[Infra] Local PV on NVMe for model weight cold starts #40

Open
opened 2026-09-11 01:27:34 +00:00 by poimen · 0 comments
Member

Downloading model weights from remote creates minutes of pod init lag. ornith pulls 35B model on every restart (ollama pull). Use Local PV on NVMe SSD with pre-cached weights. OLLAMA_MODELS=/mnt/models already set but storage is longhorn (network). Files: k8s/apps/llm-serving/ornith.yaml, k8s/infra/longhorn/

Downloading model weights from remote creates minutes of pod init lag. ornith pulls 35B model on every restart (ollama pull). Use Local PV on NVMe SSD with pre-cached weights. OLLAMA_MODELS=/mnt/models already set but storage is longhorn (network). Files: k8s/apps/llm-serving/ornith.yaml, k8s/infra/longhorn/
poimen added this to the LLM Production Hardening milestone 2026-09-11 01:27:34 +00:00
Sign in to join this conversation.