RuntimeConfig class

Runtime config for building a model — the per-call params getActiveModel collects, kept as a small holder so the provider contract stays stable as params evolve.

Constructors

RuntimeConfig({required int maxTokens, required String modelPath, String? tokenizerPath, PreferredBackend? preferredBackend, PreferredBackend? preferredVisionBackend, PreferredBackend? preferredAudioBackend, bool supportImage = false, bool supportAudio = false, int? maxNumImages, MultimodalOrder? multimodalOrder, bool? enableSpeculativeDecoding, ActivationDataType? activationDataType, int? maxConcurrentSessions, List<int>? loraRanks, Map<String, String>? artifactPaths, String? language})

const

Properties

activationDataType → ActivationDataType?

Activation type for the native LiteRT-LM engine's text decoder. Null leaves it to the model file (except a ModelType.functionGemma bundle that declares no prefer_activation_type, which gets float32 on its GPU attempt); when set, it overrides LiteRT-LM's own choice (float16 on GPU by default). Set ActivationDataType.float32 when a GPU writes wrong digits (LiteRT-LM#2814, #3012). The vision and audio encoders keep the model's own type; MediaPipe, ONNX, built-in AI and the web engines ignore it.

final

artifactPaths → Map<String, String>?

filename → on-disk path for a multi-file model bundle (TTS). The bundle backend resolves each graph/dictionary/embedding file by name from this map. Null for single-file models (inference/embedding/STT).

final

enableSpeculativeDecoding → bool?

final

hashCode → int

The hash code for this object.

no setter, inherited

language → String?

The output language, for the two model kinds that have one. Null everywhere else, and null means "the backend's own default".

final

loraRanks → List<int>?

LoRA ranks for the MediaPipe path; null falls back to the platform's supportedLoraRanks. Carried in the config so the (cached) default-engine build closure reads it per call instead of capturing a stale local.

final

maxConcurrentSessions → int?

final

maxNumImages → int?

final

maxTokens → int

final

modelPath → String

Resolved on-disk path to the model file. Core's platform createModel preamble resolves it from the active spec via the model manager and passes it here so the engine package never touches core's file-path resolution.

final

multimodalOrder → MultimodalOrder?

The order of a user message's images, audio and text. Core always sets it: getActiveModel resolves an omitted one with MultimodalOrder.forModelType, and a null in a config built by hand means the same. Honored by the native LiteRT-LM engine; the other engines ignore it (MediaPipe sends images, audio, then text).

final

preferredAudioBackend → PreferredBackend?

Backend for the AUDIO encoder, independent of preferredBackend. Null lets the engine choose, as for preferredVisionBackend, and the native LiteRT-LM engine reports it in InferenceModel.activeAudioBackend. Set one only where the section allows it: the published Gemma 4 and Gemma 3n audio encoders declare backend_constraint cpu, so PreferredBackend.gpu is refused for them (readLiteRtLmBundleInfo(path).audioEncoder?.backendConstraint shows it). Ignored by MediaPipe.

final

preferredBackend → PreferredBackend?

final

preferredVisionBackend → PreferredBackend?

Backend for the VISION encoder, independent of preferredBackend (the text/decoder backend). Null lets the engine choose: the native LiteRT-LM engine follows the bundle — the encoder's declared backend_constraint (the CPU when that allows it), else the NPU it is compiled for, else the CPU, because the vision encoder's STABLEHLO_COMPOSITE ops fail to prepare on the Metal/WebGPU delegates (LiteRT-LM#2461) — and reports its choice in InferenceModel.activeVisionBackend. Set one only where the section allows it (readLiteRtLmBundleInfo(path).visionEncoder in flutter_edge_ai_litertlm shows its backend_constraint); a backend it excludes is refused before the model loads. Ignored by the MediaPipe engine (whole-model backend, no per-encoder split).

final

runtimeType → Type

A representation of the runtime type of the object.

no setter, inherited

supportAudio → bool

Inference: whether the model is created with audio. Embedders: as supportImage, for audio.

final

supportImage → bool

Inference: whether the model is created with vision. Embedders: the media the caller declared (getActiveEmbedder(supportImage:)), widened by core to everything the cached embedder already served.

final

tokenizerPath → String?

Resolved on-disk path to the tokenizer. Used by embedding and STT backends; null for inference, and null for a self-contained embedder (EmbeddingModelSpec.isSelfContained, e.g. a .litertlm bundle that carries its own tokenizer). The spec carries source identities (network/asset/file); core resolves them to on-disk paths via the model manager and passes the resolved tokenizer path here (install-vs-runtime separation).

final

Methods

noSuchMethod(Invocation invocation) → dynamic

Invoked when a nonexistent method or property is accessed.

inherited

toString() → String

A string representation of this object.

inherited

Operators

operator ==(Object other) → bool

The equality operator.

inherited