llama.cpp issue 28185: --tensor-split causes intermittent output degeneration on long-context MoE inference

Fields

id

sources/mgpu-llamacpp-28185-tensor-split-degeneration

URL

https://github.com/ggml-org/llama.cpp/issues/28185

Publisher

ggml-org/llama.cpp

Kind

repo

Published

2026-09-01

Accessed

2026-10-03

notes

--split-mode layer --tensor-split 4,1 --n-cpu-moe 34 on Qwen3-Coder-Next at 64K context: intermittent repetition degeneration (1 of 2 raw-API runs), 0 collapses with --tensor-split removed. Also reports MoE experts staying on one device regardless of the split ratio.

Notes

--split-mode layer --tensor-split 4,1 --n-cpu-moe 34 on Qwen3-Coder-Next at 64K context: intermittent repetition degeneration (1 of 2 raw-API runs), 0 collapses with --tensor-split removed. Also reports MoE experts staying on one device regardless of the split ratio.

Cross-references

This record references no other records.