Does llama cpp split mode tensor cause issues?

A user reports that using tensor split mode in llama.cpp causes looping issues with tool calls and reasoning traces when running Qwen 27B and Gemma 4 26B (MoE) models across an RTX 5080 and two RTX 5060 Ti GPUs.

• The issue was observed specifically with the tensor split mode setting.
• Models tested include Qwen 27B and Gemma 4 26B (MoE).
• Hardware configuration involved one RTX 5080 and two RTX 5060 Ti cards.
• Layer split mode functioned correctly without these errors.

The author seeks to determine if this is a known issue or understand the underlying cause of the looping behavior.