Log in and start training
GreenAI Stack is the tested, versioned image that boots on every instance and cluster: drivers, CUDA, frameworks and inference engines that work together. Above it, the orchestrators, CLI and API your platform team expects.
# Install the CLI and log in
$ pip install greenai
$ greenai auth login
# Launch an 8x B200 instance
$ greenai instances launch \
--type gpu_8x_b200_sxm6 \
--region us-west-2 \
--ssh-key laptop \
--filesystem datasets
# Or reserve a cluster
$ greenai clusters create \
--gpu h200 --count 64 \
--term 3m --scheduler slurmWhat's on the image
Released monthly, tested as a whole on every GPU we sell. Pin a version for reproducibility or track latest. Bring your own image if you prefer.
| Layer | Included |
|---|---|
| Operating system | Ubuntu 24.04 LTS, hardened kernel, unattended security updates |
| GPU drivers | NVIDIA driver 580 series, Fabric Manager, DCGM, GPUDirect RDMA and Storage |
| CUDA platform | CUDA 13, cuDNN 9, NCCL 2.27 with topology-tuned configs, cuBLAS, TensorRT |
| Frameworks | PyTorch 2.9, JAX 0.7, TensorFlow 2.20, NVIDIA NeMo, Megatron-LM, DeepSpeed |
| Libraries | Hugging Face transformers, datasets, accelerate, PEFT, TRL; xFormers; FlashAttention 3 |
| Inference | vLLM, SGLang, TensorRT-LLM, Triton Inference Server |
| Containers | Docker and Podman with the NVIDIA container toolkit, NGC catalog access, private registry |
| Tooling | Jupyter, Weights & Biases and MLflow clients, nvtop, py-spy, Nsight Systems |
Current release: GreenAI Stack 26.08. Versions shown are indicative and updated monthly; exact pins are listed in the release notes shipped with each image.
Your scheduler, our operations
Managed Slurm
The scheduler HPC teams know, with Pyxis and Enroot for containers, fair-share by project, and topology-aware placement on the InfiniBand fabric.
Managed Kubernetes
Upstream Kubernetes with the NVIDIA GPU Operator and Network Operator, Kueue or Volcano for batch, and RDMA-enabled pods out of the box.
Bare metal
Just SSH and a network. Run Ray, Run:ai, SkyPilot, dstack or whatever your team has standardized on. We keep the hardware healthy.
Everything the dashboard does, the API does
A REST API with an OpenAPI spec, a Python SDK, the CLI and a Terraform provider. Instances, clusters, filesystems, SSH keys, billing and carbon reports.
- Scoped API keys per project with rate limits and audit logs
- Webhooks for instance lifecycle, job completion and spend thresholds
- Terraform provider for reproducible environments and CI
- Carbon endpoint for per-job energy and emissions data
# Launch with the REST API
$ curl -X POST https://api.greenaibox.com/v1/instances \
-H "Authorization: Bearer $GREENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"type": "gpu_8x_h200_sxm",
"region": "ca-east-1",
"name": "sft-run-12",
"ssh_keys": ["laptop"],
"filesystems": ["datasets"]
}'
# Terraform
resource "greenai_instance" "train" {
type = "gpu_8x_h200_sxm"
region = "ca-east-1"
image = "greenai-stack:26.08"
}See what your GPUs are doing
Per-GPU telemetry
Utilization, memory, SM clocks, temperature, power and ECC errors at one-second resolution.
Fabric health
Link errors, congestion and NCCL collective timings across the InfiniBand partition.
Prometheus and Grafana
Scrape endpoints and pre-built dashboards, or ship metrics to Datadog, New Relic or your own stack.
Energy and carbon
Metered power and hourly grid mix per job, in the dashboard and the API.
Try the stack
Launch an instance and see how much setup you don't have to do.