Files
LlamaFactory/src/llmtuner/tuner/ppo/utils.py
hiyouga be4d2822ea fix PPO trainer #551 , update readme
Former-commit-id: faead74849470cebae9e37cde5fab2a71b32aa43
2023-08-18 11:43:10 +08:00

18 lines
781 B
Python

from typing import TYPE_CHECKING, Literal
if TYPE_CHECKING:
from trl import AutoModelForCausalLMWithValueHead
def replace_model(model: "AutoModelForCausalLMWithValueHead", target: Literal["default", "reward"]) -> None:
if target == "reward": # save default head temporarily
valuehead_state_dict = model.v_head.state_dict()
setattr(model, "default_head_weight", valuehead_state_dict["summary.weight"])
setattr(model, "default_head_bias", valuehead_state_dict["summary.bias"])
model.pretrained_model.set_adapter(target) # set the LoRA adapter to be active
model.v_head.load_state_dict({
"summary.weight": getattr(model, "{}_head_weight".format(target)),
"summary.bias": getattr(model, "{}_head_bias".format(target))
})