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

1"""Hooked Transformer RMS Norm Pre Component. 

2 

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

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 

15class RMSNormPre(nn.Module): 

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

17 """RMSNormPre - LayerNormPre without the centering and bias (RMS = Root Mean Square)""" 

18 super().__init__() 

19 self.cfg = HookedTransformerConfig.unwrap(cfg) 

20 self.eps = self.cfg.eps 

21 

22 # Adds a hook point for the normalisation scale factor 

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

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

25 

26 def forward( 

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

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

29 if self.cfg.dtype not in [torch.float32, torch.float64]: 29 ↛ 30line 29 didn't jump to line 30, because the condition on line 29 was never true

30 x = x.to(torch.float32) 

31 

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

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

34 ) 

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