LDLINGLINGLING
commented
5 months ago if training_args.use_lora: if training_args.use_lora and training_args.tune_llm: raise ValueError("The model cannot simultaneously adjust LLM parameters and apply LoRA.")
rank0_print("Currently using LoRA for fine-tuning the MiniCPM-V model.")
for name, param in model.llm.named_parameters():
param.requires_grad = False
lora_config = LoraConfig(
r=lora_args.lora_r,
lora_alpha=lora_args.lora_alpha,
target_modules=lora_args.lora_target_modules,
lora_dropout=lora_args.lora_dropout,
bias=lora_args.lora_bias,
layers_to_transform=lora_args.lora_layers_to_transform,
task_type="CAUSAL_LM",
)
if not hasattr(model, 'get_input_embeddings'):
def get_input_embeddings(self):
return self.llm.get_input_embeddings()
model.get_input_embeddings = MethodType(get_input_embeddings, model)
if lora_args.q_lora:
model = prepare_model_for_kbit_training(
model, use_gradient_checkpointing=training_args.gradient_checkpointing
)
model = get_peft_model(model, lora_config)
model.base_model.resampler.requires_grad_(True)This is the new code I saw. The resample module is fully fine-tuned in lora, and the language model is highly efficient fine-tuning in lora.
qyc-98
是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?
该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?
当前行为 | Current Behavior
With default config resampler is neither a LoRA module nor unfrozen for full finetuning. Seems to be the fixable by adding
modules_to_save=['resampler'],toLoraConfig.期望行为 | Expected Behavior
Resampler is finetuned during LoRA training.
复现方法 | Steps To Reproduce
run
finetune_lora.sh运行环境 | Environment
备注 | Anything else?
No response