fix

train_llama.py

@@ -16,8 +16,10 @@ print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")
 MODEL_ID = "meta-llama/Llama-2-7b-hf"
 tokenizer = LlamaTokenizer.from_pretrained(MODEL_ID)
 
+# Set pad token to existing <|endoftext|> (ID 2) instead of adding a new one
 if tokenizer.pad_token is None:
-    tokenizer.add_special_tokens({'pad_token': '[PAD]'})
+    tokenizer.pad_token = tokenizer.eos_token  # Use <|endoftext|> as pad token
+    tokenizer.pad_token_id = tokenizer.eos_token_id  # Should be 2
 
 # Quantization config
 quantization_config = BitsAndBytesConfig(load_in_8bit=True)