Skip to content

README.md

Latest commit

 

History

History
58 lines (41 loc) · 1.94 KB

README.md

File metadata and controls

58 lines (41 loc) · 1.94 KB
Raw

Browser WebGPU demo (Small Q8)

Experimental static site for running UniFlow-Audio Small (dit-Q4_0 + pack T5 Q4_0) in Chrome/Edge via ggml WebGPU (Emscripten). Q4 keeps the download smaller for demos.

Weights are not stored in git. The page fetches them from audiohacking/uniflow-audio-gguf.

Build

Needs Emscripten ≥ 4.0.10 (CI uses 4.0.23). Current ggml-webgpu needs a newer Dawn emdawnwebgpu package than Emscripten 4.0.10–4.0.12 ship by default — make web / the Pages workflow fetch a pinned package via scripts/fetch_emdawnwebgpu.sh.

Built with ASYNCIFY (not JSPI) so WebGPU WaitAny works without WebAssembly.promising / Chrome experimental flags.

# Homebrew emscripten, or source emsdk_env.sh
export PATH="/opt/homebrew/opt/emscripten/bin:$PATH"

cd /path/to/uniflow.cpp
git submodule update --init --recursive

make web
# outputs: web/uniflow-web.{js,wasm}

Local serve

npx --yes serve web -p 8080
# open http://localhost:8080  (Chrome with WebGPU)

GitHub Pages

Workflow: .github/workflows/pages-webgpu.yml

  1. Settings → Pages → Build and deployment → Source = GitHub Actions (once).
  2. On push to main (or Actions → Deploy WebGPU demo → Run workflow):
    • build job: recursive submodule checkout → Emscripten + WebGPU compile → upload artifact
    • deploy job: actions/deploy-pages only
  3. Site URL is printed on the deploy job (environment: github-pages).

Users need a WebGPU browser and enough RAM for ~520 MB weights + WASM heap.

API (WASM)

Export Role
initBackend() Init WebGPU + CPU sched
loadModels(t5, dit, vae, instructions) MEMFS paths
generateFromTokens(ids, duration, steps, cfg, sway, seed) PCM length
getPcm() / pcmSampleRate() / freePcm() Copied float32 PCM

Tokenization is done in JS (@huggingface/transformers).