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""" 

7from typing import Dict, Optional, Union 

8 

9import torch 

10import torch.nn as nn 

11from jaxtyping import Float 

12 

13from transformer_lens.components import LayerNorm, LayerNormPre 

14from transformer_lens.factories.activation_function_factory import ( 

15 ActivationFunctionFactory, 

16) 

17from transformer_lens.hook_points import HookPoint 

18from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

19from transformer_lens.utilities.activation_functions import ActivationFunction 

20 

21 

22class CanBeUsedAsMLP(nn.Module): 

23 # The actual activation function 

24 act_fn: ActivationFunction 

25 

26 # The full config object for the model 

27 cfg: HookedTransformerConfig 

28 

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

30 d_mlp: int 

31 

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

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

34 

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

36 ln: Optional[nn.Module] 

37 

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

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

40 

41 Args: 

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

43 

44 Raises: 

45 ValueError: If there is a misconfiguration 

46 """ 

47 super().__init__() 

48 self.cfg = HookedTransformerConfig.unwrap(cfg) 

49 if self.cfg.d_mlp is None: 49 ↛ 50line 49 didn't jump to line 50, because the condition on line 49 was never true

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

51 

52 self.d_mlp = self.cfg.d_mlp 

53 

54 def forward( 

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

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

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

58 return x 

59 

60 def select_activation_function(self) -> None: 

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

62 for activation functions setup. 

63 

64 Raises: 

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

66 """ 

67 

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

69 

70 if self.cfg.is_layer_norm_activation(): 

71 self.hook_mid = HookPoint() 

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

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

74 else: 

75 self.ln = LayerNormPre(self.cfg)