mlx-bun library API
    Preparing search index...

    Interface DenoisingGraph<Tensor, State>

    Denoising has canvas state and feedback, independent of AR token caches. Returned tensors/state are owned by the run. Inputs remain borrowed. A fused method may bypass this graph and implement InferenceMethod directly.

    interface DenoisingGraph<Tensor, State> {
        canvasLength: number;
        descriptor: GraphDescriptor;
        embedScale: number;
        vocabSize: number;
        closeState(state: State): void;
        decoderLogits(
            canvas: Tensor,
            state: State,
            feedback: Tensor | null,
        ): Tensor;
        dequantEmbedWeight(): Tensor;
        extendPrefill(tokens: Tensor, state: State): void;
        prefill(promptIds: number[], vision?: Tensor): State;
        softEmbeddings(logits: Tensor, weight: Tensor): Tensor;
    }

    Type Parameters

    • Tensor
    • State
    Index
    canvasLength descriptor embedScale vocabSize
    closeState decoderLogits dequantEmbedWeight extendPrefill prefill softEmbeddings
    canvasLength: number
    descriptor: GraphDescriptor
    embedScale: number
    vocabSize: number

    Settings
    Member Visibility

    On This Page
    Properties
    Methods