Update handler.py

handler.py

@@ -6,9 +6,9 @@ class EndpointHandler():
   def __init__(self, path="ceiteach/chart-no-pretrain-llama31-unsloth"):
     base_model_name = "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
     model_name = "ceiteach/chart-no-pretrain-llama31-unsloth"
-    model = AutoModelForCausalLM.from_pretrained(base_model_name)
+    model = AutoModelForCausalLM.from_pretrained(base_model_name, use_safetensors=True, load_in_4bit=True, device_map="auto")
     model.load_adapter(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name, max_model_length=4096)
     self.model = model
     self.tokenizer = tokenizer