media_queue:
                media_item = temp_media_queue.pop(0)
                if media_item["type"] == "video":
                     user_content.append(media_item)
                else:
                    gr.Warning(f"Placeholder <video> found, but next media is an image: {media_item['path']}. Skipping placeholder.")
                    user_content.append({"type": "text", "text": part})
                    temp_media_queue.insert(0, media_item)
            elif part.strip():
                user_content.append({"type": "text", "text": part.strip()})
            elif part in ["<image>", "<video>"] and not temp_media_queue:
                 gr.Warning(f"Placeholder {part} found, but no more media items available.")
                 user_content.append({"type": "text", "text": part})

        user_content.extend(temp_media_queue)

    else:
        if text:
             user_content.append({"type": "text", "text": text})
        user_content.extend(media_queue)


    resulting_messages = [{"role": "user", "content": user_content}]


    print("resulting_messages", resulting_messages)

    try:
        inputs = processor.apply_chat_template(
            resulting_messages,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors="pt",
        )
    except Exception as e:
        gr.Error(f"Error during input processing: {e}")
        print(f"Processor Error: {e}")
        print("Problematic message structure:", resulting_messages)
        return


    inputs = inputs.to(model.device)


    if "pixel_values" in inputs:
        inputs["pixel_values"] = inputs["pixel_values"].to(model.dtype)


    streamer = TextIteratorStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True)
    generation_args = dict(inputs, streamer=streamer, max_new_tokens=max_tokens)
    generated_text = ""

    thread = Thread(target=model.generate, kwargs=generation_args)
    thread.start()

    buffer = ""

    for new_text in streamer:
        buffer += new_text
        yield buffer

    thread.join()


examples=[
              [{"text": "Where do the severe droughts happen according to this diagram?", "files": ["example_images/examples_weather_events.png"]}],
              [{"text": "What art era this artpiece <image> and this artpiece <image> belong to?", "files": ["example_images/rococo.jpg", "example_images/rococo_1.jpg"]}],
              [{"text": "Describe this image.", "files": ["example_images/mosque.jpg"]}],
              [{"text": "When was this purchase made and how much did it cost?", "files": ["example_images/fiche.jpg"]}],
              [{"text": "What is the date in this document?", "files": ["example_images/document.jpg"]}],
              [{"text": "What is happening in the video?", "files": ["example_images/short.mp4"]}],
      ]
demo = gr.ChatInterface(fn=model_inference, title="SmolVLM2: The Smollest Video Model Ever 📺",
                description="Play with [SmolVLM2-500M-Video-Instruct](https://huggingface.co/HuggingFaceTB/SmolVLM2-500M-Video-Instruct) in this demo. To get started, upload an image/video and text or try one of the examples. This demo doesn't use history for the chat, so every chat you start is a new conversation.",
                examples=examples,
                textbox=gr.MultimodalTextbox(label="Query Input", file_types=["image", ".mp4"], file_count="multiple"), stop_btn="Stop Generation", multimodal=True,
                cache_examples=False,
                additional_inputs=[gr.Slider(minimum=100, maximum=500, step=50, value=200, label="Max Tokens")],
                type="messages"
                )



demo.launch(debug=True)