Fix LoRA tokenizer for Llama and chatglm (#11186)

* Set pad_token to eos_token if it's None. Otherwise, use model config.

python/llm/example/CPU/QLoRA-FineTuning/alpaca-qlora/alpaca_qlora_finetuning_cpu.py

@@ -39,7 +39,7 @@ import transformers
 from datasets import load_dataset
 import accelerate
 
-from transformers import LlamaTokenizer
+from transformers import AutoTokenizer
 from peft import (
     get_peft_model_state_dict,
     set_peft_model_state_dict,
@@ -198,13 +198,12 @@ def train(
     model = model.to("cpu")
     print(f"Model moved to rank {os.environ.get('LOCAL_RANK')}")
 
-    tokenizer = LlamaTokenizer.from_pretrained(base_model)
+    tokenizer = AutoTokenizer.from_pretrained(base_model)
     print(f"Tokenizer loaded on rank {os.environ.get('LOCAL_RANK')}")
 
-    tokenizer.pad_token_id = (
-        0  # unk. we want this to be different from the eos token
-    )
-    tokenizer.padding_side = "left"  # Allow batched inference
+    # For Llama family
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
 
     print(model)
 

python/llm/example/GPU/LLM-Finetuning/LoRA/alpaca_lora_finetuning.py

@@ -180,10 +180,9 @@ def train(
     tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
     print(f"Tokenizer loaded on rank {os.environ.get('LOCAL_RANK')}")
 
-    tokenizer.pad_token_id = (
-        0  # unk. we want this to be different from the eos token
-    )
-    tokenizer.padding_side = "left"  # Allow batched inference
+    # For Llama family
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
 
     print(model)
 

python/llm/example/GPU/LLM-Finetuning/QA-LoRA/alpaca_qalora_finetuning.py

@@ -192,11 +192,9 @@ def train(
     tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
     print(f"Tokenizer loaded on rank {os.environ.get('LOCAL_RANK')}")
 
-    tokenizer.pad_token_id = (
-        0  # unk. we want this to be different from the eos token
-    )
-    tokenizer.padding_side = "left"  # Allow batched inference
-
+    # For Llama family
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
     print(model)
 
     # Prepare a IPEX-LLM compatible Peft model

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

@@ -192,6 +192,9 @@ def train(
 
     tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
     print(f"Tokenizer loaded on rank {os.environ.get('LOCAL_RANK')}")
+    # For Llama family
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
     print(model)
 
     # Prepare a IPEX-LLM compatible Peft model