app.py

# app.py
# --- ENV FIX: MUSS GANZ OBEN STEHEN, vor JEDEM anderen Import! ---
import os
os.environ.pop("TORCH_LOGS", None)                 # falls ungültig gesetzt -> torch-Import fixen
os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")  # Dynamo/JIT aus (vermeidet Gemma3-Issues)
os.environ.setdefault("CUDA_VISIBLE_DEVICES", "0")

import gradio as gr
import spaces
import torch
from PIL import Image
from transformers import pipeline

# Optional: Dynamo-Fehler unterdrücken, falls doch aktiv
try:
    import torch._dynamo as dynamo
    dynamo.config.suppress_errors = True
except Exception:
    pass

# -----------------------------
# Config
# -----------------------------
MODEL_ID = os.getenv("MODEL_ID", "olemeyer/dentalai-nxt-medgemma4b")  # ggf. "google/medgemma-4b-it"
HF_TOKEN = os.getenv("HF_TOKEN")  # im Space als Secret setzen

def _device_index() -> int:
    return 0 if torch.cuda.is_available() else -1

torch.set_grad_enabled(False)
if hasattr(torch, "set_float32_matmul_precision"):
    torch.set_float32_matmul_precision("high")

pipe = pipeline(
    task="image-text-to-text",
    model=MODEL_ID,
    token=HF_TOKEN,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else None,
    device=_device_index(),  # 0=CUDA:0, -1=CPU
)

# -----------------------------
# Utils
# -----------------------------
def _resize_long_side_max(img: Image.Image, max_long_side: int = 800) -> Image.Image:
    w, h = img.size
    long_side = max(w, h)
    if long_side <= max_long_side:
        return img
    scale = max_long_side / float(long_side)
    new_w = int(round(w * scale))
    new_h = int(round(h * scale))
    return img.resize((new_w, new_h), Image.LANCZOS)

# -----------------------------
# Inferenz (Trainingsformat: NUR user, Bild zuerst)
# -----------------------------
@spaces.GPU
def infer(image: Image.Image):
    if image is None:
        return "Bitte ein Bild hochladen."

    # Bild auf lange Seite 800px skalieren
    image = _resize_long_side_max(image, 800)

    messages = [
        {
            "role": "user",
            "content": [
                {"index": 0, "type": "image", "text": None, "image": image},
                {
                    "index": None,
                    "type": "text",
                    "text": "Analysiere dieses Zahnbild und extrahiere alle verfügbaren Informationen als JSON.",
                },
            ],
        }
    ]

    try:
        out = pipe(
            text=messages,
            max_new_tokens=2048,   # angefordert
            temperature=0.1,       # angefordert
        )
    except TypeError:
        # Fallback für alternative Parameternamen
        out = pipe(
            messages=messages,
            max_new_tokens=2048,
            temperature=0.1,
        )
    except Exception as e:
        return f"Fehler bei der Inferenz: {e}"

    # Ausgabe normalisieren
    try:
        first = out[0]
        gen = first.get("generated_text", first)
        if isinstance(gen, list) and gen:
            last_msg = gen[-1]
            content = last_msg.get("content", "")
            if isinstance(content, list):
                texts = [c.get("text", "") for c in content if isinstance(c, dict) and c.get("type") == "text"]
                joined = "\n".join(t for t in texts if t)
                return joined or str(last_msg)
            if isinstance(content, str):
                return content
            return str(last_msg)
        if isinstance(gen, str):
            return gen
        return str(gen)
    except Exception:
        return str(out)

# -----------------------------
# UI (simpel)
# -----------------------------
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🦷 DentalAI – Trainingsformat\nNur Bild hochladen → **Analysieren**.")
    with gr.Row():
        with gr.Column(scale=1):
            image_in = gr.Image(type="pil", label="Bild", height=360)
            run_btn = gr.Button("Analysieren", variant="primary")
        with gr.Column(scale=1):
            output = gr.Textbox(label="Ausgabe", lines=22, show_copy_button=True)

    run_btn.click(fn=infer, inputs=[image_in], outputs=[output], api_name="predict")

demo.queue().launch()