Skip to main content
Version: v0.0.0

Supported devices

Accelerator labels​

We believe that the types of accelerators used to build AI data centers will continue to diversify, and the MoAI Inference Framework provides features to efficiently utilize clusters with heterogeneous accelerators.

Each accelerator is identified by labels prefixed with moai.moreh.io. Users can either specify the accelerator on which a model should run, or allow the framework to automatically select the optimal accelerator. Below is the list of accelerator labels supported by the MoAI Inference Framework.

LabelAccelerator
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi355x
AMD Instinct MI355X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi350x
AMD Instinct MI350X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi325x
AMD Instinct MI325X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi308x
AMD Instinct MI308X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi300x
AMD Instinct MI300X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi300x-hf
AMD Instinct MI300X HF GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi300a
AMD Instinct MI300A APU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi250x
Half of AMD Instinct MI250X GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi250
Half of AMD Instinct MI250 GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi210
AMD Instinct MI210 GPU
moai.moreh.io/accelerator.vendor: amd
moai.moreh.io/accelerator.model: mi100
AMD Instinct MI100 GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: h100
NVIDIA H100 SXM GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: h100-pcie
NVIDIA H100 PCIe GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: h100-80gb-hbm3
NVIDIA H100 80GB HBM3 GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: h800
NVIDIA H800 GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: h800-pcie
NVIDIA H800 PCIe GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: a100-80gb-pcie
NVIDIA A100 80GB PCIe GPU
moai.moreh.io/accelerator.vendor: nvidia
moai.moreh.io/accelerator.model: a100-40gb-pcie
NVIDIA A100 40GB PCIe GPU
moai.moreh.io/accelerator.vendor: tenstorrent
moai.moreh.io/accelerator.model: blackhole
Tenstorrent Blackhole NPU
moai.moreh.io/accelerator.vendor: tenstorrent
moai.moreh.io/accelerator.model: wormhole
Tenstorrent Wormhole NPU

Checking available accelerators​

You can use the following command to see which accelerators are currently available in the cluster and how they are labeled by the MoAI Inference Framework.

kubectl get nodes -L moai.moreh.io/accelerator.vendor,moai.moreh.io/accelerator.model
NAME         STATUS     ROLES           AGE    VERSION   ACCELERATOR.VENDOR   ACCELERATOR.MODEL
cpu-001 Ready control-plane 1d v1.32.5
mi250-001 Ready <none> 1d v1.32.5 amd mi250
mi300x-001 Ready <none> 1d v1.32.5 amd mi300x

To list nodes that match a specific label, use the -l option:

kubectl get nodes -l moai.moreh.io/accelerator.model=mi300x
NAME         STATUS   ROLES    AGE   VERSION
mi300x-001 Ready <none> 1d v1.32.5