Enforce image token in tiny mode

app.py

@@ -309,6 +309,9 @@ def score_binary_logits(logits: torch.Tensor, tokenizer) -> tuple[float, str]:
 
 
 def build_prompt(processor_obj, user_text: str) -> str:
+    if MODEL_MODE == "tiny":
+        # Tiny LLaVA forward expects an explicit image placeholder token.
+        return f"<image>\n{user_text}"
     return user_text
 
 
@@ -324,6 +327,16 @@ def infer_image(image: Image.Image):
         inputs = processor_obj(text=prompt_text, images=prepared_image, return_tensors="pt", padding=True)
         inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
 
+        if MODEL_MODE == "tiny":
+            image_token_id = getattr(model_obj.config, "image_token_index", None)
+            if image_token_id is not None and "input_ids" in inputs:
+                image_token_count = int((inputs["input_ids"] == image_token_id).sum().item())
+                if image_token_count == 0:
+                    # Defensive recovery if tokenizer template/path strips the image placeholder.
+                    fallback_prompt = f"<image>\n{TINY_PROMPT}"
+                    inputs = processor_obj(text=fallback_prompt, images=prepared_image, return_tensors="pt", padding=True)
+                    inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
+
         # Keep pixel dtype stable across CPU/GPU to avoid backend kernel errors.
         if "pixel_values" in inputs:
             inputs["pixel_values"] = inputs["pixel_values"].to(