Speech / ONNX WebGPU

CPU + WebGPU twins for browser speech (STT/TTS/VAD) and ONNX embeddings. Honest accel: publish webgpu only when ORT AppendExecutionProvider("WebGPU") succeeds.

	CPU	WebGPU
ORT archive	`third_party/onnxruntime-wasm/`	`third_party/onnxruntime-wasm-webgpu/`
Vendor script	`wasm/scripts/vendor-onnxruntime-wasm.sh`	`wasm/scripts/vendor-onnxruntime-wasm-webgpu.sh`
WASM out	`racommons-onnx-sherpa.{js,wasm}`	`racommons-onnx-sherpa-webgpu.{js,wasm}`

Separate trees keep provenance (webgpu=on only on the GPU archive) and ORT build caches from colliding. One script with a --webgpu flag would work; twins match the llama pattern.

WebGPU vendor also stages Dawn emdawn JS under onnxruntime-wasm-webgpu/emdawn/ so the ORT source build/ tree can be deleted after vendoring without breaking --onnx-webgpu links.

ONNX.register({ acceleration: 'auto'|'cpu'|'webgpu', threads }) — badge ×N is pthread/intra-op threads (clamped 1–8), not multiple GPUs. Default example uses threads: 2.

Do not regress: WebGPU twin must compile rac_runtime_onnxrt with RAC_ONNXRT_EP_WEBGPU_ENABLED (onnxrt CMakeLists). Never #define RAC_ONNXRT_EP_WEBGPU (enum collision). Final-wasm COMPILE_DEFS alone is not enough.

Build / release

bash

# from sdk/runanywhere-web
source emsdk/emsdk_env.sh
npm run vendor:wasm:speech          # CPU ORT + WebGPU ORT + Sherpa
npm run build:wasm:all              # includes --onnx-webgpu

# demo
cd ../../examples/web/RunAnywhereAI && npm run release:build

Gate: both speech WASM pairs non-empty; demo release.sh requires the WebGPU pair in dist; browser COI shows Speech/Embeddings: WebGPU and ONNX.lastFallbackReason == null (or __RUNANYWHERE_ONNX_DIAG__). Forced 'webgpu' must throw if the twin/probe is broken.

Thread soak (optional): measure wall time / Aborted() at threads 1/2/4 on Whisper Tiny, Canary, Nemotron under COOP/COEP.