From 657813346b1709303b61fee4d3cec6b93f48cfc5 Mon Sep 17 00:00:00 2001 From: araleza <70412719+araleza@users.noreply.github.com> Date: Sun, 31 Aug 2025 18:51:50 +0100 Subject: [PATCH] Changed u16 pow() factor from 16.0 to 8.0, seems to better distribute the exp_avg[_sq] values --- library/adamw_fused.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/library/adamw_fused.py b/library/adamw_fused.py index d94df654..b439c5d9 100644 --- a/library/adamw_fused.py +++ b/library/adamw_fused.py @@ -73,7 +73,7 @@ def adamw_offload_step_param(self, p, group): # the min and max of the results were noted, and then the values were scaled # to the 0-65535 range for storage. This process is reversed here. - u16power = 16.0 # This value worked acceptably in testing to spread the values more evenly + u16power = 8.0 # This value worked acceptably in testing to spread the values more evenly exp_avg_min = state['exp_avg_min'] exp_avg_max = state['exp_avg_max']