Skip to content

v0.1.0: measured WebGPU and WASM browser inference

Latest

Choose a tag to compare

@lucioduran lucioduran released this 09 Nov 20:39

Initial public release of a small, inspectable support-ticket classifier with native WebGPU and WebAssembly kernels, explicit offline caching and reproducible browser evidence.

Verified source: 7279202a64b6a047ca5f2ad88efd0360e876b333.
CI: https://github.com/lucioduran/webgpu-local-inference/actions/runs/42695016736

Real in-app Chromium 157 execution on one Mac Pro with an AMD RDNA 1 adapter produced 36/36 correct predictions per backend in each of the online and cached-reload phases. All 144 scores matched the float32 reference exactly. The full measured workload ran under the shared exclusive benchmark lock. Raw evidence includes download sizes and timing, initialization, individual inference durations, cancellation, explicit device destruction and injected unavailable-adapter fallback to real WASM.

WebGPU online initialization was 49.2 ms and median inference 0.4 ms. WASM initialization was 0.6 ms, with 34/36 timings below timer resolution. No GPU speedup or general accuracy claim is made. The task is a fixed synthetic fixture; offline verification is a static origin outage with the report API reachable. One physical device and one browser environment were measured. Safari connection failures are preserved; Safari compatibility is not claimed. Memory figures are logical allocation sizes, not measured process memory.

Validation: 10 tests, lint, formatting, deterministic artifact build, source/hash checks, offline numerical replay and offline installed-package smoke all passed. The release's browser sources match measured revision 615e4051094b1b9ec865d920503a1516947cae71; later commits only add verification, packaging, metadata handling and documentation.

Offline lightweight replay from the source checkout with Node 24.8.0:

node scripts/replay.mjs

Attached package includes browser assets, model, binary and evidence. The JSON report and summary are also attached separately. MIT licensed code, original fixtures and model. No runtime third-party services or paid compute.