LoRA adapters

Overview

A LoRA adapter is a small set of weights applied on top of a loaded base model, changing the model’s behavior without reloading multi-gigabyte weights. The lora namespace has three verbs.

Verbs

suspend fun apply(adapterId: String, scale: Float? = null)
suspend fun remove(adapterId: String? = null)
suspend fun list(): LoraState

import com.runanywhere.sdk.public.RunAnywhere
import com.runanywhere.sdk.public.api.llm
import com.runanywhere.sdk.public.api.lora

// A base language model must already be loaded.
RunAnywhere.lora.apply("medical-qa-v1", scale = 0.8f)

println(RunAnywhere.llm.generate("What are the symptoms of diabetes?").text)

apply() looks the adapter up by id, downloads its weights when they are absent, and layers it onto the loaded base model. It throws SDKException when the adapter is unknown, incompatible, or the apply fails. scale is optional. Left null, the adapter’s own recommended scale applies, falling back to 1.0.

Scale

scale controls how strongly the adapter shifts generation. 0.0 loads it inert, 1.0 is full strength, and values above 1.0 amplify.

Removing

RunAnywhere.lora.remove("medical-qa-v1")  // one adapter
RunAnywhere.lora.remove()                 // all of them

Inspecting state

val state = RunAnywhere.lora.list()
state.applied                              // List<AppliedAdapter>

val activeId = state.applied.firstOrNull()?.id

AppliedAdapter carries id and scale.

Registering and discovering adapters

Commons keeps the adapter catalog behind a separate ABI, so registration, discovery, and explicit downloads still come from RunAnywhere.loraCatalog in com.runanywhere.sdk.public.extensions. That accessor is deprecated and kept for one release, but it is the only path to those verbs today.

import ai.runanywhere.proto.v1.LoraAdapterCatalogEntry
import com.runanywhere.sdk.public.extensions.loraCatalog

RunAnywhere.loraCatalog.register(
    LoraAdapterCatalogEntry(
        id = "medical-qa-v1",
        name = "Medical QA adapter",
        description = "Fine-tuned for medical question answering",
        url = "https://huggingface.co/your-org/medical-qa-lora/resolve/main/adapter.gguf",
        filename = "medical-qa.gguf",
        compatible_models = listOf("qwen-0.5b", "qwen-1.5b"),
        size_bytes = 25_000_000L,
        default_scale = 1f,
    ),
)

val adapters = RunAnywhere.loraCatalog.adaptersForModel("qwen-0.5b")

Field	Type	Description
`id`	`String`	The id you pass to `lora.apply()`
`name`	`String`	Display name
`description`	`String`	What the adapter does
`url`	`String`	Direct download URL for the `.gguf`
`filename`	`String`	On-disk filename
`compatible_models`	`List<String>`	Base model ids
`size_bytes`	`Long`	`0` when unknown
`default_scale`	`Float`	Recommended scale
`local_path`	`String?`	Set after download or import

Field names are snake_case because the entry is a generated proto type: url not downloadUrl, size_bytes not fileSize. loraCatalog also carries allRegistered(), queryCatalog(), download(entry, onProgress), and importAdapter(sourcePath) for a local file. lora.apply() downloads on its own, so an explicit download() is only for showing progress.

ViewModel

class LoraViewModel : ViewModel() {
    var adapters by mutableStateOf<List<LoraAdapterCatalogEntry>>(emptyList())
        private set
    var activeId by mutableStateOf<String?>(null)
        private set

    fun apply(entry: LoraAdapterCatalogEntry) {
        viewModelScope.launch {
            try {
                RunAnywhere.lora.apply(entry.id, entry.default_scale.takeIf { it > 0f })
                activeId = entry.id
            } catch (e: CancellationException) {
                throw e
            } catch (e: Exception) {
                // surface e.message
            }
        }
    }

    fun clear() {
        viewModelScope.launch {
            RunAnywhere.lora.remove()
            activeId = null
        }
    }

    private suspend fun reload(baseModelId: String) {
        adapters = RunAnywhere.loraCatalog.adaptersForModel(baseModelId)
        activeId = RunAnywhere.lora.list().applied.firstOrNull()?.id?.takeIf { it.isNotEmpty() }
    }
}

Notes

Applying or removing an adapter recreates the inference context, so the KV cache is cleared and conversation history does not carry across a swap. Every lora verb throws SDKException with ERROR_CODE_NOT_INITIALIZED before RunAnywhere.initialize().

LLM generation

Text generation with options

LLM streaming

Streaming text generation

RAG

Retrieval-augmented generation

Best practices

Memory and lifecycle

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Vibe Coding

Overview

Verbs

Scale

Removing

Inspecting state

Registering and discovering adapters

ViewModel

Notes

LLM generation

LLM streaming

RAG

Best practices

​Overview

​Verbs

​Scale

​Removing

​Inspecting state

​Registering and discovering adapters

​ViewModel

​Notes

​Related

LLM generation

LLM streaming

RAG

Best practices

Overview

Verbs

Scale

Removing

Inspecting state

Registering and discovering adapters

ViewModel

Notes

Related