Newer version of KoboldCPP cannot fit as many layers in VRAM?

[PhoenixGameDevelopment]

I am using KoboldCPP on my linux system, and I have noticed that the amount of layers that I can offload to VRAM is a lot less (And a lot less reliable) on more recent versions.

I have three GPU's (A6000 Ampere, 2 3090s).

With version 1.76 I can offload, for example, 83 layers of a particular model. I am also using 2.0,1.0,1.0 as my tensorsplit.

With version 1.93.1, with the same settings, I first of all have to change my tensor split to 1.0,2.0,1.0 (It seems that the order has changed?).

In addition, I can only load between 76 and 80 layers of the same model before I get OOM's. Not only that, but when I successfully loaded the model with 80 layers offloaded, I then restarted, and got an OOM with 80 layers, I had to reduce the layer count (with no other changes made).

I tried again under identical conditions with version 1.76, and 83 layers of the model loaded with no issues.

Can anyone confirm this behaviour?