Parabun

A fork of Bun with four additional runtime modules — a worker pool, SIMD primitives for typed arrays, a compute-only GPU surface, and a GGUF/Llama runtime — and parse-time extensions for purity, memoization, and reactive bindings in .pts / .pjs files. Regular .ts and .js files behave the same as in upstream Bun.

Linux and macOS. Windows build is in progress. parabun self-update refreshes an existing install along with the VS Code extension.

Installs the VS Code extension into any of code, cursor, or kiro found on $PATH. The extension provides the .pts / .pjs TextMate grammar and an LSP with hover, go-to-definition, purity diagnostics, memo hints, and operator documentation.

Runtime modules

import { pmap } from "bun:parallel";

pure function score(row) { return row.reduce((a, b) => a + b * b, 0); }

const rows = new Float32Array(new SharedArrayBuffer(1_000_000 * 4));
// ...fill rows...
const scores = await pmap(score, rows, { concurrency: 8 });

import { mulScalar, add, dot, sum } from "bun:simd";

const y = mulScalar(new Float32Array([1, 2, 3, 4]), 3); // [3, 6, 9, 12]
const z = add(a, b);
const d = dot(u, v);
const s = sum(a);

import gpu from "bun:gpu";

const mat = gpu.alloc(M * K, "f32");
// ...fill mat...
const held = gpu.hold(mat);                   // uploaded once
for (const q of queries) {
  const scores = gpu.matVec(held, q, M, K); // no copy
}
gpu.release(held);

import llm from "bun:llm";

using m = await llm.LLM.load("./Llama-3.2-1B-Instruct-Q4_K_M.gguf");

for await (const piece of m.chat([
  { role: "system", content: "You are helpful and concise." },
  { role: "user", content: "What is the capital of France?" },
])) {
  process.stdout.write(piece);
}

Example: LangChain VectorStore

ParabunVectorStore extends VectorStore from @langchain/core and implements the addVectors and similaritySearchVectorWithScore methods, so call sites that accept any VectorStore work against it without changes.

import { MemoryVectorStore }
  from "langchain/vectorstores/memory";

const store = new MemoryVectorStore(emb);
await store.addVectors(vectors, docs);
const hits = await store
  .similaritySearchVectorWithScore(q, 10);

import { ParabunVectorStore }
  from "./parabun-store.pjs";

const store = new ParabunVectorStore(emb);
await store.addVectors(vectors, docs);
const hits = await store
  .similaritySearchVectorWithScore(q, 10);

add_ms is higher because rows are packed into a single SAB Float32Array and normalized in place — one-time O(N·D) work amortized across subsequent queries. Top-K indices and scores match LangChain's to four decimal places.

Language extensions — .pts / .pjs

Files ending in .pts, .ptsx, .pjs, or .pjsx are parsed with additional desugarings. All output is standard JS; no runtime support is required, and the runtime modules above do not depend on any of this syntax. GitHub's TextMate grammars do not cover .pts; the VS Code / Cursor / Kiro extension provides the grammar and an LSP.

pure and memo

A pure function is rejected at parse time if it mutates an outer variable, reads this, or calls a known-impure global. Prefix pure with memo — or drop pure entirely and write memo as the declarator — and the result is cached by argument identity: 0-arg singleton, 1-arg Map, multi-arg nested Map chain. Recursive self-references route through the outer wrapper, so fib below runs the body 21 times for fib(20), not 21,891.

// declarator form — `memo` implies pure + function
memo fib(n: number): number {
  return n < 2 ? n : fib(n - 1) + fib(n - 2);
}

// arrow form — same thing as an expression prefix
const normalize = memo (s: string) => s.trim().toLowerCase();

// async dedupes concurrent in-flight calls, evicts on reject
memo async fetchProfile(id: string) { return await db.users.get(id); }

signal, effect, ~>

signal NAME = <rhs> desugars to a Signal binding; bare reads rewrite to .get(), assignments to .set(). If the RHS references another in-scope signal, the binding auto-promotes to a read-only derived(). effect { ... } tracks every signal it reads as a dependency and re-runs on change. A ~> B is reactive binding — it desugars to effect(() => { B = A; }), so B stays in step with A and whatever signals A reads from.

signal count = 0;
signal doubled = count * 2;   // auto-derived

effect { console.log(count, doubled); }

count++;                           // effect re-runs: 1, 2

// bind signal value into a DOM-ish sink — updates track dep changes
count ~> el.innerHTML;

|>, ..!, ..&, ..=

x |> f is f(x). pure functions passed through |> get inlined at parse time — no call overhead. ..! / ..& are .catch / .finally in suffix position. ..= is = await in a declaration and disambiguates to an inclusive-range literal otherwise (0..5 excludes 5, 0..=5 includes it).

pure function sq(x: number) { return x * x; }

const result = 5 |> sq |> sq;   // 625 — both calls inlined

const json ..= fetch("/api").then(r => r.json())
  ..! err => console.error(err)      // .catch
  ..& () => console.log("done"); // .finally

for (const i of 0..=9) emit(i);                    // [0..9]

defer and arena

defer EXPR schedules EXPR to run when the enclosing block exits (return, throw, fall-through). Multiple defers dispose in LIFO order. defer await EXPR inside an async function awaits the cleanup. arena { ... } runs the block with the GC paused, then frees everything allocated inside on exit — useful for tight numeric loops with short-lived intermediate allocations.

function readConfig(path: string) {
  const fd = fs.openSync(path);
  defer fs.closeSync(fd);              // runs on every exit path
  return JSON.parse(fs.readFileSync(fd));
}

arena {
  const buf = new Float32Array(1_000_000);
  // ...numeric work...
}                                        // buf freed here, no GC pressure

Full grammar in LLMs.md, and the LSP carries arity-based "could be memo" / "memo probably not worth it" hints plus full purity diagnostics.

Roadmap

Parabun's positioning is to open typical JS performance bottlenecks via multithreading and GPU. The modules above are the foundation; the next set attack the most common "I have to shell out / use Python / write native code" pain points in JS-land.

Each module ships behind a compile-time feature flag. Production builds slim to only what your app imports — heavy codecs (ffmpeg, Apache Arrow) are opt-in, server / Lambda / Workers builds stay minimal.

bun:llm serves as proof-of-concept for the stack — built on bun:gpu + bun:simd + bun:parallel. Parabun is positioned as a perf runtime, not an AI runtime.