v3.5.6
WebGPU Inference Support for Electron
-
Consistency Testing Added
WebGPU inference may produce inconsistent results in Electron apps, particularly on some Intel GPUs. In contrast, the same models typically run reliably in Chromium-based browsers. -
Automatic Fallback to WASM
The SDK now performs a consistency check using reference data after initialization. If inconsistencies are detected, it will automatically fall back to the WASM backend. -
Available for Balanced Preset Only
The WebGPU consistency test is implemented only when using theBalancedpreset. -
Enable via Config
sdk.config({ test_inference: true, });
-
Native Inference Modules for Electron
Native inference modules are available for Electron, offering better performance and lower resource usage compared to the WASM backend. See documentation for more details.
Bug Fixes & Stability Improvements
- Improved resource caching logic
- Enhanced precision in getMetrics, especially during the first second of execution
- General bug fixes and overall stability improvements