Coverage for transformer_lens/components/layer_norm.py: 93%

1"""Hooked Transformer Layer Norm Component. 

2 

3This module contains all the component :class:`LayerNorm`. 

4""" 

5from typing import Dict, Optional, Union 

6 

7import torch 

8import torch.nn as nn 

9from jaxtyping import Float 

10 

11from transformer_lens.hook_points import HookPoint 

12from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

13 

14 

15class LayerNorm(nn.Module): 

16 def __init__(self, cfg: Union[Dict, HookedTransformerConfig], length: Optional[int] = None): 

17 """ 

18 LayerNorm with optional length parameter 

19 

20 length (Optional[int]): If the dimension of the LayerNorm. If not provided, assumed to be d_model 

21 """ 

22 super().__init__() 

23 self.cfg = HookedTransformerConfig.unwrap(cfg) 

24 self.eps = self.cfg.eps 

25 if length is None: 

26 self.length = self.cfg.d_model 

27 else: 

28 self.length = length 

29 

30 self.w = nn.Parameter(torch.ones(self.length, dtype=self.cfg.dtype)) 

31 self.b = nn.Parameter(torch.zeros(self.length, dtype=self.cfg.dtype)) 

32 

33 # Adds a hook point for the normalisation scale factor 

34 self.hook_scale = HookPoint() # [batch, pos, 1] 

35 # Hook_normalized is on the LN output 

36 self.hook_normalized = HookPoint() # [batch, pos, length] 

37 

38 def forward( 

39 self, 

40 x: Union[ 

41 Float[torch.Tensor, "batch pos d_model"], 

42 Float[torch.Tensor, "batch pos head_index d_model"], 

43 ], 

44 ) -> Union[ 

45 Float[torch.Tensor, "batch pos d_model"], 

46 Float[torch.Tensor, "batch pos head_index d_model"], 

47 ]: 

48 if self.cfg.dtype not in [torch.float32, torch.float64]: 48 ↛ 49line 48 didn't jump to line 49, because the condition on line 48 was never true

49 x = x.to(torch.float32) 

50 

51 x = x - x.mean(-1, keepdim=True) # [batch, pos, length] 

52 scale: Float[torch.Tensor, "batch pos 1"] = self.hook_scale( 

53 (x.pow(2).mean(-1, keepdim=True) + self.eps).sqrt() 

54 ) 

55 x = x / scale # [batch, pos, length] 

56 return self.hook_normalized(x * self.w + self.b).to(self.cfg.dtype)