Should I use both of CPT lora and SFT lora for inference?

[BlueAnthony]

Reminder

• I have read the README and searched the existing issues.

System Info

LLaMA Factory: version 0.7.2.dev0

Reproduction

The following is my llama3_lora_sft.yaml.

model

model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
cache_dir: /home/llama-factory/
adapter_name_or_path: saves/llama3-8b/lora/cpt_bat32_20epoch_outlook

method

stage: sft
do_train: true
finetuning_type: lora
lora_target: q_proj,v_proj

dataset

dataset: openorca,lima
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/llama3-8b/lora/sft_bat32_10epoch_outlook
logging_steps: 10
save_steps: 100
plot_loss: true
overwrite_output_dir: true

train

per_device_train_batch_size: 8
gradient_accumulation_steps: 4
learning_rate: 0.0001
num_train_epochs: 10.0
lr_scheduler_type: cosine
warmup_steps: 0.1
fp16: true

eval

val_size: 0.1
per_device_eval_batch_size: 1
evaluation_strategy: steps
eval_steps: 500

Expected behavior

Can I use the LoRA generated from the CPT stage for SFT like the above yaml?
And during inference, should I list the paths for these two LoRAs in the YAML file in sequence?
It seems that when both the CPT and SFT LoRAs are added simultaneously, the inference results become talking nonsense. However, if only the SFT LoRA is used, the responses are more normal.

Others

No response