Common Workflows

Use your deployed MSP management API for management operations and its inference Gateway for model requests. Check the deployed version and required permissions. Example addresses are placeholders.

These are Human integration examples for the complete SDK surface. The msp-operations Skill must check its current capability manifest before using any command or method. In the first Skill release, workload creation, upload, update, cancellation, restart, stop, and deletion remain disabled pending contract tests, sandbox end-to-end evidence, and Agent evaluation.

Upload and Deploy a Model

The upload helper validates files, calculates content metadata, transfers the directory, and completes the model asset before returning it. Deployment creation then uses the asset id.

Python

model = client.my_models.upload(
    "./model",
    name="example-model",
    model_type="LLM",
)

request = {
    "name": "example-deployment",
    "modelAssetId": model["id"],
    "backend": "sglang",
    "servingMode": "standard",
    "gpuType": "your-gpu-type",
    "replicas": 1,
    "clusterId": 1,
}

preview = client.deployments.preview(request)
if preview.get("creatable"):
    deployment = client.deployments.create(request)
    running = client.wait_for_deployment(deployment["id"])

TypeScript

const model = await client.myModels.upload<{ id: number }>("./model", {
  name: "example-model",
  modelType: "LLM",
});

const request = {
  name: "example-deployment",
  modelAssetId: model.id,
  backend: "sglang",
  servingMode: "standard",
  gpuType: "your-gpu-type",
  replicas: 1,
  clusterId: 1,
};

const preview = await client.deployments.preview(request);
if (preview.creatable) {
  const deployment = await client.deployments.create(request);
  const running = await client.waitForDeployment(deployment.id);
}

Choose the cluster and accelerator values returned by the discovery and preview methods. The placeholder values above are not deployment recommendations.

Upload a Dataset and Start Training

Build the request from the active Training capabilities, a selected cluster, the uploaded Dataset, and a deployable base model. Preview the exact resource specification before creating the job.

Python

dataset = client.datasets.upload(
    "./train.jsonl",
    name="example-dataset",
    dataset_type="training",
    training_category="sft-llm",
)

preview = client.training.preview({
    "clusterId": 1,
    "resourceSpecId": "resource-spec-id",
})

request = {
    "clientToken": "training-request-001",
    "displayName": "example-training",
    "outputModelName": "example-output",
    "recipeId": "recipe-id",
    "recipeVersion": "recipe-version",
    "baseModelRef": {"type": "recipe_model", "id": "model-id"},
    "datasetRefs": [{"datasetId": str(dataset["id"]), "role": "train"}],
    "placement": {"clusterId": "1", "resourceSpecId": "resource-spec-id"},
    "params": {},
}

if preview.get("decision") == "FIT":
    job = client.training.create(request)
    completed = client.wait_for_training_job(job["jobId"])

TypeScript

const dataset = await client.datasets.upload<{ id: number }>("./train.jsonl", {
  name: "example-dataset",
  datasetType: "training",
  trainingCategory: "sft-llm",
});

const preview = await client.training.preview({
  clusterId: 1,
  resourceSpecId: "resource-spec-id",
});

const request = {
  clientToken: "training-request-001",
  displayName: "example-training",
  outputModelName: "example-output",
  recipeId: "recipe-id",
  recipeVersion: "recipe-version",
  baseModelRef: { type: "recipe_model", id: "model-id" },
  datasetRefs: [{ datasetId: String(dataset.id), role: "train" }],
  placement: { clusterId: "1", resourceSpecId: "resource-spec-id" },
  params: {},
};

if (preview.decision === "FIT") {
  const job = await client.training.create<{ jobId: string }>(request);
  const completed = await client.waitForTrainingJob(job.jobId);
}

Replace the placeholder recipe, model, cluster, and resource identifiers with values returned by the current environment. Do not copy resource selections between environments. outputModelName is only the output suffix. Keep it within the model-specific limit returned by Training capabilities so the final <base>-FT-<suffix> name remains deployable.

Create an Evaluation

Python

request = {
    "kind": "benchmark",
    "modelType": "LLM",
    "models": [{
        "type": "external",
        "provider_key_id": "provider-key-id",
        "model_id": "model-id",
    }],
    "dataset": "evaluation-dataset",
    "maxSamples": 100,
    "clusterId": 1,
}

preview = client.evaluations.preview({"clusterId": request["clusterId"]})
if preview.get("decision") == "FIT":
    job = client.evaluations.create(request)
    client.wait_for_evaluation_job(job["jobId"])
    report = client.evaluations.report(job["jobId"])

TypeScript

const request = {
  kind: "benchmark",
  modelType: "LLM",
  models: [{
    type: "external",
    provider_key_id: "provider-key-id",
    model_id: "model-id",
  }],
  dataset: "evaluation-dataset",
  maxSamples: 100,
  clusterId: 1,
};

const preview = await client.evaluations.preview({ clusterId: request.clusterId });
if (preview.decision === "FIT") {
  const job = await client.evaluations.create<{ jobId: string }>(request);
  await client.waitForEvaluationJob(job.jobId);
  const report = await client.evaluations.report(job.jobId);
}

Use the report method for ordinary evaluations. Media-oriented evaluations may return a separate artifact reference for download.

Safety Rules

• Preview compute requirements before creating workloads.
• Treat create, update, stop, restart, cancel, and delete as state-changing operations.
• Use idempotency fields supplied by the API when retrying creation after an uncertain result.
• Fetch pre-signed object URLs without adding the platform bearer token.
• Verify final state with get, list, or a waiter instead of relying only on the initial response.