Platform · Models

KERNSTACK Training

Training on KERNSTACK is stated plainly because the platform has no separate training resource to sell: model training and fine-tuning jobs run as workers — long compute processes in the project, with bucket volumes mounting the dataset and the resulting model written back to a bucket, ready for a custom inference endpoint to serve. The value is colocation: the job runs beside the tables and buckets it trains on, inside the same declared project.

Training (Workers)

Talk to us Read Training Docs → Software Lifecycle Cloud →

Why teams replace assembled cloud services with KERNSTACK Training

Teams otherwise run a GPU account disconnected from the app: data exported to a training platform, models exported back, lineage reconstructed from notebook names. On KERNSTACK the dataset, the job, the artefact, and the serving endpoint are one project’s resources, reviewed in one plan.

How a training job is declared

A worker carries the training code (python is the common runtime), declares its sizing, and mounts volumes: a dataset volume over the bucket prefix holding training data, an output path for checkpoints and the final model. A workflow can chain the honest pipeline — prepare data, train, evaluate, and a human step before the new model version is promoted to the endpoint that serves it.

The limits, without varnish

This is product-scale training: fine-tunes, classifiers, embeddings adapters, periodic retraining over business data. Frontier-scale pretraining across accelerator fleets is a specialised business this page will not pretend to replace. Worker sizing (up to 1024 MiB memory and 8 cores in the standard declaration) tells you the shape of job that belongs here; the meter is the ordinary compute rate, so a training run’s cost is knowable to the GiB-second.

Datasets stay in the zone: training reads the same buckets the product writes
Model artefacts land in a bucket with versions in the key path; endpoints point at a prefix
Promotion to serving can pause at a human step — no model reaches users unreviewed

SDKs & Language Integration

Every KERNSTACK environment vendors typed SDK packages into .kernstack/ so both engineers and coding agents in Agent Sandboxes can call Training without external credentials:

Go SDK: github.com/coding-berlin/kernstack-sdk-go/sdk_workers
Node.js SDK: kernstack-sdk
Python SDK: kernstack_sdk
PHP SDK: Kernstack\Workers

Training in the Project Graph

Declarestack.jsontraining

RunTrainingintelligence

MeterOne credit balanceEU zone

stack.json — training configuration plan · deploy

{
  "resources": {
    "buckets": [{ "name": "datasets", "volumes": [{ "name": "data", "source_prefix": "training/" }] }, { "name": "models" }],
    "workers": [{ "name": "train", "runtime": "python", "memory_mb": 1024, "cpu_cores": 8,
      "volumes": [{ "name": "data" }] }],
    "inference_endpoints": [{ "name": "ranker", "runtime": "python",
      "model_bucket": "models", "model_prefix": "ranker-v7/" }]
  }
}

Metering · compute, storage, or network 1 Shared Balance

Compute is 0.01 credits per GiB-second on the shared environment balance.

Replaces: SageMaker Training · no native equivalent · no direct equivalent

FAQ

Frequently asked questions

Is there a dedicated training/GPU product?

No, and this page exists partly to say so. Training runs as workers with dataset volumes. If your workload needs multi-node accelerator fleets, use a specialist platform and keep the data export deliberate.

How does a trained model get served?

The job writes the artefact to a bucket (versioned by prefix), and a custom inference endpoint declares that bucket and prefix as its model source. Training and serving are two declarations in one project.

What does a training run cost?

Its compute at 0.01 credits per GiB-second plus dataset storage at the storage rate. Long runs are therefore budgetable in advance from expected duration and sizing, on the same balance as the product.

Connected in the Project Graph

Related capabilities, architectures, and comparisons

Ship your stack on the Software Lifecycle Cloud.

Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.

Talk to us See pricing & credits →