revert: Qwen2.5-VL-7B crashes ZeroGPU container, use SmolVLM-Instruct 2.25B (known good)

app.py

@@ -10,7 +10,7 @@ from PIL import Image
 import json
 import re
 import torch
-from transformers import AutoProcessor, AutoModelForImageTextToText
+from transformers import AutoProcessor, AutoModelForVision2Seq
 
 # ZeroGPU integration — no-op outside HF Spaces
 try:
@@ -23,7 +23,11 @@ except ImportError:
         return fn
 
 
-MODEL_ID = "Qwen/Qwen2.5-VL-7B-Instruct"
+# SmolVLM-Instruct (2.25B). Bigger than the 500M we tested earlier —
+# much better at fine-grained vision tasks like spotting queens.
+# Qwen2.5-VL-7B kept crashing the ZeroGPU container at the OCI runtime
+# layer; reverting to the SmolVLM family which is known to work here.
+MODEL_ID = "HuggingFaceTB/SmolVLM-Instruct"
 
 _model = None
 _processor = None
@@ -35,10 +39,7 @@ def get_model():
     if _model is None:
         print(f"Loading {MODEL_ID} ...")
         _processor = AutoProcessor.from_pretrained(MODEL_ID)
-        _model = AutoModelForImageTextToText.from_pretrained(
-            MODEL_ID,
-            torch_dtype=torch.bfloat16,
-        )
+        _model = AutoModelForVision2Seq.from_pretrained(MODEL_ID)
         _model.eval()
         print("Model loaded.")
     return _model, _processor
@@ -104,7 +105,7 @@ def build_narrative(r: dict, raw: str) -> str:
 **Notes:** {r['notes']}
 
 ---
-*Powered by Qwen2.5-VL-7B on ZeroGPU. Fully local, no cloud APIs.*
+*Powered by SmolVLM-Instruct (2.25B) on ZeroGPU. Fully local, no cloud APIs.*
 
 <details><summary>Raw model output</summary>
 
@@ -123,27 +124,24 @@ def analyze_frame(image: Image.Image, hive_name: str):
     model, processor = get_model()
 
     device = "cuda" if torch.cuda.is_available() else "cpu"
-    dtype = torch.bfloat16 if device == "cuda" else torch.float32
+    dtype = torch.float16 if device == "cuda" else torch.float32
     model = model.to(device=device, dtype=dtype)
 
     messages = [
         {
             "role": "user",
             "content": [
-                {"type": "image", "image": image},
+                {"type": "image"},
                 {"type": "text", "text": INSPECTION_PROMPT},
             ],
         }
     ]
 
     try:
-        inputs = processor.apply_chat_template(
-            messages,
-            tokenize=True,
-            add_generation_prompt=True,
-            return_dict=True,
-            return_tensors="pt",
+        prompt = processor.apply_chat_template(
+            messages, add_generation_prompt=True
         )
+        inputs = processor(text=prompt, images=[image], return_tensors="pt")
         inputs = inputs.to(device)
 
         with torch.no_grad():
@@ -238,7 +236,7 @@ with gr.Blocks(title="Apiarist - Hive Frame Inspector") as app:
 - 🎯 Vision-language model fine-tuned for honeybees
 - 📓 Built in 10 days for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon)
 
-**Stack**: Qwen2.5-VL-7B on ZeroGPU, served via Gradio.
+**Stack**: SmolVLM-Instruct (2.25B) on ZeroGPU, served via Gradio.
 """
             )