LLM: update qlora alpaca example to change lora usage (#9835)

* update example

* fix style

python/llm/example/GPU/QLoRA-FineTuning/alpaca-qlora/alpaca_qlora_finetuning.py

@@ -196,10 +196,18 @@ def train(
     else:
         # According to the QLoRA paper, using "nf4" could yield better model quality than "int4"
         # Default 4-bit format for qa-lora is sym_int4
+        if training_mode == "lora":
+            model = AutoModelForCausalLM.from_pretrained(
+                base_model,
+                load_in_low_bit="bf16",
+                optimize_model=False,
+                torch_dtype=torch.bfloat16,
+                modules_to_not_convert=["lm_head"],
+            )
+        else:
+            # use bnb_config for qlora/qalora/relora, which use 4bit for base model
             if training_mode == "qalora":
                 low_bit_format = "int4"
-        elif training_mode == "lora":
-            low_bit_format = "bf16"
             else:
                 low_bit_format = "nf4"
             bnb_config = BitsAndBytesConfig(