mlx-bun library API
    Preparing search index...
    interface ExecutionCapabilities {
        checkpoints: boolean;
        compiledDecode?: boolean;
        grammarBatch: boolean;
        groupedMethods?: readonly string[];
        mediaBatch?: boolean;
        mediaPrefixCache?: boolean;
        method: "autoregressive" | "denoising";
        sharedCheckpoints?: boolean;
        sharedFill?: boolean;
        sharedGrammarJump?: boolean;
        sharedGrammarProposals?: boolean;
        sharedSpeculativeAdapters?: boolean;
        sharedSpeculativeEcho?: boolean;
        speculativeLogprobs?: boolean;
    }
    Index
    checkpoints: boolean
    compiledDecode?: boolean

    The graph owns a compiled decode step.

    grammarBatch: boolean

    Executor coordination, not a model capability: the shared group drives each row's own grammar state (wait for its mask, advance it with the emitted token, stop the row when it terminates). The shared sampler masks any model's logits; grammar jump and speculation are qualified separately.

    groupedMethods?: readonly string[]

    Methods supplied by the model's shared execution binding.

    mediaBatch?: boolean

    Prepared media can enter the ordinary shared execution group.

    mediaPrefixCache?: boolean
    method: "autoregressive" | "denoising"
    sharedCheckpoints?: boolean

    Bound ordinary driver can restore/capture this request configuration.

    sharedFill?: boolean

    A grouped method can consume this request's strict known continuations.

    sharedGrammarJump?: boolean

    An AR method appends grammar-accepted spans without speculative verification.

    sharedGrammarProposals?: boolean

    A shared method accepts request-owned grammar continuation proposals.

    sharedSpeculativeAdapters?: boolean

    The grouped provider supports the target's mounted adapter context.

    sharedSpeculativeEcho?: boolean

    A mounted provider can consume verified external continuations.

    speculativeLogprobs?: boolean