app.py

import os

os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")

import spaces  # MUST come before any torch / CUDA-touching import
import torch
import torchvision.transforms as T
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoTokenizer

# Register the custom SpatioLM model classes with AutoModel
import spatiolm.models  # noqa: F401  – registers InternVL3RChatModel, DA3Model, etc.
from spatiolm.models import InternVL3RChatModel

MODEL_ID = "xiaomi-research/SpatioLM-Understanding-InternVL3.5"

# ---------------------------------------------------------------------------
# Image preprocessing – replicated from lmms_eval.models.simple.internvl2
# (we avoid pulling in the full lmms-eval dependency for a single function)
# ---------------------------------------------------------------------------
IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)
INPUT_SIZE = 448


def _build_transform(input_size=INPUT_SIZE):
    transform = T.Compose([
        T.Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
    ])
    return transform


def _find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float("inf")
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def _dynamic_preprocess(image, min_num=1, max_num=6, image_size=INPUT_SIZE, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    target_ratios = set(
        (i, j)
        for n in range(min_num, max_num + 1)
        for i in range(1, n + 1)
        for j in range(1, n + 1)
        if i * j <= max_num and i * j >= min_num
    )
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    target_aspect_ratio = _find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size
    )

    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size,
        )
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image, input_size=INPUT_SIZE, max_num=6):
    """Load and preprocess an image into a tensor of tiles.

    Args:
        image: a PIL.Image or a path to an image file.
        input_size: pixel size of each tile.
        max_num: maximum number of tiles.

    Returns:
        torch.Tensor of shape (num_tiles, 3, input_size, input_size).
    """
    if isinstance(image, str):
        image = Image.open(image).convert("RGB")
    else:
        image = image.convert("RGB")
    transform = _build_transform(input_size=input_size)
    images = _dynamic_preprocess(
        image, image_size=input_size, use_thumbnail=True, max_num=max_num
    )
    pixel_values = [transform(img) for img in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values


# ---------------------------------------------------------------------------
# Model loading (module scope, eager .to("cuda") per ZeroGPU rules)
# ---------------------------------------------------------------------------
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    use_fast=False,
)
model = InternVL3RChatModel.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).eval().to("cuda")


# ---------------------------------------------------------------------------
# Inference
# ---------------------------------------------------------------------------
@spaces.GPU(duration=30)
def answer(
    image: Image.Image,
    question: str,
    max_new_tokens: int = 512,
    temperature: float = 0.0,
    top_p: float = 1.0,
) -> str:
    """Answer a spatial-reasoning question about an image.

    Args:
        image: the input image to analyse.
        question: a natural-language question about the scene (e.g.
            "Which object is closer to the camera?").
        max_new_tokens: maximum number of tokens to generate.
        temperature: sampling temperature; 0 for greedy.
        top_p: nucleus-sampling probability mass.

    Returns:
        The model's textual answer.
    """
    if image is None:
        return "Please upload an image."
    if not question.strip():
        return "Please enter a question."

    pixel_values = load_image(image).to(dtype=torch.bfloat16, device="cuda")
    num_patches_list = [pixel_values.shape[0]]

    gen_config = {
        "max_new_tokens": int(max_new_tokens),
        "do_sample": temperature > 0,
        "temperature": float(temperature) if temperature > 0 else 1.0,
        "top_p": float(top_p),
    }

    response = model.chat(
        tokenizer,
        pixel_values,
        question,
        gen_config,
        num_patches_list=num_patches_list,
    )
    return response


# ---------------------------------------------------------------------------
# Gradio UI
# ---------------------------------------------------------------------------
import gradio as gr

CSS = """
#col-container { max-width: 1100px; margin: 0 auto; }
.dark .gradio-container { color: var(--body-text-color); }
"""

EXAMPLE_IMAGES = [
    "cafe_interior.jpg",
    "living_room_blue_couch.jpg",
    "library_interior.jpg",
    "city_skyline_toronto.jpg",
    "spiral_staircase.jpg",
]

EXAMPLES = [
    ["cafe_interior.jpg", "Which table is closer to the camera, the one on the left or the one on the right?"],
    ["living_room_blue_couch.jpg", "Describe the spatial arrangement of the furniture in this room."],
    ["library_interior.jpg", "How many shelves can you see, and which one is closest to the viewer?"],
    ["spiral_staircase.jpg", "Does this staircase spiral clockwise or counter-clockwise when viewed from above?"],
]

with gr.Blocks() as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(
            "# SpatioLM: Spatial Understanding with Vision-Language Models\n"
            "Ask spatial-reasoning questions about any image. SpatioLM enhances "
            "physical spatial intelligence in VLMs without requiring 3D input at "
            "inference time.\n\n"
            "[Paper](https://arxiv.org/abs/2608.01899) | "
            "[Code](https://github.com/xiaomi-research/spatio-lm) | "
            "[Model](https://huggingface.co/xiaomi-research/SpatioLM-Understanding-InternVL3.5)"
        )

        with gr.Row():
            with gr.Column(scale=1):
                image_input = gr.Image(type="pil", label="Input Image")
                question_input = gr.Textbox(
                    label="Question",
                    placeholder="e.g. Which object is closer to the camera?",
                    lines=2,
                )
                run_btn = gr.Button("Ask", variant="primary")
            with gr.Column(scale=1):
                output_text = gr.Textbox(
                    label="Answer",
                    lines=10,
                    interactive=False,
                )

        with gr.Accordion("Advanced settings", open=False):
            max_tokens = gr.Slider(
                minimum=32, maximum=2048, value=512, step=32,
                label="Max new tokens",
            )
            temp_slider = gr.Slider(
                minimum=0.0, maximum=2.0, value=0.0, step=0.1,
                label="Temperature (0 = greedy)",
            )
            top_p_slider = gr.Slider(
                minimum=0.1, maximum=1.0, value=1.0, step=0.05,
                label="Top-p",
            )

        gr.Examples(
            examples=EXAMPLES,
            inputs=[image_input, question_input],
            outputs=output_text,
            fn=answer,
            cache_examples=True,
            cache_mode="lazy",
        )

    run_btn.click(
        fn=answer,
        inputs=[image_input, question_input, max_tokens, temp_slider, top_p_slider],
        outputs=output_text,
        api_name="answer",
    )

if __name__ == "__main__":
    demo.launch(mcp_server=True, theme=gr.themes.Citrus(), css=CSS)