Enums and Constants

The LLM sections on this page reflect ZeticMLange Android 1.5.9.

This page documents the enums and configuration types most commonly used from the Android SDK.

Package

com.zeticai.mlange

ModelMode

Controls inference mode selection for general, non-LLM models.

enum class ModelMode {
    RUN_AUTO,
    RUN_SPEED,
    RUN_ACCURACY
}

LLMModelMode

Controls automatic target selection for ZeticMLangeLLMModel.

enum class LLMModelMode {
    RUN_AUTO,
    RUN_SPEED,
    RUN_ACCURACY,
}

Value	Description
`RUN_AUTO`	Default. Uses metadata-based selection to balance practical speed and quality.
`RUN_SPEED`	Prioritizes lower latency.
`RUN_ACCURACY`	Prioritizes higher score or lower loss.

val model = ZeticMLangeLLMModel(
    context = context,
    personalKey = PERSONAL_KEY,
    name = MODEL_NAME,
    modelMode = LLMModelMode.RUN_SPEED,
)

LLMDataSetType

Optional dataset hint for RUN_ACCURACY.

enum class LLMDataSetType {
    MMLU,
    TRUTHFULQA,
    CNN_DAILYMAIL,
    GSM8K,
}

Value	Description
`MMLU`	Massive Multitask Language Understanding.
`TRUTHFULQA`	Truthfulness and informativeness benchmark.
`CNN_DAILYMAIL`	Summarization benchmark.
`GSM8K`	Grade-school math reasoning benchmark.

val model = ZeticMLangeLLMModel(
    context = context,
    personalKey = PERSONAL_KEY,
    name = MODEL_NAME,
    modelMode = LLMModelMode.RUN_ACCURACY,
    dataSetType = LLMDataSetType.MMLU,
)

Pass null when you do not want to specify a dataset.

LLMKVCacheCleanupPolicy

Controls how the conversation KV cache is handled when it reaches capacity.

enum class LLMKVCacheCleanupPolicy {
    CLEAN_UP_ON_FULL,
    DO_NOT_CLEAN_UP
}

Value	Description
`CLEAN_UP_ON_FULL`	Clears the conversation context when the KV cache is full. Default.
`DO_NOT_CLEAN_UP`	Keeps the context as-is. You must call `cleanUp()` before the next conversation.

LLMInitOption

Configures LLM runtime initialization.

data class LLMInitOption(
    val kvCacheCleanupPolicy: LLMKVCacheCleanupPolicy = LLMKVCacheCleanupPolicy.CLEAN_UP_ON_FULL,
    val nCtx: Int = 2048,
)

Field	Type	Description
`kvCacheCleanupPolicy`	`LLMKVCacheCleanupPolicy`	KV-cache cleanup behavior during generation.
`nCtx`	`Int`	Requested context length.

nCtx is not guaranteed to be applied exactly as provided. The runtime can adjust it internally depending on the model or backend.

LLMTarget

Manual runtime selection target for ZeticMLangeLLMModel.

LLMTarget.LLAMA_CPP

Value	Description
`LLAMA_CPP`	Use this value for explicit GGUF selection.

LLMQuantType

GGUF quantization types used with LLMTarget.LLAMA_CPP.

enum class LLMQuantType {
    GGUF_QUANT_ORG,
    GGUF_QUANT_F16,
    GGUF_QUANT_BF16,
    GGUF_QUANT_Q8_0,
    GGUF_QUANT_Q6_K,
    GGUF_QUANT_Q4_K_M,
    GGUF_QUANT_Q3_K_M,
    GGUF_QUANT_Q2_K,
    GGUF_QUANT_NUM_TYPES,
}

APType

Processor type used with explicit runtime selection.

enum class APType {
    CPU,
    GPU,
    NPU,
    NA,
}

Device / runtime	Supported values
Qualcomm Android + `LLAMA_CPP`	`CPU`, `GPU`, `NPU`
Other Android devices + `LLAMA_CPP`	`CPU`

ModelCacheHandlingPolicy

Managed artifact cache policy for downloaded model files.

enum class ModelCacheHandlingPolicy {
    REMOVE_OVERLAPPING,
    KEEP_EXISTING,
}

Value	Description
`REMOVE_OVERLAPPING`	Default. Removes overlapping cached aliases for the same model entry.
`KEEP_EXISTING`	Keeps older cached aliases and artifacts.

Detailed semantics for ModelCacheHandlingPolicy are currently TBD. See Cache Management.

ModelLoadingStatus

Status values delivered by the Android-only onStatusChanged callback.

enum class ModelLoadingStatus {
    UNKNOWN,
    PENDING,
    DOWNLOADING,
    TRANSFERRING,
    COMPLETED,
    FAILED,
    CANCELED,
    WAITING_FOR_WIFI,
    NOT_INSTALLED,
    REQUIRES_USER_CONFIRMATION,
}