upgrade SmolVLM-500M -> Qwen2.5-VL-7B for real vision quality on ZeroGPU

app.py

@@ -1,9 +1,8 @@
 """
 Apiarist - Offline AI inspector for honeybee hive frames.
 
-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.
+Day 7: upgraded SmolVLM-500M -> Qwen2.5-VL-7B for real vision quality.
+ZeroGPU gives us a Blackwell GPU with plenty of VRAM, so 7B is cheap.
 """
 
 import gradio as gr
@@ -11,9 +10,9 @@ from PIL import Image
 import json
 import re
 import torch
-from transformers import AutoProcessor, AutoModelForVision2Seq
+from transformers import AutoProcessor, AutoModelForImageTextToText
 
-# ZeroGPU integration — gracefully degrades to a no-op outside HF Spaces
+# ZeroGPU integration — no-op outside HF Spaces
 try:
     import spaces
 
@@ -24,7 +23,7 @@ except ImportError:
         return fn
 
 
-MODEL_ID = "HuggingFaceTB/SmolVLM-500M-Instruct"
+MODEL_ID = "Qwen/Qwen2.5-VL-7B-Instruct"
 
 _model = None
 _processor = None
@@ -36,7 +35,10 @@ def get_model():
     if _model is None:
         print(f"Loading {MODEL_ID} ...")
         _processor = AutoProcessor.from_pretrained(MODEL_ID)
-        _model = AutoModelForVision2Seq.from_pretrained(MODEL_ID)
+        _model = AutoModelForImageTextToText.from_pretrained(
+            MODEL_ID,
+            torch_dtype=torch.bfloat16,
+        )
         _model.eval()
         print("Model loaded.")
     return _model, _processor
@@ -52,10 +54,11 @@ HEALTH: good, watch, or alarm
 NOTES: one short sentence describing what you see
 
 Definitions:
-- Queens are noticeably larger bees with long abdomens.
-- Varroa mites are small reddish-brown parasites on bees or comb cells.
-- Swarm cells are peanut-shaped cells hanging from the bottom edge of the comb.
-- Brood pattern is solid when cells are tightly packed and consistent, spotty when scattered."""
+- Queens are noticeably larger bees with elongated abdomens, often appearing distinct from worker bees.
+- Varroa mites are small reddish-brown parasites visible on bees or comb cells.
+- Swarm cells are peanut-shaped cells hanging from the bottom or edges of the comb.
+- Brood pattern is solid when capped cells are tightly packed and consistent, spotty when scattered with empty cells.
+- Be honest about uncertainty — only say "yes" when you can clearly see the feature."""
 
 
 def parse_response(text: str, hive_name: str) -> dict:
@@ -101,7 +104,7 @@ def build_narrative(r: dict, raw: str) -> str:
 **Notes:** {r['notes']}
 
 ---
-*Powered by SmolVLM-500M on ZeroGPU — lightweight stand-in. Qwen2.5-VL-7B + specialist YOLO detector are the planned final stack.*
+*Powered by Qwen2.5-VL-7B on ZeroGPU. Fully local, no cloud APIs.*
 
 <details><summary>Raw model output</summary>
 
@@ -119,26 +122,28 @@ def analyze_frame(image: Image.Image, hive_name: str):
 
     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
+    dtype = torch.bfloat16 if device == "cuda" else torch.float32
     model = model.to(device=device, dtype=dtype)
 
     messages = [
         {
             "role": "user",
             "content": [
-                {"type": "image"},
+                {"type": "image", "image": image},
                 {"type": "text", "text": INSPECTION_PROMPT},
             ],
         }
     ]
 
     try:
-        prompt = processor.apply_chat_template(
-            messages, add_generation_prompt=True
+        inputs = processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            add_generation_prompt=True,
+            return_dict=True,
+            return_tensors="pt",
         )
-        inputs = processor(text=prompt, images=[image], return_tensors="pt")
         inputs = inputs.to(device)
 
         with torch.no_grad():
@@ -229,11 +234,11 @@ with gr.Blocks(title="Apiarist - Hive Frame Inspector") as app:
                 """
 **Apiarist** is a fully-offline vision AI for backyard beekeepers.
 
-- 🔌 No cloud APIs — runs entirely on your laptop
-- 🎯 Fine-tuned for honeybees
+- 🔌 No cloud APIs — runs entirely on the laptop
+- 🎯 Vision-language model fine-tuned for honeybees
 - 📓 Built in 10 days for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon)
 
-**Stack**: SmolVLM-500M on ZeroGPU. Qwen2.5-VL-7B + YOLOv8 are the planned upgrades.
+**Stack**: Qwen2.5-VL-7B on ZeroGPU, served via Gradio.
 """
             )