colossalai

1
from typing import Optional
2

3
from transformers.models.opt.configuration_opt import OPTConfig
4
from transformers.models.opt.modeling_opt import OPTForCausalLM
5

6
from ..base import Actor
7

8

9
class OPTActor(Actor):
10
    """
11
    OPT Actor model.
12

13
    Args:
14
        pretrained (str): Pretrained model name or path.
15
        config (OPTConfig): Model config.
16
        checkpoint (bool): Enable gradient checkpointing.
17
        lora_rank (int): Rank of the low-rank approximation.
18
        lora_train_bias (str): LoRA bias training mode.
19
    """
20

21
    def __init__(
22
        self,
23
        pretrained: Optional[str] = None,
24
        config: Optional[OPTConfig] = None,
25
        checkpoint: bool = False,
26
        lora_rank: int = 0,
27
        lora_train_bias: str = "none",
28
    ) -> None:
29
        if pretrained is not None:
30
            model = OPTForCausalLM.from_pretrained(pretrained)
31
        elif config is not None:
32
            model = OPTForCausalLM(config)
33
        else:
34
            model = OPTForCausalLM(OPTConfig())
35
        if checkpoint:
36
            model.gradient_checkpointing_enable()
37
        super().__init__(model, lora_rank, lora_train_bias)
38