ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Fix LoRA tokenizer for Llama and chatglm (#11186)

Browse source 
* Set pad_token to eos_token if it's None. Otherwise, use model config.
This commit is contained in:
Qiyuan Gong 2024-06-03 15:35:38 +08:00 committed by GitHub
parent 3eb13ccd8c
commit 15a6205790
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 14 additions and 15 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

11
python/llm/example/CPU/QLoRA-FineTuning/alpaca-qlora/alpaca_qlora_finetuning_cpu.py
View file

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

7
python/llm/example/GPU/LLM-Finetuning/LoRA/alpaca_lora_finetuning.py
View file

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

8
python/llm/example/GPU/LLM-Finetuning/QA-LoRA/alpaca_qalora_finetuning.py
View file

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

3
python/llm/example/GPU/LLM-Finetuning/QLoRA/alpaca-qlora/alpaca_qlora_finetuning.py
View file

@ -192,6 +192,9 @@ def train(
tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True) tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
print(f"Tokenizer loaded on rank {os.environ.get('LOCAL_RANK')}") 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) print(model)
# Prepare a IPEX-LLM compatible Peft model # Prepare a IPEX-LLM compatible Peft model