Core API

The public API is in io.github.kmplitert.core. These are the interfaces commonly used from shared code; platform-level bindings should not be treated as stable cross-platform abstractions.

LiteRTCompiler

LiteRTCompiler(
    filePath: String,
    accelerator: LiteRTAccelerator = LiteRTAccelerator.CPU
)

Method	Purpose
`suspend init()`	Load and prepare the model. Call before other operations.
`suspend getInputTensorType(inputName)`	Query an input's element type and layout by name.
`suspend getOutputTensorType(outputName)`	Query an output's element type and layout by name.
`suspend getInputBufferRequirements(inputName)`	Query input buffer requirements.
`suspend getOutputBufferRequirements(outputName)`	Query output buffer requirements.
`suspend getInputBuffers(signatureIndex = 0)`	Create the signature's input buffer list.
`suspend getOutputBuffers(signatureIndex = 0)`	Create the signature's output buffer list.
`suspend run(inputs, outputs, signatureIndex = 0)`	Run inference with the supplied buffers.
`suspend close()`	Release compiler resources. Do not reuse the instance afterward.

filePath is a model file path on local targets and typically an accessible model URL on the web. The model defines names, ordering, and signatures; see platform capabilities for signature differences.

TFBuffer

TFBuffer extends AutoCloseable. Writes are regular functions; reads are suspending:

Write	Read	Kotlin array
`writeInt`	`readInt`	`IntArray`
`writeFloat`	`readFloat`	`FloatArray`
`writeInt8`	`readInt8`	`ByteArray`
`writeBoolean`	`readBoolean`	`BooleanArray`
`writeLong`	`readLong`	`LongArray`

The caller must ensure the type and length match the model. See lifecycle for resource ownership.

Types and enums

Type	Description
`LiteRTAccelerator`	`CPU`, `GPU`, `NPU`; actual capability depends on the platform.
`LiteRTTensorType`	`elementType` and nullable `layout`.
`LiteRTElementType`	LiteRT element types and native `value`; `fromNative(value)` throws for unknown values. Not all enum values have `TFBuffer` read/write methods.
`LiteRTLayout`	`dimensions`, `strides`, `rank`, `hasStrides`; `calculateDefaultStrides` computes row-major element strides.
`LiteRTBufferRequirements`	`supportedTypes`, `bufferSize` in bytes, and `strides`; see stride considerations.
`LiteRTTensorBufferType`	Host and platform buffer types; `type` is the native value and `from(type)` converts it.

See the quick start for examples, image and audio pipelines for optional preprocessing APIs, and the Web initialization API for web-specific setup.