RunIntegrations Early access

Bring the scheduler you already use.

Every Fantasti cluster is yours, with its own Kubernetes API. Run Ray, SkyPilot or NVIDIA OSMO on it the way their docs describe, track runs in managed MLflow, and add applications from a catalog.

Request access Talk to the solutions desk

infra: k8s/fantasti-acme

Being built for the first cohort. NVIDIA OSMO and Applications are planned.

TAB 01Interfaces to one cluster Source · Fantasti

Metaflow Runs as Fantasti Flows. Planned

CLI

Preview API · subject to change

export KUBECONFIG="$HOME/.kube/fantasti-acme.yaml"

kubectl get nodes -L nvidia.com/gpu.product,nvidia.com/gpu.count

Output · illustrative

NAME          STATUS  ROLES   GPU.PRODUCT  GPU.COUNT
gpu-node-01   Ready   <none>  NVIDIA-H200  8
gpu-node-02   Ready   <none>  NVIDIA-H200  8

§01 · On this page Kubernetes

CLI

Preview API · subject to change

helm repo add kuberay https://ray-project.github.io/kuberay-helm/
helm install kuberay-operator kuberay/kuberay-operator

# ray-values.yaml: a GPU worker group, min 0
helm install raycluster kuberay/ray-cluster -f ray-values.yaml

§02 · On this page Ray

CLI

Preview API · subject to change

sky check kubernetes                      # finds your Fantasti context
sky gpus list --infra k8s/fantasti-acme   # GPUs in your cluster
sky jobs launch -n sft-01 sft.yaml     # recovers after node loss

§03 · On this page SkyPilot

CLI

Preview API · subject to change

osmo pool list
osmo workflow submit train.yaml --pool fantasti-acme

§04 · On this page NVIDIA OSMO

Python

Preview API · subject to change

# MLFLOW_TRACKING_URI and credentials are already set
import mlflow

mlflow.set_experiment("sft-01")
with mlflow.start_run():
    mlflow.log_params({"lr": 2e-5, "epochs": 3})
    mlflow.log_metric("eval_loss", eval_loss)
    mlflow.log_artifact("checkpoints/last/config.json")

§05 · On this page MLflow

Python

Preview API · subject to change

app = fx.apps.install("jupyterlab", cluster="acme-train", gpu="L40S")
print(app.url)

§06 · On this page Applications

Control plane: your own, one per cluster
GPU nodes: dedicated · NVIDIA GPU labels
Fabric: one InfiniBand fabric per cluster
Capacity: on-demand · spot pools with max_price

§01 Kubernetes Early access

Your cluster, your API server.

Each account gets an isolated cluster with its own control plane and dedicated GPU nodes. You receive a kubeconfig. Helm, operators and GitOps tools work unchanged, and kubectl get pods -A shows your workloads and nothing else.

GPU labelsnodes carry the standard NVIDIA GPU product labels
One fabricnodes in a cluster share one InfiniBand fabric
Spot poolsadd spot node pools with a max price

GPU Clusters MCP server

Preview API · subject to change

import fantasti

fx = fantasti.Client()                # reads FANTASTI_API_KEY

cluster = fx.clusters.create(
    name="sft-01",
    gpu="H200:8",                     # 8-GPU nodes
    nodes=2,                          # on one InfiniBand fabric
    capacity=fantasti.Spot(max_price="follow"),
    interfaces=["kubernetes", "skypilot"],
)
cluster.wait("ready")
cluster.kubeconfig.save("~/.kube/fantasti-acme.yaml")

# Console: Clusters → acme → Connect → download kubeconfig
export KUBECONFIG="$HOME/.kube/fantasti-acme.yaml"

# your workloads and your cluster's system pods, nothing else
kubectl get pods -A

kubectl get nodes -L nvidia.com/gpu.product,nvidia.com/gpu.count

Output · illustrative

NAME          STATUS  ROLES   GPU.PRODUCT  GPU.COUNT
gpu-node-01   Ready   <none>  NVIDIA-H200  8
gpu-node-02   Ready   <none>  NVIDIA-H200  8

{
  "mcpServers": {
    "fantasti": { "type": "http", "url": "https://mcp.fantasti.ai/mcp" }
  }
}

Clusters tool group · planned

clusters_list
clusters_get
clusters_create    asks for confirmation
clusters_delete    asks for confirmation

curl -X POST https://api.fantasti.ai/v1/clusters \
  -H "Authorization: Bearer $FANTASTI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "name": "acme-train",
        "gpu": "H200:8",
        "nodes": 2,
        "capacity": { "mode": "spot", "max_price": "follow" },
        "interfaces": ["kubernetes", "skypilot"]
      }'

Response · illustrative

{ "id": "cl_8f2k…", "status": "placing", "fabric": "single" }

Planned Cluster tools for MCP clients such as coding agents, through the MCP server.

§02 Ray Early access

Ray on your GPUs.

Run Ray clusters with KubeRay on your Fantasti cluster, with GPU worker groups that scale from zero. Workspaces use the same Ray clusters, so code moves from notebook to job without changes.

Preview API · subject to change

# Plain Ray, nothing Fantasti-specific
import ray
ray.init()  # connects to the Ray cluster on your Fantasti cluster

ds = ray.data.read_parquet("s3://acme-data/prompts/")
ds = ds.map_batches(  # GPU workers scale up from 0
    Embedder, concurrency=8, num_gpus=1, batch_size=256,
)
ds.write_parquet("s3://acme-data/embeddings/")

helm repo add kuberay https://ray-project.github.io/kuberay-helm/
helm install kuberay-operator kuberay/kuberay-operator

# ray-values.yaml: a GPU worker group, min 0
helm install raycluster kuberay/ray-cluster -f ray-values.yaml

Workspaces

Workers from zero. Illustrative.

Plain Ray code calls ray.init() and maps Embedder over a dataset with 8 workers of 1 GPU each. It connects to the Ray head that KubeRay runs on your Fantasti cluster. The GPU worker group starts at zero workers, scales to 8 while the job runs and returns to zero. Connections: Your code to Ray head; Ray head to GPU worker group.

DriverYour codePlain Ray
ray.init() # plain Rayds.map_batches(Embedder, concurrency=8, num_gpus=1)
Ray headKubeRay
kind
RayCluster
runs on
your cluster
workers
min 0 · max 8
GPU worker groupFrom zero
1 GPU each

DriverYour codePlain Ray
ray.init() # plain Rayds.map_batches(Embedder, concurrency=8, num_gpus=1)
Ray headKubeRay
kind
RayCluster
runs on
your cluster
workers
min 0 · max 8
GPU worker groupFrom zero
1 GPU each

DriverYour codePlain Ray
ray.init() # plain Rayds.map_batches(Embedder, concurrency=8, num_gpus=1)
Ray headKubeRay
kind
RayCluster
runs on
your cluster
workers
min 0 · max 8
GPU worker groupFrom zero
1 GPU each

Scales up
Your job

API One per cluster Reviewed 2026-10-09

1 Kubernetes API per cluster, and it is yours.

One isolated cluster per account

EQ 01 Source · Fantasti platform configuration

HGX H200 8-GPU baseboard, line drawing — One cluster, every interface. Illustrative · example names.

§03 SkyPilot Early access

Bring your SkyPilot YAML.

Fantasti issues a Kubernetes context for your cluster. Point SkyPilot at it and sky launch, sky jobs launch and sky serve work as they do anywhere else, with failover across the GPU types you list.

infra: k8s/fantasti-acme

SkyPilot’s own “workspaces” are team permissions. Fantasti Workspaces are GPU development environments; they are different things.

YAML

Preview API · subject to change

# sft.yaml: "fantasti-acme" is the kube context for your isolated cluster
name: sft-01
num_nodes: 2

resources:
  infra: k8s/fantasti-acme
  accelerators: {H200:8, B200:8}   # either works; SkyPilot takes what is free
  job_recovery:
    strategy: FAILOVER
    max_restarts_on_errors: 2

workdir: .
setup: pip install -r requirements.txt
run: |
  MASTER=$(echo "$SKYPILOT_NODE_IPS" | head -n1)
  torchrun --nnodes=$SKYPILOT_NUM_NODES --node_rank=$SKYPILOT_NODE_RANK \
    --nproc_per_node=$SKYPILOT_NUM_GPUS_PER_NODE \
    --master_addr=$MASTER --master_port=29500 train.py --resume

CLI

Preview API · subject to change

sky check kubernetes                      # finds your Fantasti context
sky gpus list --infra k8s/fantasti-acme   # GPUs in your cluster
sky jobs launch -n sft-01 sft.yaml     # recovers after node loss

§04 NVIDIA OSMO

NVIDIA OSMO for physical-AI workflows. Planned

Use OSMO on your Fantasti cluster to run simulation, synthetic-data and training workflows as one pipeline. Fantasti installs it and exposes your GPUs as OSMO pools.

YAML

Preview API · subject to change

# train.yaml: an OSMO workflow on your Fantasti cluster's GPU pool
workflow:
  name: policy-train
  resources:
    gpu_task:
      cpu: 16
      gpu: 1
      memory: 64Gi
      storage: 100Gi
  tasks:
  - name: train
    image: nvcr.io/nvidia/pytorch:24.01-py3
    command: ["bash", "-c", "python train.py --epochs 3"]
    resource: gpu_task

CLI

Preview API · subject to change

osmo pool list
osmo workflow submit train.yaml --pool fantasti-acme

One pipeline, one pool. Schematic · not to scale.

Three workflows, simulation, synthetic data and training, run in order as one pipeline. NVIDIA OSMO, installed by Fantasti on your cluster, submits their tasks to one pool named fantasti-acme, which is the GPU nodes of your cluster. The task train of the workflow policy-train holds 1 of the 8 GPUs of the first node. Connections: Simulation to Synthetic data; Synthetic data to Training; Simulation to NVIDIA OSMO; Synthetic data to NVIDIA OSMO; Training to NVIDIA OSMO; NVIDIA OSMO to gpu-node-01: submits tasks.

OSMO poolfantasti-acmeYour GPU nodes

01Simulation
02Synthetic data
03Training
NVIDIA OSMOPlanned
Installed by Fantasti on your cluster.
workflow
policy-train
pool
fantasti-acme
gpu-node-018 GPUs
task train
gpu-node-028 GPUs
free

OSMO poolfantasti-acme

01Simulation
02Synthetic data
03Training
NVIDIA OSMOPlanned
Installed by Fantasti on your cluster.
workflow
policy-train
pool
fantasti-acme
gpu-node-018 GPUs
task train
gpu-node-028 GPUs
free

One pipeline
Planned

§05 MLflow Early access

Managed MLflow.

Managed MLflow tracking and model registry, wired into Workspaces, Flows, Batch Inference, Serverless jobs and Clusters.

One run, on the record. Illustrative.

A workload on any of your nodes, in Workspaces, jobs or batch runs, already has MLFLOW_TRACKING_URI and credentials set. It logs the parameter lr and the metric eval_loss to the experiment sft-01, which holds 3 runs. Run 03 is registered as version 3 of the model sft-01. Connections: Any workload to sft-01: logs; sft-01 to sft-01: registers.

Any workloadOn your nodes
- Workspaces
- Jobs
- Batch runs
MLFLOW_TRACKING_URI
set
credentials
set
mlflow.log_params({"lr": 2e-5})mlflow.log_metric("eval_loss", …)
Experimentsft-013 runs
run lr eval_loss03 2e-5 0.41202 5e-5 0.44701 1e-4 0.503
artifacts
checkpoints/last
Modelsft-01Registry
v3 run 03v2 run 02v1 run 01

Any workloadOn your nodes
- Workspaces
- Jobs
- Batch runs
MLFLOW_TRACKING_URI
set
credentials
set
mlflow.log_params({"lr": 2e-5})mlflow.log_metric("eval_loss", …)
Experimentsft-013 runs
run lr eval_loss03 2e-5 0.41202 5e-5 0.44701 1e-4 0.503
artifacts
checkpoints/last
Modelsft-01Registry
v3 run 03v2 run 02v1 run 01

Any workloadOn your nodes
- Workspaces
- Jobs
- Batch runs
MLFLOW_TRACKING_URI
set
credentials
set
mlflow.log_params({"lr": 2e-5})mlflow.log_metric("eval_loss", …)
Experimentsft-013 runs
run lr eval_loss03 2e-5 0.41202 5e-5 0.44701 1e-4 0.503
artifacts
checkpoints/last
Modelsft-01Registry
v3 run 03v2 run 02v1 run 01

Registered
Logged by your code

Python

Preview API · subject to change

# MLFLOW_TRACKING_URI and credentials are already set
import mlflow

mlflow.set_experiment("sft-01")
with mlflow.start_run():
    mlflow.log_params({"lr": 2e-5, "epochs": 3})
    mlflow.log_metric("eval_loss", eval_loss)
    mlflow.log_artifact("checkpoints/last/config.json")

01Tracking

Experiments and runs

Log parameters, metrics and artifacts from any workload.
02Registry

Model versions

Register, version and promote models.
03Wired in

Set on every node

Workspaces, jobs and batch runs get MLFLOW_TRACKING_URI and credentials.

§06 Applications 5 categories

Applications, one click from your GPUs. Planned

TAB 02Applications by category Source · Fantasti

Applications by category: 5 categories, 11 applications, and how each category runs.
Category	Applications	Managed app	On your cluster	Runs as
Notebooks and IDEs	JupyterLab code-server	Yes	Yes	Managed app or on your cluster
Inference servers	vLLM SGLang NVIDIA Triton	No	Yes	On your cluster
Vector databases	Qdrant	Yes	Yes	Managed app or on your cluster
Interfaces	Open WebUI ComfyUI	Yes	No	Managed app
Schedulers	Ray Cluster SkyPilot API server NVIDIA OSMO	No	Yes	On your cluster

Python

Preview API · subject to change

app = fx.apps.install("jupyterlab", cluster="acme-train", gpu="L40S")
print(app.url)

§07 Q&A 5 questions

Sheet: 01
Title: Integrations · applications and answers
Reviewed: 2026-10-09

§08 Request Early access

Keep your scheduler. Change where it runs.

Integrations is in early access. Request access and tell us which scheduler you run.

Request access

Bring the scheduler you already use.

Your cluster, your API server.

Ray on your GPUs.

Bring your SkyPilot YAML.

NVIDIA OSMO for physical-AI workflows. Planned

Managed MLflow.

Experiments and runs

Model versions

Set on every node

Applications, one click from your GPUs. Planned

Questions and answers.

Do I need to change my Ray or SkyPilot code?

Is my cluster shared with other accounts?

Can my scheduler run on spot capacity?

Which integrations are in early access?

What about Metaflow?

Keep your scheduler. Change where it runs.