vLLM Recipes

Loading models…

Launcher

Checkpoint

Hardware

Strategy

Launcher settings

SLURM_ACCOUNT (--set; or edit the machine profile)

SLURM_PARTITION

MODEL_HOST_PATH — model weights dir on the docker host (bind-mounted at the same path)

K8S_NAMESPACE — namespace the task pods run in

MODEL_PVC_CLAIM — optional: existing checkpoint PVC (mounted read-only at /models)

Use local checkpoint (skip HF download)

Redirect caches to shared storage (CACHE_DIR)

IFACE_NAME — cluster network interface for the multi-node rendezvous

Run mode

sflow run

workflow.yaml