Coverage for transformer_lens/components/layer_norm_pre.py: 90%

1"""Hooked Transformer Layer Norm Pre Component. 

2 

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

4""" 

5from typing import Dict, 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 

15# LayerNormPre 

16# I fold the LayerNorm weights and biases into later weights and biases. 

17# This is just the 'center and normalise' part of LayerNorm 

18# Centering is equivalent to just deleting one direction of residual space, 

19# and is equivalent to centering the weight matrices of everything writing to the residual stream 

20# Normalising is a funkier non-linear operation, that projects the residual stream onto the unit hypersphere 

21class LayerNormPre(nn.Module): 

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

23 """LayerNormPre - the 'center and normalise' part of LayerNorm. Length is 

24 normally d_model, but is d_mlp for softmax. Not needed as a parameter. This 

25 should only be used in inference mode after folding in LayerNorm weights""" 

26 super().__init__() 

27 self.cfg = HookedTransformerConfig.unwrap(cfg) 

28 self.eps = self.cfg.eps 

29 

30 # Adds a hook point for the normalisation scale factor 

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

32 # Hook Normalized captures LN output - here it's a vector with std 1 and mean 0 

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

34 

35 def forward( 

36 self, 

37 x: Union[ 

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

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

40 ], 

41 ) -> Union[ 

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

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

44 ]: 

45 if self.cfg.dtype not in [torch.float32, torch.float64]: 45 ↛ 46line 45 didn't jump to line 46, because the condition on line 45 was never true

46 x = x.to(torch.float32) 

47 

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

49 scale: Union[ 

50 Float[torch.Tensor, "batch pos 1"], 

51 Float[torch.Tensor, "batch pos head_index 1"], 

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

53 return self.hook_normalized(x / scale).to(self.cfg.dtype)