wire @spaces.GPU decorator for ZeroGPU; auto cuda/cpu device handling

app.py

@@ -1,9 +1,9 @@
 """
 Apiarist - Offline AI inspector for honeybee hive frames.
 
-Day 5: switched from Moondream2 (version-pinning hell) to SmolVLM-500M
-(HuggingFace's own model, standard transformers API). Will swap up to
-Qwen2.5-VL-7B once GPU access lands.
+Day 6: SmolVLM-500M + ZeroGPU (free on-demand GPU). The @spaces.GPU
+decorator tells HF to allocate a GPU just for the inference call,
+then release it.
 """
 
 import gradio as gr
@@ -13,6 +13,16 @@ import re
 import torch
 from transformers import AutoProcessor, AutoModelForVision2Seq
 
+# ZeroGPU integration — gracefully degrades to a no-op outside HF Spaces
+try:
+    import spaces
+
+    def gpu(fn):
+        return spaces.GPU(duration=60)(fn)
+except ImportError:
+    def gpu(fn):
+        return fn
+
 
 MODEL_ID = "HuggingFaceTB/SmolVLM-500M-Instruct"
 
@@ -21,14 +31,12 @@ _processor = None
 
 
 def get_model():
+    """Lazy-load model on first call. Stays on CPU until moved by analyze_frame."""
     global _model, _processor
     if _model is None:
         print(f"Loading {MODEL_ID} ...")
         _processor = AutoProcessor.from_pretrained(MODEL_ID)
-        _model = AutoModelForVision2Seq.from_pretrained(
-            MODEL_ID,
-            torch_dtype=torch.float32,  # CPU prefers FP32
-        )
+        _model = AutoModelForVision2Seq.from_pretrained(MODEL_ID)
         _model.eval()
         print("Model loaded.")
     return _model, _processor
@@ -93,7 +101,7 @@ def build_narrative(r: dict, raw: str) -> str:
 **Notes:** {r['notes']}
 
 ---
-*Powered by SmolVLM-500M — lightweight stand-in. Qwen2.5-VL-7B + specialist YOLO detector land once GPU access is granted.*
+*Powered by SmolVLM-500M on ZeroGPU — lightweight stand-in. Qwen2.5-VL-7B + specialist YOLO detector are the planned final stack.*
 
 <details><summary>Raw model output</summary>
 
@@ -104,12 +112,18 @@ def build_narrative(r: dict, raw: str) -> str:
 """
 
 
+@gpu
 def analyze_frame(image: Image.Image, hive_name: str):
     if image is None:
         return None, "Upload a frame photo first.", ""
 
     model, processor = get_model()
 
+    # ZeroGPU allocates a GPU only inside this @gpu-decorated call.
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    dtype = torch.float16 if device == "cuda" else torch.float32
+    model = model.to(device=device, dtype=dtype)
+
     messages = [
         {
             "role": "user",
@@ -125,6 +139,7 @@ def analyze_frame(image: Image.Image, hive_name: str):
             messages, add_generation_prompt=True
         )
         inputs = processor(text=prompt, images=[image], return_tensors="pt")
+        inputs = inputs.to(device)
 
         with torch.no_grad():
             generated = model.generate(
@@ -218,7 +233,7 @@ with gr.Blocks(title="Apiarist - Hive Frame Inspector") as app:
 - 🎯 Fine-tuned for honeybees
 - 📓 Built in 10 days for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon)
 
-**Stack**: SmolVLM (placeholder) → Qwen2.5-VL-7B + YOLOv8 (final), Gradio, llama.cpp.
+**Stack**: SmolVLM-500M on ZeroGPU. Qwen2.5-VL-7B + YOLOv8 are the planned upgrades.
 """
             )