Backends

textsift picks the fastest available backend at create() time. The public API (PrivacyFilter.create, redact, detect) is identical across all of them — the backend is an implementation detail.

Backend	Where it runs	Compute
metal-direct	macOS (Node)	Hand-written MSL kernels via Obj-C bridge
vulkan-direct	Linux (Node)	Hand-written GLSL → SPIR-V via glslangValidator
dawn-direct	Windows (Node) + Linux fallback	Tint → D3D12 / Vulkan via statically-linked Dawn
webgpu	Browsers	Custom WGSL kernels (int4 matmul, banded attention, sparse MoE). Requires shader-f16.
wasm	Anywhere with no GPU	Custom Zig + SIMD128 → WASM. Multi-thread when COOP/COEP headers set.

Backend decision tree

 PrivacyFilter.create({ backend?: ... })
              │
              ▼
 ┌────────────────────────────────────────┐
 │ explicit backend ?                     │
 │   "webgpu" / "wasm" → use it           │
 └────────────┬───────────────────────────┘
              │ (no explicit backend or "auto")
              ▼
        ┌─────────────┐
        │  Node?      │
        └──┬────────┬─┘
        yes│        │no
           ▼        ▼
   ┌─────────────┐  ┌─────────────────────────────┐
   │ try native  │  │ navigator.gpu + shader-f16? │
   │  (Metal /   │  └──────┬──────────────────┬───┘
   │  Vulkan /   │     yes │              no  │
   │  Dawn)      │         ▼                  ▼
   │             │   backend:"webgpu"   backend:"wasm"
   │ if fails →  │
   │   wasm      │
   └─────────────┘

Compatibility

	Chromium 147	Firefox 129	Safari 19	Node 22 (macOS)	Node 22 (Linux)	Node 22 (Win)
webgpu	✅	⚠️ shader-f16 preview	⚠️ shader-f16 preview	—	—	—
metal-direct	—	—	—	✅	—	—
vulkan-direct	—	—	—	—	✅ (Mesa loader)	—
dawn-direct	—	—	—	—	✅ fallback	✅
wasm	✅	✅	✅	✅ fallback	✅ fallback	✅ fallback

textsift’s WASM backend is a from-scratch implementation that loads model_q4f16.onnx directly. Within the standard JS-app ecosystem this is the only working CPU path — transformers.js with device: "wasm" fails session creation on this model because ORT-Web has no implementations for the GatherBlockQuantized / MatMulNBits ONNX contrib ops the int4 export uses. Other runtimes (onnxruntime-node with custom ops, web-llm) can in principle run it but require setup textsift doesn’t.

Output parity

All backends produce byte-identical spans on the same input. Conformance is enforced kernel-by-kernel against the canonical browser WGSL fixtures (15/15 pass for both Metal-direct and Vulkan-direct). Logit magnitudes drift by up to ~0.2 RMS due to fp16 rounding differences accumulating across 8 transformer layers, but argmax (and hence Viterbi + span decode) is preserved.

Switching backends at runtime

A PrivacyFilter instance is bound to one backend for its lifetime. To switch, dispose and recreate:

const wasm = await PrivacyFilter.create({ backend: "wasm" });
await wasm.redact(text);
wasm.dispose();

const gpu = await PrivacyFilter.create({ backend: "webgpu" });

Both share the same model cache (OPFS in browsers, filesystem at ~/.cache/textsift/ in Node) for the 770 MB weights — switching is a ~1-second warmup, not a full re-download.

Memory footprint

Backend	Resident memory during inference
GPU (any: webgpu / metal-direct / vulkan-direct / dawn-direct)	~30 MB JS heap + ~800 MB GPU buffers
wasm	~800 MB WASM linear memory

GPU backends are the most memory-efficient since weights live in GPU buffers and JS holds only kernel-dispatch state.