// Exact generation throughput, from Ollama's own final metrics. // // The done:true chunk of /api/chat carries eval_count (generated tokens) and // eval_duration (nanoseconds spent generating). Those are authoritative for // the completed attempt - wall-clock time would fold in model load and prompt // evaluation, and chunk counts are not token counts on batched streams // (issue #4). Pure functions only; the provider reports, the extension layer // formats and displays. import type { OllamaChunk } from "./wire.js"; export interface GenerationMetrics { outputTokens: number; evalDurationNs: number; tokensPerSecond: number; promptTokens?: number; promptEvalDurationNs?: number; } type MetricFields = Pick< OllamaChunk, "eval_count" | "eval_duration" | "prompt_eval_count" | "prompt_eval_duration" >; function isPositiveFinite(n: unknown): n is number { return typeof n === "number" && Number.isFinite(n) && n > 0; } /** * Exact metrics from a done:true chunk, or undefined when they can't produce * a finite rate (older/remote servers omit them; a zero duration would divide * to Infinity). Undefined means "show nothing", never a made-up number. */ export function parseGenerationMetrics( chunk: MetricFields, ): GenerationMetrics | undefined { const outputTokens = chunk.eval_count; const evalDurationNs = chunk.eval_duration; if (!isPositiveFinite(outputTokens) || !isPositiveFinite(evalDurationNs)) { return undefined; } const metrics: GenerationMetrics = { outputTokens, evalDurationNs, tokensPerSecond: outputTokens / (evalDurationNs / 1e9), }; if (isPositiveFinite(chunk.prompt_eval_count)) { metrics.promptTokens = chunk.prompt_eval_count; } if (isPositiveFinite(chunk.prompt_eval_duration)) { metrics.promptEvalDurationNs = chunk.prompt_eval_duration; } return metrics; } /** Completed-generation status: Ollama-reported, so one decimal is honest. */ export function formatExactThroughput(metrics: GenerationMetrics): string { return `${metrics.tokensPerSecond.toFixed(1)} tok/s · ${metrics.outputTokens.toLocaleString("en-US")} tok`; } /** Live status: an estimate, so it is marked and carries no decimals. */ export function formatEstimatedThroughput(tokensPerSecond: number): string { return `≈${Math.round(tokensPerSecond)} tok/s`; }