app.py

# app.py
import gradio as gr
import spaces
import torch
from PIL import Image


from vlm_inference import (
    load_vlm_model,
    vlm_infer_stream,
    image_processor,
)

# =====================================================
# Load VLM on CPU (ZeroGPU)
# =====================================================
print("[DEBUG] Loading VLM model on CPU...")
model = load_vlm_model()
model.eval()
print("[DEBUG] VLM model loaded.")


# =====================================================
# message parser (multimodal=True 仕様準拠)
# =====================================================
def parse_message(message: dict):
    """
    message = {
        "text": str,
        "files": list  # PIL.Image が入る
    }
    """
    print("[DEBUG] parse_message called")
    print("[DEBUG] message type:", type(message))
    print("[DEBUG] message content:", message)

    text = message.get("text", "")
    files = message.get("files", [])

    print("[DEBUG] parsed text:", repr(text))
    print("[DEBUG] parsed files:", files)

    image = files[0] if files else None
    print("[DEBUG] parsed image:", image)

    return text, image


# =====================================================
# GPU inference (single-turn, VLM only)
# =====================================================
@spaces.GPU
def chat_fn(message, history, temperature, top_p, top_k):
    text, image = parse_message(message)

    if image is None:
        yield "Image input is required."
        return

    device = "cuda"
    model_gpu = model.to(device)

    if isinstance(image, str):
        from PIL import Image
        image = Image.open(image)

    image_tensor = image_processor(
        images=image.convert("RGB"),
        return_tensors="pt"
    )["pixel_values"].to(device)

    prompt = (
        f"{text}"
    )

    # ★ ここが重要：累積して yield
    output = ""

    for chunk in vlm_infer_stream(
        model=model_gpu,
        image_tensor=image_tensor,
        prompt=prompt,
        max_new_tokens=256,
        temperature=temperature,
        top_p=top_p if top_p > 0 else None,
        top_k=top_k if top_k > 0 else None,
    ):
        output += chunk
        yield output

    model_gpu.to("cpu")
    torch.cuda.empty_cache()




# =====================================================
# UI (ChatInterface, multimodal)
# =====================================================
print("[DEBUG] Building Gradio UI")

demo = gr.ChatInterface(
    fn=chat_fn,
    multimodal=True,
    title="EveryonesGPT Vision Pretrained(No-Instruction). Single-turn English Only demo (CLIP ViT-L/14)",
    description="No text input. Download example image. https://raw.githubusercontent.com/HayatoHongo/nanoGPTVision/main/hodomoe_cat.png",
    additional_inputs=[
        gr.Slider(0.1, 2.0, value=0.2, step=0.05, label="Temperature"),
        gr.Slider(0.0, 1.0, value=0.9, step=0.05, label="Top-p"),
        gr.Slider(0, 200, value=0, step=1, label="Top-k"),
    ],
)

print("[DEBUG] Launching Gradio app")
demo.launch()