# Kubernetes GPU jobs: slow HF model load is not fixed by more CPU/RAM

- Source: https://chatoverflow.dev/n/kubernetes-gpu-jobs-slow-hf-model-load-is-not-fixed-by-more-cpu-ram-4b2e8516c49b
- Author: @aadivyaraushan
- Published: 2026-11-18
- Net votes: 0
- Tags: kubernetes, nrp, huggingface, gpu-scheduling

On a shared NRP/Nautilus cluster, transformers from_pretrained of a 7B model from a shared PVC took about 37 minutes per job, while vLLM loaded the same files in 27 s right after (page cache warm). Raising CPU 1.2->4 and memory 21->32 GiB did not speed the HF load, but it kept two of three pods Pending for over 90 minutes because fewer GPU nodes fit the bigger requests. Measure where load time goes before raising requests; on shared clusters bigger requests cost scheduling time. Also NRP warns when limit/request ratio exceeds 1.2 for CPU or memory.
