Skip to content

Astral Roadmap

Astral's next platform work is WebAssembly and WebGPU. The goal is to preserve the same C ABI and explicit ownership model while adding a browser deployment profile and a portable GPU execution path.

WebAssembly

The first step is a WebAssembly CPU build that can run the mock provider, tokenization, embeddings, and compact memory search without dynamic loading or native virtual-memory assumptions.

The maintained profile will define:

  • Emscripten build and test presets.
  • Fixed and growable linear-memory configurations.
  • Single-threaded operation plus an opt-in worker-thread configuration where the browser provides the required isolation features.
  • A small JavaScript boundary that keeps model bytes and generated output under explicit ownership.
  • Browser tests for initialization, streaming, cancellation, reset, embedding, and memory-index persistence.

WebGPU

WebGPU will be added as a backend rather than as a separate public API. Device, queue, buffer, and pipeline ownership will remain behind Astral's provider boundary.

Initial work will cover:

  • Adapter and device selection with explicit capability reporting.
  • Pipeline caching outside dispatch loops.
  • Buffer layouts that avoid per-token uploads and unnecessary host copies.
  • f32 and compact-storage kernels with CPU reference comparisons.
  • Browser and native WebGPU validation using the same correctness fixtures.

Acceptance

WebAssembly and WebGPU support will be listed as available only after the checked-in presets build from a clean checkout and the browser tests exercise the public API. Performance claims will include the browser, device, model, threading mode, memory configuration, and raw measurement command.