export type LocalInferenceProfileMode = "balanced" | "low-impact";
export interface LocalInferenceProfile {
    mode: LocalInferenceProfileMode;
    promptCacheMiB: number;
    maxLoadedModels: number;
    parallelRequests: number;
    maxQueuedRequests: number;
    keepAlive: string;
    kvCacheType: "q8_0";
    flashAttention: boolean;
    generationThreads: number;
    batchThreads: number;
}
interface LocalInferenceHostCapacity {
    mode: LocalInferenceProfileMode;
    totalMemoryBytes: number;
    logicalCpuCount: number;
}
export interface LocalInferenceHostProbes {
    totalMemoryBytes?: () => number;
    logicalCpuCount?: () => number;
}
/**
 * One provider-neutral owner for local inference resource policy. Runtime adapters translate this
 * validated profile into their native arguments; they must not invent independent cache,
 * residency, concurrency, or thread defaults.
 */
export declare function deriveLocalInferenceProfile(args: LocalInferenceHostCapacity): LocalInferenceProfile;
/** Resolves host capacity exactly once through the provider-neutral profile owner. Runtime
 * adapters pass their injectable probes here instead of reproducing host-default selection. */
export declare function deriveHostLocalInferenceProfile(mode: LocalInferenceProfileMode, probes?: LocalInferenceHostProbes): LocalInferenceProfile;
export declare function ollamaEnvironmentForLocalInferenceProfile(profile: LocalInferenceProfile): Record<string, string>;
export {};
//# sourceMappingURL=local-inference-profile.d.ts.map