本地processor

gradio_ui/agent/vision_agent.py

@@ -33,35 +33,19 @@ class VisionAgent:
         
         # load the image caption model and processor
         self.caption_processor = AutoProcessor.from_pretrained(
-            "microsoft/Florence-2-base", 
+            "processor", 
             trust_remote_code=True
         )
         
-        # load the model according to the device type
         try:
-            if self.device.type == 'cuda':
+            self.caption_model = AutoModelForCausalLM.from_pretrained(
-                # CUDA device uses float16
+                caption_model_path, 
-                self.caption_model = AutoModelForCausalLM.from_pretrained(
+                torch_dtype=torch.float32,
-                    caption_model_path, 
+                trust_remote_code=True
-                    torch_dtype=torch.float16,
+            ).to(self.device)
-                    trust_remote_code=True
-                ).to(self.device)
-            elif self.device.type == 'mps':
-                # MPS device uses float32 (MPS has limited support for float16)
-                self.caption_model = AutoModelForCausalLM.from_pretrained(
-                    caption_model_path, 
-                    torch_dtype=torch.float32,
-                    trust_remote_code=True
-                ).to(self.device)
-            else:
-                # CPU uses float32
-                self.caption_model = AutoModelForCausalLM.from_pretrained(
-                    caption_model_path, 
-                    torch_dtype=torch.float32,
-                    trust_remote_code=True
-                ).to(self.device)
             
         except Exception as e:
+            print(f"Model loading failed for path: {caption_model_path}")
             raise e
         self.prompt = "<CAPTION>"
         

processor/preprocessor_config.json

@@ -0,0 +1,33 @@
+{
+  "auto_map": {
+    "AutoProcessor": "microsoft/Florence-2-base--processing_florence2.Florence2Processor"
+  },
+  "crop_size": {
+    "height": 768,
+    "width": 768
+  },
+  "do_center_crop": false,
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.485,
+    0.456,
+    0.406
+  ],
+  "image_processor_type": "CLIPImageProcessor",
+  "image_seq_length": 577,
+  "image_std": [
+    0.229,
+    0.224,
+    0.225
+  ],
+  "processor_class": "Florence2Processor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 768,
+    "width": 768
+  }
+}