mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2026-01-04 15:14:19 +00:00
7548ca5a54
* add normalize function to value_head in bloom rm * add normalization to value_function in gpt_rm * add normalization to value_head of opt_rm * add Anthropic/hh-rlhf dataset * Update __init__.py * Add LogExpLoss in RM training * Update __init__.py * update rm trainer to use acc as target * update example/train_rm * Update train_rm.sh * code style * Update README.md * Update README.md * add rm test to ci * fix tokenier * fix typo * change batchsize to avoid oom in ci * Update test_ci.sh
39 lines
1.2 KiB
Python
39 lines
1.2 KiB
Python
from typing import Optional
|
|
|
|
import torch.nn as nn
|
|
from transformers import OPTConfig, OPTModel
|
|
|
|
from ..base import RewardModel
|
|
|
|
|
|
class OPTRM(RewardModel):
|
|
"""
|
|
OPT Reward model.
|
|
|
|
Args:
|
|
pretrained (str): Pretrained model name or path.
|
|
config (OPTConfig): Model config.
|
|
checkpoint (bool): Enable gradient checkpointing.
|
|
lora_rank (int): Rank of the low-rank approximation.
|
|
lora_train_bias (str): LoRA bias training mode.
|
|
"""
|
|
|
|
def __init__(self,
|
|
pretrained: Optional[str] = None,
|
|
config: Optional[OPTConfig] = None,
|
|
checkpoint: bool = False,
|
|
lora_rank: int = 0,
|
|
lora_train_bias: str = 'none') -> None:
|
|
if pretrained is not None:
|
|
model = OPTModel.from_pretrained(pretrained)
|
|
elif config is not None:
|
|
model = OPTModel(config)
|
|
else:
|
|
model = OPTModel(OPTConfig())
|
|
if checkpoint:
|
|
model.gradient_checkpointing_enable()
|
|
|
|
value_head = nn.Linear(model.config.word_embed_proj_dim, 1)
|
|
value_head.weight.data.normal_(mean=0.0, std=1/(model.config.word_embed_proj_dim + 1))
|
|
super().__init__(model, value_head, lora_rank, lora_train_bias)
|