How to give a user a GPU limit per node (for example 3 GPUs on gpu1 and 3 on gpu2) instead
of one limit across the whole cluster, and why the setup tried on 18 September 2026 was removed
again the same day.
Since 10 August 2026 every job on the Koios compute nodes gets its own /tmp, on the node's
NVMe disk, that no other job can see and that is removed when the job ends. The rollout started
on 30 July.
On 6 May 2026 two of gpu2's eight A100s were split into small MIG instances. The trial was rolled back on 13–14 May. The CPU topology changes made at the same time stayed.