mig-configure
Configurar las particiones NVIDIA MIG (GPU de múltiples instancias) en la DGX Station GB300, incluyendo habilitar el modo MIG, elegir un diseño de perfil, crear instancias y…
npx skills add https://github.com/nvidia/dgx-spark-playbooks --skill mig-configureMIG Configuration on DGX Station
Configure MIG (Multi-Instance GPU) partitions on the DGX Station GB300.
Steps
-
Find the GB300 GPU index. Run:
nvidia-smi --query-gpu=index,name --format=csv,noheader -
Check current MIG state:
nvidia-smi -i <GB300_INDEX> -q | grep -i "MIG Mode" -
If MIG is already enabled, show current instances:
nvidia-smi mig -lgi -i <GB300_INDEX> nvidia-smi mig -lci -i <GB300_INDEX>If the user wants to reconfigure, destroy existing instances first (step 6).
-
If MIG is not enabled, enable it. All GPU processes must be stopped first:
# Check for running GPU processes sudo fuser -v /dev/nvidia* # Enable MIG sudo nvidia-smi -i <GB300_INDEX> -mig 1 # Verify nvidia-smi -i <GB300_INDEX> -q | grep -i "MIG Mode" -
Show available profiles and help the user choose a layout:
nvidia-smi mig -lgip -i <GB300_INDEX>Common GB300 MIG profiles:
ID Profile name (driver-dependent) Approx. memory Use case 19 1g.35gb(59x) ·1g.31gb(61x)~30 GB Small models (7-8B), dev/test 20 1g.35gb+me·1g.31gb+me~30 GB Same + media extensions 15 1g.70gb~68 GB Slightly larger inference 14 2g.70gb~68 GB Medium models (14-30B) 9 3g.139gb(59x) ·3g.126gb(61x)~137 GB Large models (70B quantized) 5 4g.139gb·4g.126gb~137 GB Large models, more compute 0 7g.278gb(59x) ·7g.251gb(61x)~276 GB Full GPU as single instance Profile names depend on your driver version; the profile IDs do not. Always read the exact names and sizes on your box with
nvidia-smi mig -lgip -i <GB300_INDEX>, and create instances by ID. (Driver 59x reports the…35gb/139gb/278gbnames; 61x reports…31gb/126gb/251gbfor the same IDs.)Suggest layouts based on the user's workload (use the stable IDs). Examples:
- Two models (70B + smaller): one
3g+ two1g.70gb→ IDs9,15,15 - Many small models: three
1g→ IDs19,19,19 - One large model with isolation: the full
7g→ ID0
MIG layouts are constrained by fixed memory-slice placement, not just total memory — never sum nominal GB and assume any combination fits. A
3g + 2g + 2glayout (9,14,14) is not realizable, for example, because the second2ghas no legal placement after a3g. Andnvidia-smi mig -lgipFree/Totaltracks compute (GPC) slices, so it overstates the number of instances you can actually create (QA observed only 3 creatable1ginstances on a 61x Station even thoughFree/Totalreported 7). Always validate a specific layout withnvidia-smi mig -lgippbefore relying on it.Ask the user what models they want to run before suggesting a layout.
- Two models (70B + smaller): one
-
Create (or recreate) instances:
If reconfiguring, destroy existing instances first:
sudo nvidia-smi mig -dci -i <GB300_INDEX> sudo nvidia-smi mig -dgi -i <GB300_INDEX>Then create the new layout:
sudo nvidia-smi mig -cgi <PROFILE_IDS> -C -i <GB300_INDEX> -
Get the MIG device UUIDs:
nvidia-smi -LNote the
MIG-<uuid>entries — these are used to target specific MIG instances. -
Show the user how to use MIG devices:
# Bare metal export CUDA_VISIBLE_DEVICES=MIG-<uuid> # Docker docker run --gpus '"device=MIG-<uuid>"' ... -
Report the final layout to the user with UUIDs and suggested docker commands for each instance.
Disabling MIG
If the user wants to return to full-GPU mode:
# Stop all workloads using MIG instances first
sudo nvidia-smi mig -dci -i <GB300_INDEX>
sudo nvidia-smi mig -dgi -i <GB300_INDEX>
sudo nvidia-smi -i <GB300_INDEX> -mig 0
Do not run
nvidia-fabricmanageron DGX Station. It has a single GB300 over NVLink-C2C (no NVSwitch fabric), so Fabric Manager is not installed andsystemctl start nvidia-fabricmanagerfails with "Unit not found." NVLink-C2C re-initializes automatically after MIG is disabled. If MIG mode is stuck in a "pending" state, reset the GPU instead:sudo nvidia-smi -i <GB300_INDEX> --gpu-reset.