tt-device-plugin

Kubernetes Device Plugin for Tenstorrent AI accelerators.

Exposes Tenstorrent NPU/AI accelerator cards to Kubernetes workloads via the standard Device Plugin API. Forked from squat/generic-device-plugin and pre-configured for Tenstorrent hardware.

Supported Hardware

Generation	Model	Chips per card	Device file
Grayskull	e75, e150	1	/dev/tenstorrent/N
Wormhole	n150	1	/dev/tenstorrent/N
Wormhole	n300, n300s	2 (L+R, Ethernet-linked)	/dev/tenstorrent/N
BlackHole	p100, p150	1	/dev/tenstorrent/N

All models expose devices via /dev/tenstorrent/N. Each device file represents one allocatable unit:

• Single-chip cards (e75, e150, n150, p100, p150): one device file = one chip.
• Dual-chip cards (n300, n300s): one device file = one card (both chips, L+R, accessible via the left chip's PCIe connection).

The plugin supports up to 8 devices per node (/dev/tenstorrent/0–7). Device files that do not exist on a node are automatically excluded from the advertised capacity.

Kubernetes Resource

tenstorrent.com/device

Quick Start

Deploy the DaemonSet to your cluster:

kubectl apply -f https://raw.githubusercontent.com/bricksum/tt-device-plugin/main/manifests/tenstorrent/daemonset.yaml

Verify that devices are registered on the node:

kubectl describe node <node-name> | grep tenstorrent

Expected output (example with 4 devices):

tenstorrent.com/device:  4

Usage

Request Tenstorrent devices in a Pod spec using the tenstorrent.com/device resource:

apiVersion: v1
kind: Pod
metadata:
  name: tt-workload
spec:
  containers:
  - name: workload
    image: <your-image>
    resources:
      limits:
        tenstorrent.com/device: "1"

The requested device file is mounted into the container at the same path (/dev/tenstorrent/N).

Examples

Single device (1 card):

resources:
  limits:
    tenstorrent.com/device: "1"

Multiple devices (e.g. 4 N300S cards for a mesh workload):

resources:
  limits:
    tenstorrent.com/device: "4"

Node selector for Tenstorrent nodes (recommended when the cluster has mixed hardware):

spec:
  nodeSelector:
    tenstorrent.com/device: "true"   # set this label on Tenstorrent nodes
  containers:
  - name: workload
    resources:
      limits:
        tenstorrent.com/device: "1"

How It Works

The plugin is a DaemonSet running in the kube-system namespace. It uses squat/generic-device-plugin to watch /dev/tenstorrent/0–7 and registers each existing device file as an allocatable unit under the tenstorrent.com domain via the Kubernetes Device Plugin gRPC API.

When a Pod requests tenstorrent.com/device: N, the scheduler selects a node with sufficient capacity, and the kubelet mounts the allocated device files into the container with the appropriate permissions.

Requirements

• Kubernetes 1.26+
• TT-KMD kernel driver installed on the node
• Device files present at /dev/tenstorrent/

Manifest

The DaemonSet manifest is at manifests/tenstorrent/daemonset.yaml.

The container image used is squat/generic-device-plugin (upstream, unmodified). Only the configuration — device paths and resource domain — is customized in this fork.

Upstream

This repository is a fork of squat/generic-device-plugin. For generic usage, advanced configuration options (--device, --config, USB devices, count, optional paths), and the full flag reference, see the upstream repository.

Maintained by

Bricksum Engineering

Bricksum is an AI infrastructure company specializing in Tenstorrent NPU deployments and Kubernetes-based AI accelerator orchestration.

License

Copyright 2026 Bricksum Engineering Copyright 2020 the generic-device-plugin authors

Licensed under the Apache License, Version 2.0.