gpu-time

A compact neural parser for English schedules. One small trained model reads a natural-language time expression and returns concrete dates, time ranges, and RFC 5545 recurrence rules. The public API returns dates and rules without internal token labels or syntax trees.

pnpm add gpu-time

import { parse } from "gpu-time";

const result = await parse("Sat Sun 1pm-8pm Mon 10pm-12am", {
  reference: "2026-09-09T12:00:00+06:00",
  timeZone: "Asia/Dhaka",
});

console.log(result.occurrences);
// Saturday and Sunday: 13:00–20:00
// Monday: 22:00–Tuesday 00:00
console.log(result.rrules);
console.log(result.diagnostics);

parse(text, context) and parseMany(texts, context) are the convenience entry points. context must carry a reference instant and a timeZone; an unknown timezone rejects the call. The result holds occurrences (ISO start, optional end, allDay), rrules, truncated, diagnostics, the backend that ran, and timings. Diagnostics report known problems, but model errors can still produce incorrect dates.

For a reusable instance or explicit backend selection, use defineParser({ backend }) and release its GPU resources with dispose():

import { defineParser } from "gpu-time";

const parser = await defineParser({ backend: "webgpu" });
const results = await parser.parseMany(texts, context);
parser.dispose();

Automatic mode tries WebGPU at 32 inputs or 512 tokens per batch. Smaller batches use the CPU. If WebGPU fails in automatic mode, the parser uses CPU and reports a fallbackReason. Explicit backend: "webgpu" requests reject on failure. GPU calls reuse the device, pipelines, and weights.

The model can return incorrect dates, even without a diagnostic. Before you use a result, make sure that its dates match the intended schedule. See the project repository for architecture, benchmarks, and model limitations.