backend.ai-accelerator-habana 26.4.4rc3

Backend.AI Accelerator Plugin for Intel® Gaudi® HPUs

Backend.AI Accelerator Plugin for Habana

Just install this along with Backend.AI agents, using the same virtual environment. This will allow the agents to detect Habana devices on their hosts and make them available to Backend.AI kernel sessions.

$ pip install -e <Path to this repository>

This open-source edition of Habana plugins support allocation of one or more Habana devices to a container, slot-by-slot.

Compatibility Matrix

TBD

Generating Docker macvlan interfaces for inter-container network

#! /bin/bash
set -e
for pci_bus_id in $(hl-smi --format=csv,noheader -Q bus_id); do
    for net_dev in $(grep PCI_SLOT_NAME /sys/class/net/*/device/uevent | grep "${pci_bus_id}" | awk -F: '{print $1}' | awk -F/ '{print $5}'); do
        docker network create --attachable -o parent="${net_dev}" -d macvlan "gaudi_${net_dev}" >/dev/null
        echo "Docker network gaudi_${net_dev} created"
    done
done