Coverage for transformer_lens/components/rms_norm.py: 81%

1"""Hooked Transformer RMS Norm Component. 

2 

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

4""" 

5 

6from typing import Dict, Optional, Union 

7 

8import torch 

9import torch.nn as nn 

10from jaxtyping import Float 

11 

12from transformer_lens.hook_points import HookPoint 

13from transformer_lens.HookedTransformerConfig import HookedTransformerConfig 

14 

15 

16class RMSNorm(nn.Module): 

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

18 """ 

19 RMSNorm - LayerNorm without the centering and bias (RMS = Root Mean Square) 

20 

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

22 """ 

23 super().__init__() 

24 self.cfg = HookedTransformerConfig.unwrap(cfg) 

25 self.eps = self.cfg.eps 

26 if length is None: 26 ↛ 29line 26 didn't jump to line 29 because the condition on line 26 was always true

27 self.length = self.cfg.d_model 

28 else: 

29 self.length = length 

30 

31 self.w = nn.Parameter(torch.ones(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 self.hook_normalized = HookPoint() # [batch, pos, length] 

36 

37 def forward( 

38 self, x: Float[torch.Tensor, "batch pos length"] 

39 ) -> Float[torch.Tensor, "batch pos length"]: 

40 if self.cfg.dtype not in [torch.float32, torch.float64]: 40 ↛ 41line 40 didn't jump to line 41 because the condition on line 40 was never true

41 x = x.to(torch.float32) 

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

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

44 ) 

45 x = self.hook_normalized(x / scale).to(self.cfg.dtype) # [batch, pos, length] 

46 

47 if x.device != self.w.device: 47 ↛ 48line 47 didn't jump to line 48 because the condition on line 47 was never true

48 self.to(x.device) 

49 

50 return x * self.w