repository·main·Indexed 23 days ago
https://github.com/kitops-ml/kitopsKitOps is a CNCF open-source tool for packaging, versioning, and securely sharing AI/ML projects using OCI-compliant ModelKits. It enables teams to treat AI assets like container images for standardized deployment, security scanning, and lifecycle management. The project includes the Kit CLI for managing ModelKits, the PyKitOps Python SDK for programmatic integration, and support for cryptographic signatures via Cosign and AI Bill of Materials (AI SBOM) for compliance.
Because KitOps follows OCI standards, it is compatible with a wide range of tools across the ML lifecycle.
KitOps works with major platforms for model tracking, experimentation, and orchestration, including:
Any platform that accepts containers can serve KitOps ModelKits, such as:
KitOps integrates with standard CI/CD and storage tools:
A ModelKit is a self-contained, immutable, and versioned bundle used to package AI/ML projects. It is built on OCI (Open Container Initiative) technology, allowing it to be stored and retrieved from any standard container registry.
A ModelKit can include:
Key Properties:
A Kitfile is a YAML-based manifest that serves as the blueprint for an AI project packaged as a ModelKit in KitOps. It defines the components of your project—such as models, datasets, code, prompts, and documentation—allowing you to package, version, and deploy them via any OCI-compatible registry.
Key benefits include:
ModelKit is a standardized, OCI-compliant packaging format designed for AI projects. It bundles various AI project artifacts—such as models, datasets, code, prompts, agent skill files, MCP server configurations, and documentation—into a single, versioned, and shareable artifact.
Key characteristics include:
A ModelKit is a comprehensive bundle of AI/ML artifacts (models, datasets, code, and parameters). It is composed of three primary elements:
application/vnd.kitops.modelkit.model.v1.tarapplication/vnd.kitops.modelkit.modelpart.v1.tarapplication/vnd.kitops.modelkit.dataset.v1.tarapplication/vnd.kitops.modelkit.code.v1.tarapplication/vnd.kitops.modelkit.docs.v1.tarapplication/vnd.kitops.modelkit.config.v1+json) that details the properties, relationships, and intended uses of the included artifacts.The Kit init container supports Sigstore [Cosign] verification to ensure ModelKit integrity. You can use either local keys or keyless OIDC verification.
initContainer.COSIGN_KEY environment variable to the path where the key is mounted.Set the following environment variables in the initContainer:
COSIGN_CERT_IDENTITY: The expected identity.COSIGN_CERT_OIDC_ISSUER: The OIDC issuer URL.ModelKits are managed using a hierarchical naming system:
sha256:a9561eb1b190625c9adb5a9513e72c4dedafc1cb2d4c5236c9a6957ec7dfd5a9).ModelKitID.[a-zA-Z0-9_.-]. or -; maximum length of 128 characters.repository:tag (e.g., in myllm:3.1.4, myllm is the repository).:8080).., one or two underscores __, or one or more dashes -). Components cannot start or end with a separator.Every KitOps function returns a CancellablePromise. You can call .cancel() to terminate the underlying process. This is useful for:
push or pull.SIGINT for graceful interruption.When cancelled, the promise rejects with an AbortError (specifically a DOMException with name: 'AbortError').
import { push } from '@kitops/kitops-ts';
async function pushWithTimeout(ref, timeoutMs) {
const op = push(ref);
const timer = setTimeout(() => op.cancel(), timeoutMs);
try {
await op;
} catch (err) {
if (err instanceof DOMException && err.name === 'AbortError') {
console.error(`Push timed out after ${timeoutMs}ms`);
process.exit(1);
}
throw err;
} finally {
clearTimeout(timer);
}
}
await pushWithTimeout('registry.example.com/org/my-model:v1.0.0', 5 * 60 * 1000);import { pull } from '@kitops/kitops-ts';
const op = pull('registry.example.com/org/my-model:v1.0.0');
process.once('SIGINT', () => {
console.log('\nInterrupted — cancelling pull…');
op.cancel();
});
try {
await op;
console.log('Pull complete.');
} catch (err) {
if (err instanceof DOMException && err.name === 'AbortError') {
console.log('Pull cancelled.');
process.exit(130); // standard exit code for SIGINT
}
throw err;
}When using the inspect() method, KitOps returns an OCI image manifest. The Manifest object contains the schemaVersion, mediaType, and an artifactType. The config field holds the Kitfile descriptor, while the layers array contains descriptors for each content layer. Metadata can be attached via annotations.
interface Manifest {
schemaVersion: number;
mediaType: string;
artifactType?: string;
config: ManifestDescriptor;
layers: ManifestDescriptor[];
annotations: ManifestAnnotations;
}A FilterFlag is a string expression used by unpack and info commands to target specific content within a ModelKit.
Accepted Formats:
'model': Selects the entire model layer.'datasets:training': Selects a named dataset (e.g., datasets:training).'docs:./README.md': Selects a specific file within a layer (e.g., docs:./README.md).'model,datasets:validation': Selects multiple layers using comma-separated values.type FilterFlag = string