Skip to main content

GPU Cluster Health Management

Project description

Trainy Konduktor Logo

Konduktor is a source available Run:AI alternative built on Kubernetes. Konduktor uses existing open source tools to build a platform that makes it easy for ML Researchers to submit batch jobs and for administrative/infra teams to easily manage GPU clusters.

How it works

Konduktor uses a combination of open source projects. Where tools exist with MIT, Apache, or another compatible open license, we want to use and even contribute to that tool. Where we see gaps in tooling, we build it and open source ourselves.

Architecture

Konduktor can be self-hosted and run on any certified Kubernetes distribution or managed by us. Contact as at founders@trainy.ai if you are just interested in the managed version. We're focused on tooling for clusters with NVIDIA cards for now but in the future we may expand to our scope to support other accelerators.

architecture

For ML researchers

  • Skypilot - user friendly batch job framework, where users only need to specify the resource requirements of their job and a script to launch that makes simple to scale work across multiple nodes. Works with most ML application frameworks. Examples
num_nodes: 100

resources:
  accelerators: H100:8
  cloud: kubernetes
  labels:
    kueue.x-k8s.io/queue-name: user-queue
    kueue.x-k8s.io/priority-class: low-priority

setup: |
  pip install torch torchvision

run: |
  torchrun \
  --nproc_per_node 8 \
  --rdzv_id=1 --rdzv_endpoint=$master_addr:1234 \
  --rdzv_backend=c10d --nnodes $num_nodes \
  torch_ddp_benchmark.py --distributed-backend nccl

For cluster administrators

  • DCGM Exporter, GPU operator, Network Operator - For installing NVIDIA driver, container runtime, and exporting node health metrics.
  • Kueue - centralized creation of job queues, gang-scheduling, and resource quotas and sharing across projects.
  • Prometheus - For publishing metrics about node health and workload queues.
  • OpenTelemetry - For pushing logs from each node
  • Grafana, Loki - Visualizations for metrics/logging solution.

Community & Support

Contributor Guide

Format your code with

poetry install --with dev
bash format.sh

Project details


Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

konduktor_nightly-0.1.0.dev20240805104313.tar.gz (12.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

File details

Details for the file konduktor_nightly-0.1.0.dev20240805104313.tar.gz.

File metadata

File hashes

Hashes for konduktor_nightly-0.1.0.dev20240805104313.tar.gz
Algorithm Hash digest
SHA256 49da9a54a307df051e181dbe1e17e6c1198393faf61cc716ca583efdcd3e3ddc
MD5 f1fa6083c68975b360702695805646d7
BLAKE2b-256 0f4e2fbff4f44e124525429e3a6ae6c06f33f5aaf7f9e60a9e28eb68f972a6e4

See more details on using hashes here.

File details

Details for the file konduktor_nightly-0.1.0.dev20240805104313-py3-none-any.whl.

File metadata

File hashes

Hashes for konduktor_nightly-0.1.0.dev20240805104313-py3-none-any.whl
Algorithm Hash digest
SHA256 db0a34b4a0c3387a0fbed25ec95c29bf66fb73601ef2ac2671d69b0d4dbc3d2a
MD5 b59e92a22142658886eb031ce1eb8b58
BLAKE2b-256 963dedd3ed52b99b4c3f5378f6c7654a9f813b7626ca8bd6ab1aa3131ecd028b

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page