Hardware Acceleration

Request LiteRTAccelerator.CPU, GPU, or NPU when creating a compiler. CPU is the default.

val compiler = LiteRTCompiler(modelPath, LiteRTAccelerator.CPU)

These are requests, not guarantees that the device, model operators, runtime, plugins, or browser support the hardware. Establish a CPU correctness baseline first, then test performance and fallback on the target device.

Implementation	Current behavior and considerations
Android	Passes the selection to Android LiteRT; support depends on runtime and device.
JVM	Maps to LiteRT hardware options; the native loader logs a CPU fallback warning when the GPU acceleration plugin is unavailable.
Native	Maps to LiteRT C API hardware options; plugins, dynamic libraries, and device capabilities must match.
JS / WasmJS	CPU uses WASM; GPU requests select WASM when WebGPU is unavailable; NPU requests use WebNN options. Verify WebNN on the target browser.

Successful initialization does not necessarily mean the entire model runs on the requested hardware. For performance-sensitive applications, measure initialization, data transfer, and repeated inference separately, and inspect runtime logs.