/**
 * Inference scheduling benchmark harness (3.0.0 foundation).
 *
 * Synthetic-backend benchmark over the real SharedInferenceScheduler. It
 * measures scheduling behavior (queue latency, TTFT, busy-slot utilization,
 * avoidable idle) independently of real Qwen. Real-Qwen measurements are a
 * separate QA artifact; synthetic capacity profiles are labeled as such.
 */
export interface InferenceSchedulerBenchmarkOptions {
    capacity: number;
    agentCount: number;
    requestsPerAgent: number;
    /** Synthetic generation wall time per request. */
    generationMs: number;
    /** Synthetic tool time between an agent's sequential requests. */
    toolMs?: number;
    /** Deterministic queue-wait deadline (0 = unbounded). */
    queueWaitTimeoutMs?: number;
}
export interface InferenceSchedulerBenchmarkResult {
    profile: {
        backend: "synthetic";
        capacity: number;
        agentCount: number;
        requestsPerAgent: number;
        generationMs: number;
        toolMs: number;
    };
    requestCount: number;
    queueWaitMs: {
        p50: number;
        p95: number;
        max: number;
    };
    ttftMs: {
        p50: number;
        p95: number;
        max: number;
    };
    perRequestDecodeTokensPerSec: {
        p50: number;
        p95: number;
    };
    aggregateModelTokensPerSec: number;
    busySlotPercent: number;
    avoidableIdleMs: number;
    totalWallMs: number;
    completedCount: number;
    failedCount: number;
}
/**
 * Run a synthetic concurrency-profile benchmark. Each logical agent issues
 * `requestsPerAgent` sequential generations through the scheduler (capacity
 * `capacity`). All agents run concurrently, so queueing emerges naturally.
 */
export declare function runInferenceSchedulerBenchmark(options: InferenceSchedulerBenchmarkOptions): Promise<InferenceSchedulerBenchmarkResult>;
//# sourceMappingURL=benchmark.d.ts.map