Crate model_runtime 

model-runtime

Generic model infrastructure for the workspace.

This crate owns model identity, sources, bundle materialization, downloads, model-specific artifact metadata, and job helpers. Generic artifact storage and validation live in moritzbrantner-jobs-core; model bundles and Hugging Face download logic stay here. Domain crates should hide this layer behind operations such as object detection, transcription, OCR, segmentation, embeddings, and classification.

Runtime Surface

• model.executionPlan validates a ModelAccessJobRequest and returns a pure job/access plan with runtime execution metadata.
• model.bundlePlan plans bundle manifest layout and artifact references without downloading or materializing files.
• model.jobManifest projects a planned model access job into a deterministic JobManifest.
• model.presets lists preset ids and model spec summaries.
• model.spec validates a model spec and returns safe names, task, source, requested files, and revision.

Default surface operations are deterministic and planning-only. They do not download models, spawn background jobs, run native inference, execute external commands, write files, or access the network.

Local ONNX Defaults

moritzbrantner-model-runtime owns materialization for the server-only local model workflows used by text QA, image classification, and image captioning. The current ONNX-first presets are:

• roberta-base-squad2-onnx: onnx-community/roberta-base-squad2-ONNX.
• vit-base-patch16-224-onnx: Xenova/vit-base-patch16-224.
• vit-gpt2-image-captioning-onnx: Xenova/vit-gpt2-image-captioning.

Use resolve_or_download_bundle when a workflow should load an existing bundle first and auto-download missing files into .model-runtime on native server/CLI runtimes. WASM callers should keep using validation/import paths.

Re-exports

pub use jobs::plan_model_access;

pub use jobs::plan_model_bundle;

pub use jobs::ModelAccessJobRequest;

pub use jobs::ModelAccessPlan;

pub use jobs::ModelBundlePlan;

pub use jobs::ModelBundlePlanFile;

pub use jobs::ModelJobInput;

pub use jobs::ModelJobKind;

Modules

jobs

surface

Library-owned runtime surface for model-runtime.

Structs

BlueGreenPredictionTestOptions

Data type for blue/green prediction test options.

BlueGreenPredictionTestReport

Data type for blue/green prediction test report.

CudaOxideModelPlan

Data type for cuda-oxide model plan.

CudaOxideRuntimeConfig

Data type for cuda-oxide runtime config.

DownloadedModel

Data type for downloaded model.

HuggingFaceDownloader

Data type for hugging face downloader.

ModelBundle

Data type for model bundle.

ModelBundleFile

Data type for model bundle file.

ModelBundleManifest

Data type for model bundle manifest.

ModelBundleResolveOptions

Options for resolving a local bundle, downloading it when allowed.

ModelBundleStore

Data type for model bundle store.

ModelRuntimeSelection

Advanced generic runtime selection.

ModelSpec

Generic model spec.

RawPrediction

Backend-neutral raw model prediction used for runtime conformance checks.

Enums

FallbackPolicy

General fallback behavior for model-backed operations.

ModelFileRequest

Variants describing model file request.

ModelPreset

Variants describing model preset.

ModelRuntimeBackend

Variants describing model runtime backend.

ModelRuntimeError

Error type used by generic model runtime infrastructure.

ModelSource

Model source.

ModelTask

Variants describing model task.

RuntimePreference

General runtime preference for domain APIs.

Constants

CUDA_OXIDE_BOOK_URL

Documentation URL for the cuda-oxide compiler/runtime stack.

Traits

ModelDownloader

Minimal downloader seam for bundle resolution tests and alternate materializers.

Functions

compare_blue_green_predictions

Compares green and blue model predictions for runtime conformance tests.

resolve_or_download_bundle

Resolves a bundle from disk, optionally downloading and materializing it first.

resolve_or_download_bundle_with_downloader

Resolves a bundle with a caller-provided downloader seam.

Type Aliases

HuggingFaceModelSpec

Hugging Face-oriented compatibility alias for callers migrating to ModelSpec.

Result

Result type used by generic model runtime infrastructure.