Created
April 3, 2023 00:16
-
-
Save Birch-san/6902c1437fae9081561457d094242da5 to your computer and use it in GitHub Desktop.
Questionable softmax
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| from torch import FloatTensor | |
| vae_scale_factor = 8 | |
| typical_self_attn_key_length = (512/vae_scale_factor) * (512/vae_scale_factor) | |
| desired_self_attn_key_length = (200/vae_scale_factor) * (200/vae_scale_factor) | |
| key_length_factor=desired_self_attn_key_length/typical_self_attn_key_length if is_self_attn else 1. | |
| def softmax(x: FloatTensor, dim=-1) -> FloatTensor: | |
| key_tokens = x.size(-1) | |
| maxes = x.max(dim, keepdim=True).values | |
| diffs = x-maxes | |
| x_exp = diffs.exp() | |
| avg_diff = diffs.float().quantile(.175, dim=-1, keepdim=True).to(diffs.dtype) | |
| avg_diff_exp = avg_diff.exp() | |
| x_exp_sum = x_exp.sum(dim, keepdim=True) | |
| preferred_token_count = key_tokens/key_length_factor | |
| x_exp_sum = x_exp_sum + avg_diff_exp * (preferred_token_count-key_tokens) | |
| quotient = x_exp/x_exp_sum | |
| return quotient |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment