Update handler.py

handler.py

@@ -2,11 +2,15 @@ from unsloth import FastLanguageModel
 import torch
 import os
 
+
 class EndpointHandler:
     def __init__(self, path=""):
         # ระบุชื่อโมเดลที่คุณต้องการใช้งาน
         model_name = "defog/llama-3-sqlcoder-8b"
-
+        
+        torch.cuda.empty_cache() 
+        os.environ[ "PYTORCH_CUDA_ALLOC_CONF" ] = "expandable_segments:True"
+        
         # Configuration settings
         max_seq_length = 2048
         dtype = None  # Keep as None, you can change later if needed