Update app.py

app.py

@@ -4,7 +4,7 @@
 # - Uses qwen-vl-utils + AutoProcessor (multimodal) with trust_remote_code, use_fast=False
 # - Deterministic decoding for stable eval
 # - ZeroGPU only during inference (ALL CUDA work happens inside @spaces.GPU functions)
-# - Includes a ZeroGPU-safe debug tool: "LoRA ON vs OFF" comparison
+# - Debug tools: (1) LoRA ON vs OFF toggle; (2) LoRA ON vs pure BASE
 
 import os
 import logging
@@ -139,6 +139,8 @@ def try_load_model():
             if hasattr(model, "get_active_adapters"):
                 logger.info(f"Active adapters: {model.get_active_adapters()}")
             logger.info(f"PEFT config present: {hasattr(model, 'peft_config')}")
+            if hasattr(model, "peft_config"):
+                logger.info(f"PEFT keys: {list(model.peft_config.keys())}")
         except Exception:
             pass
         logger.info("LoRA adapters attached and active (not merged).")
@@ -204,6 +206,21 @@ def compare_with_without_lora(model, inputs):
 
     return with_lora, without_lora
 
+def load_base_only_model():
+    """
+    Load a fresh BASE model (no adapters) for direct comparison.
+    Call ONLY inside GPU-decorated functions.
+    """
+    model = AutoModelForVision2Seq.from_pretrained(
+        BASE_MODEL_ID,
+        torch_dtype=torch.float16,
+        device_map="cuda",
+        trust_remote_code=True,
+        low_cpu_mem_usage=True,
+    )
+    model.eval()
+    return model
+
 # ---------------------------
 # Inference (ZeroGPU)
 # ---------------------------
@@ -257,6 +274,42 @@ def debug_compare_lora(image: Optional[Image.Image], question: str) -> str:
             del model
         torch.cuda.empty_cache()
 
+# ---------------------------
+# Debug (ZeroGPU-safe): LoRA ON vs pure BASE comparison
+# ---------------------------
+@spaces.GPU(duration=ZGPU_DURATION)
+def debug_compare_vs_base(image: Optional[Image.Image], question: str) -> str:
+    if image is None:
+        return "Please upload an image first."
+    lo_model = None
+    base_model = None
+    try:
+        inputs = build_inputs(image, question)
+
+        lo_model, warn = try_load_model()
+        if lo_model is None:
+            return f"Load error (LoRA): {warn}"
+        if warn:
+            logger.warning(warn)
+
+        base_model = load_base_only_model()
+
+        lora_text = _generate_text(lo_model, inputs)
+        base_text = _generate_text(base_model, inputs)
+
+        return (
+            "=== LoRA ON (adapters attached) ===\n" + lora_text +
+            "\n\n=== BASE ONLY ===\n" + base_text
+        )
+    except Exception as e:
+        logger.exception("Debug compare vs base failed")
+        return f"Debug error: {e}"
+    finally:
+        for m in (lo_model, base_model):
+            if m is not None:
+                del m
+        torch.cuda.empty_cache()
+
 # ---------------------------
 # UI
 # ---------------------------
@@ -283,9 +336,11 @@ def create_interface() -> gr.Blocks:
 
         # Debug: LoRA ON vs OFF (GPU-decorated function)
         with gr.Row():
-            debug_btn = gr.Button("Debug: Compare LoRA ON vs OFF")
-        debug_out = gr.Textbox(label="Debug Output", lines=14)
-        debug_btn.click(fn=debug_compare_lora, inputs=[image_input, question_input], outputs=debug_out, queue=True)
+            debug_btn1 = gr.Button("Debug: Compare LoRA ON vs OFF")
+            debug_btn2 = gr.Button("Debug: Compare LoRA ON vs BASE")
+        debug_out = gr.Textbox(label="Debug Output", lines=16)
+        debug_btn1.click(fn=debug_compare_lora, inputs=[image_input, question_input], outputs=debug_out, queue=True)
+        debug_btn2.click(fn=debug_compare_vs_base, inputs=[image_input, question_input], outputs=debug_out, queue=True)
 
         demo.queue()
         gr.Markdown("Tips: Ensure good lighting and focus. Avoid uploading personally identifying information.")