Fix alternating roles and use dtype instead of torch_dtype

train.py

@@ -84,7 +84,7 @@ if tokenizer.pad_token is None:
 print("Loading model...")
 model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID,
-    torch_dtype=torch.bfloat16,
+    dtype=torch.bfloat16,
     device_map="auto",
 )