skainet-io-gguf/sk.ainet.io.gguf

Package-level declarations

Types

data class FieldParts(val size: Int, val parts: List<List<Any>>, val idxs: List<Int>, val types: List<GGUFValueType>)

enum GGMLQuantizationType : Enum<GGMLQuantizationType>

GGML quantization types.

data class GgufModelMetadata(val architecture: String?, val name: String?, val author: String?, val license: String?, val version: String?, val url: String?, val classNames: List<String>?, val numClasses: Int?, val inputSize: Int?, val contextLength: Int?, val embeddingLength: Int?, val headCount: Int?, val layerCount: Int?, val vocabSize: Int?, val tokenizerModel: String? = null, val tokenizerTokens: List<String>? = null, val tokenizerMerges: List<String>? = null, val tokenizerTokenTypes: List<Int>? = null, val bosTokenId: Int? = null, val eosTokenId: Int? = null, val rawFields: Map<String, Any?>)

Parsed model metadata from a GGUF file.

GgufModelParser

class GgufModelParser : BaseModelParser, AutoCloseable

GGUF model parser extending BaseModelParser. Provides metadata-first parsing of GGUF model files using SKaiNET's GGUF I/O capabilities.

GGUFModelReader

class GGUFModelReader : ModelReader

GgufParametersLoader

class ~~GgufParametersLoader~~(sourceProvider: () -> Source, onProgress: (current: Long, total: Long, message: String?) -> Unit = { _, _, _ -> }) : ParametersLoader

ParametersLoader implementation backed by the legacy GGUFReader.

GGUFReader

class GGUFReader(source: Source, loadTensorData: Boolean = true, decodeF16ToFloat: Boolean = true, decodeBF16ToFloat: Boolean = true)

GGUFValueType

enum GGUFValueType : Enum<GGUFValueType>

ReaderField

data class ReaderField(val offset: Int, val name: String, val parts: List<List<Any>> = emptyList(), val data: List<Int> = listOf(-1), val types: List<GGUFValueType> = emptyList())

ReaderTensor

data class ReaderTensor(val name: String, val tensorType: GGMLQuantizationType, val rawTypeValue: Int, val shape: List<UInt>, val nElements: Int, val nBytes: Int, val dataOffset: Int, val data: List<Any>, val field: ReaderField)

StreamingGgufParametersLoader

class StreamingGgufParametersLoader(sourceProvider: () -> RandomAccessSource, onProgress: (current: Long, total: Long, message: String?) -> Unit = { _, _, _ -> }) : ParametersLoader

Streaming GGUF parameters loader — the recommended path for loading GGUF models.

StreamingGGUFReader

class StreamingGGUFReader : AutoCloseable

Streaming GGUF reader that parses metadata without loading the entire file.

StreamingTensorInfo

data class StreamingTensorInfo(val name: String, val shape: List<UInt>, val tensorType: GGMLQuantizationType, val rawTypeValue: Int, val nElements: Long, val nBytes: Long, val relativeOffset: Long, var absoluteDataOffset: Long)

Tensor metadata for streaming access.

TensorNameMapper

interface TensorNameMapper

Abstracts the mapping from logical tensor roles to GGUF tensor name strings.

Properties

GGML_QUANT_SIZES

val GGML_QUANT_SIZES: Map<GGMLQuantizationType, Pair<Int, Int>>

GGUF_DEFAULT_ALIGNMENT

const val GGUF_DEFAULT_ALIGNMENT: Int = 32

GGUF_MAGIC

const val GGUF_MAGIC: UInt

This is a kotlin gguf reader related logic interpreted from python code "gguf-py/gguf/constants.py" of github repo "https://github.com/ggerganov/llama.cpp"

GGUF_VERSION

const val GGUF_VERSION: Int = 3

QK_K

const val QK_K: Int = 256

READER_SUPPORTED_VERSIONS

val READER_SUPPORTED_VERSIONS: List<Int>

This is a kotlin gguf reader interpreted from python code "gguf-py/gguf/gguf_reader.py" of github repo "https://github.com/ggerganov/llama.cpp"