Upload app.py with huggingface_hub

app.py

@@ -72,17 +72,70 @@ def download_and_load_model():
             }
 
         try:
-            # Try to load with the config from checkpoint
+            # First try: Use the checkpoint's config if available
             model = MultimodalGemmaLightning.load_from_checkpoint(
                 checkpoint_path,
                 config=config,
                 strict=False,
                 map_location="cuda" if torch.cuda.is_available() else "cpu"
             )
-        except Exception as e:
-            print(f"Failed to load with saved config: {e}")
-            # Fallback: Load without any base model dependencies
-            return "❌ Model loading failed - checkpoint may be incompatible with current setup"
+            print("✅ Loaded with checkpoint config")
+        except Exception as e1:
+            print(f"Failed with checkpoint config: {e1}")
+            try:
+                # Second try: Minimal config with no quantization
+                minimal_config = {
+                    "model": {
+                        "gemma_model_name": "microsoft/DialoGPT-small",  # Even smaller model
+                        "vision_model_name": "openai/clip-vit-base-patch32",  # Smaller CLIP
+                        "use_4bit": False,  # No quantization
+                        "projector_hidden_dim": 512,
+                        "lora": {"r": 8, "alpha": 16, "dropout": 0.1, "target_modules": ["q_proj", "v_proj"]}
+                    },
+                    "special_tokens": {"image_token": "<image>"},
+                    "training": {"projector_lr": 1e-3, "lora_lr": 1e-4}
+                }
+                model = MultimodalGemmaLightning.load_from_checkpoint(
+                    checkpoint_path,
+                    config=minimal_config,
+                    strict=False,
+                    map_location="cuda" if torch.cuda.is_available() else "cpu"
+                )
+                print("✅ Loaded with minimal config")
+            except Exception as e2:
+                print(f"Failed with minimal config: {e2}")
+                try:
+                    # Third try: Direct state dict loading
+                    print("Attempting direct state dict loading...")
+                    # Create a dummy model just to get the structure
+                    dummy_config = {
+                        "model": {
+                            "gemma_model_name": "microsoft/DialoGPT-small",
+                            "vision_model_name": "openai/clip-vit-base-patch32",
+                            "use_4bit": False,
+                            "projector_hidden_dim": 512,
+                        },
+                        "special_tokens": {"image_token": "<image>"},
+                        "training": {"projector_lr": 1e-3, "lora_lr": 1e-4}
+                    }
+                    model = MultimodalGemmaLightning(dummy_config)
+
+                    # Load only compatible weights
+                    checkpoint_state = checkpoint['state_dict']
+                    model_state = model.state_dict()
+
+                    # Filter and load compatible weights
+                    compatible_weights = {}
+                    for key, value in checkpoint_state.items():
+                        if key in model_state and model_state[key].shape == value.shape:
+                            compatible_weights[key] = value
+
+                    model.load_state_dict(compatible_weights, strict=False)
+                    print(f"✅ Loaded {len(compatible_weights)} compatible weights")
+
+                except Exception as e3:
+                    print(f"All loading methods failed: {e3}")
+                    return f"❌ Model loading failed - checkpoint incompatible. Last error: {str(e3)}"
         model.eval()
 
         # Move to appropriate device