rev app.py

app.py

@@ -28,37 +28,38 @@ def load_model_and_processor():
     global model, processor
     print("กำลังโหลดโมเดลและ processor...")
     try:
+        ### 
+        from unsloth import FastVisionModel
+        from transformers import AutoModelForVision2Seq, TextStreamer
+        ###
         # Model paths
-        base_model_path = "meta-llama/Llama-3.2-11B-Vision-Instruct"
-        adapter_path = "Aekanun/Llama-3.2-11B-Vision-Instruct-XRay"
+        ### base_model_path = "meta-llama/Llama-3.2-11B-Vision-Instruct"
+        ### adapter_path = "Aekanun/Llama-3.2-11B-Vision-Instruct-XRay"
 
         # Load processor from base model
         print("กำลังโหลด processor...")
-        processor = AutoProcessor.from_pretrained(
-            base_model_path, 
-            use_auth_token=True
-        )
+        ###processor = AutoProcessor.from_pretrained(
+        ###    base_model_path, 
+        ###    use_auth_token=True
+        ###)
 
-        # Load base model
-        print("กำลังโหลด base model...")
-        base_model = AutoModelForVision2Seq.from_pretrained(
-            base_model_path,
-            device_map="auto",
-            torch_dtype=torch.float16,  # เปลี่ยนกลับเป็น float16
-            trust_remote_code=True,
-            use_auth_token=True
+        base_model, tokenizer = FastVisionModel.from_pretrained(
+            "unsloth/Llama-3.2-11B-Vision-Instruct",
+            use_gradient_checkpointing = "unsloth"
         )
 
-        # Load adapter
-        print("กำลังโหลด adapter...")
-        model = PeftModel.from_pretrained(
-            base_model,
-            adapter_path,
-            device_map="auto",  # ให้จัดการ device map อัตโนมัติ
-            torch_dtype=torch.float16,
-            use_auth_token=True
+        print("โหลด base model และ tokenizer สำเร็จ กำลังโหลดโมเดลที่ fine-tune...")
+        
+        # ปิด FastVisionModel และโหลด model โดยตรง
+        from transformers import AutoModelForVision2Seq
+        model = AutoModelForVision2Seq.from_pretrained(
+            "Aekanun/Llama-3.2-11B-Vision-Instruct-XRay",
+            load_in_4bit = True,
+            device_map = {'': torch.cuda.current_device()},
+            torch_dtype = torch.float16
         )
         
+        FastVisionModel.for_inference(model)
         print("โหลดโมเดลสำเร็จ!")
         return True
     except Exception as e: