app.py

import os
import time
import threading
import gradio as gr
import spaces
import torch
from PIL import Image
from transformers import (
    AutoModelForImageTextToText,
    AutoProcessor,
    TextIteratorStreamer,
)
from transformers.image_utils import load_image

# Constants for text generation
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 LFM2-VL-1.6B
MODEL_ID_M = "LiquidAI/LFM2-VL-1.6B"
processor_m = AutoProcessor.from_pretrained(MODEL_ID_M, trust_remote_code=True)
model_m = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID_M,
    trust_remote_code=True,
    torch_dtype="bfloat16",
).to(device).eval()

# Load LFM2-VL-450M
MODEL_ID_T = "LiquidAI/LFM2-VL-450M"
processor_t = AutoProcessor.from_pretrained(MODEL_ID_T, trust_remote_code=True)
model_t = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID_T,
    trust_remote_code=True,
    torch_dtype="bfloat16",
).to(device).eval()

@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):
    """
    Generate responses using the selected model for image input.
    """
    if model_name == "LFM2-VL-1.6B":
        processor = processor_m
        model = model_m
    elif model_name == "LFM2-VL-450M":
        processor = processor_t
        model = model_t
    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", "image": 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,
        truncation=False,
        max_length=MAX_INPUT_TOKEN_LENGTH
    ).to(device)
    streamer = TextIteratorStreamer(processor, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = {**inputs, "streamer": streamer, "max_new_tokens": max_new_tokens}
    thread = threading.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

# Define examples for image inference
image_examples = [
    ["According to this diagram, where do severe droughts occur?", "images/1.png"],
    ["Could you describe this image?", "images/2.jpg"],
    ["Provide a description of this image.", "images/3.jpg"],
    ["Explain the movie shot in detail.", "images/4.png"],
]

# Updated CSS with model choice highlighting
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=gr.themes.Soft()) as demo:
    gr.Markdown("# **LFM2-VL by [LiquidAI](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)**")
    with gr.Row():
        with gr.Column():
            image_query = gr.Textbox(label="Query Input", placeholder="✦︎ Enter your query")
            image_upload = gr.Image(type="pil", label="Image")
            image_submit = gr.Button("Submit", elem_classes="submit-btn")
            gr.Examples(
                examples=image_examples,
                inputs=[image_query, image_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=2)
                with gr.Accordion("(Result.md)", open=False):                
                    markdown_output = gr.Markdown(label="(Result.md)")

            model_choice = gr.Dropdown(
                choices=["LFM2-VL-1.6B", "LFM2-VL-450M"],
                label="Select Model",
                value="LFM2-VL-1.6B"
            )

            gr.Markdown("**Model Info 💻** | [Report Bug](https://huggingface.co/spaces/prithivMLmods/LFM2-VL-Demo/discussions)")
            gr.Markdown("> [LFM2‑VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa) is [Liquid AI’s](https://huggingface.co/LiquidAI) first multimodal model series, featuring models with 450M and 1.6B parameters designed for efficient processing of both text and images at native resolutions up to 512×512, ideal for low-latency edge AI applications; leveraging a hybrid conv+attention LFM2 backbone and SigLIP2 NaFlex vision encoders, it delivers flexible, user-tunable inference with rapid speeds (2× faster than existing VLMs on GPU)")
            gr.Markdown("> Competitive accuracy, and dynamic image tokenization for scalable throughput, while supporting 32,768 text tokens and English language generation, and is best fine-tuned for targeted use cases using provided supervised fine-tuning tools, all released under the LFM Open License v1.0 for research and deployment scenarios not requiring safety-critical guarantees.")

    # Define the submit button action
    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])

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