Imbalanced VRAM usage between two GPUs in llama.cpp. Anyone successfully solve this?

A user on reddit.com is experiencing imbalanced VRAM usage between two GPUs when running llama.cpp, despite adjusting the --tensor-split (-ts) parameter. They note that at least 1GB of VRAM remains unusable, and even a slight adjustment to the split significantly shifts the weight to the other GPU. The user provided their llama-server command, which includes parameters like --batch-size 2048, --ctx-size 192000, --gpu-layers all, and --tensor-split 499,501, seeking a solution to this VRAM imbalance.

This report details a specific VRAM imbalance issue in llama.cpp with a --tensor-split of 499,501, unlike general discussions of multi-GPU setups.