Coverage for transformer_lens/components/mlps/can_be_used_as_mlp.py: 94%

1"""Can Be Used as MLP component. 

2 

3This module serves as the base for everything within TransformerLens that can be used like an MLP. 

4This does not necessarily mean that every component extending this class will be an MLP, but 

5everything extending this class can be used interchangeably for an MLP. 

6""" 

7 

8from typing import Dict, Optional, Union 

9 

10import torch 

11import torch.nn as nn 

12from jaxtyping import Float 

13 

14from transformer_lens.components import LayerNorm, LayerNormPre 

15from transformer_lens.factories.activation_function_factory import ( 

16 ActivationFunctionFactory, 

17) 

18from transformer_lens.hook_points import HookPoint 

19from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

20from transformer_lens.utilities.activation_functions import ActivationFunction 

21 

22 

23class CanBeUsedAsMLP(nn.Module): 

24 # The actual activation function 

25 act_fn: ActivationFunction 

26 

27 # The full config object for the model 

28 cfg: HookedTransformerConfig 

29 

30 # The d mlp value pulled out of the config to make sure it always has a value 

31 d_mlp: int 

32 

33 # The middle hook point will be None unless it specifically should be used 

34 hook_mid: Optional[HookPoint] # [batch, pos, d_mlp] 

35 

36 # The layer norm component if the activation function is a layer norm 

37 ln: Optional[nn.Module] 

38 

39 def __init__(self, cfg: Union[Dict, HookedTransformerConfig]): 

40 """The base init for all MLP like components 

41 

42 Args: 

43 config (Union[Dict, HookedTransformerConfig]): The config for this instance 

44 

45 Raises: 

46 ValueError: If there is a misconfiguration 

47 """ 

48 super().__init__() 

49 self.cfg = HookedTransformerConfig.unwrap(cfg) 

50 if self.cfg.d_mlp is None: 50 ↛ 51line 50 didn't jump to line 51 because the condition on line 50 was never true

51 raise ValueError("d_mlp must be set to use an MLP") 

52 

53 self.d_mlp = self.cfg.d_mlp 

54 

55 def forward( 

56 self, x: Float[torch.Tensor, "batch pos d_model"] 

57 ) -> Float[torch.Tensor, "batch pos d_model"]: 

58 """The format for all forward functions for any MLP""" 

59 return x 

60 

61 def select_activation_function(self) -> None: 

62 """This function should be called by all components in their init to get everything needed 

63 for activation functions setup. 

64 

65 Raises: 

66 ValueError: If the configure activation function is not supported. 

67 """ 

68 

69 self.act_fn = ActivationFunctionFactory.pick_activation_function(self.cfg) 

70 

71 if self.cfg.is_layer_norm_activation(): 

72 self.hook_mid = HookPoint() 

73 if self.cfg.normalization_type == "LN": 

74 self.ln = LayerNorm(self.cfg, self.d_mlp) 

75 else: 

76 self.ln = LayerNormPre(self.cfg)