Native Modules

Overview

Off Grid uses native modules to bridge React Native (JavaScript) to high-performance C++ and platform-specific inference engines.

Module	Platform	Language	Purpose
llama.rn	Android + iOS	C++	Text + vision LLM inference (llama.cpp)
whisper.rn	Android + iOS	C++	Speech-to-text transcription (whisper.cpp)
LocalDreamModule	Android	Kotlin + C++	Stable Diffusion (MNN/QNN backends)
CoreMLDiffusionModule	iOS	Swift	Stable Diffusion (Core ML + ANE)
PDFExtractorModule	Android	Kotlin	PDF text extraction (PdfiumAndroid)
PDFExtractorModule	iOS	Swift	PDF text extraction (PDFKit)
DownloadManagerModule	Android	Kotlin	Background downloads (DownloadManager)

llama.rn (Text + Vision Inference)

Upstream: github.com/mybigday/llama.rn
Wraps: llama.cpp (Georgi Gerganov’s GGUF inference engine)

Platform Support

Feature	Android	iOS
Text Generation	✅ llama.cpp (OpenCL GPU)	✅ llama.cpp (Metal GPU)
Vision (multimodal)	✅ mmproj CLIP encoder	✅ mmproj CLIP encoder
Tool Calling	✅ Model-dependent	✅ Model-dependent
Flash Attention	✅ Auto-disabled with GPU	✅ Always available
KV Cache Quantization	✅ f16/q8_0/q4_0	✅ f16/q8_0/q4_0

TypeScript API

import { LlamaContext } from 'llama.rn';

class LLMService {
  private context: LlamaContext | null = null;

  async loadModel(modelPath: string, mmProjPath?: string): Promise<void> {
    const { settings } = useAppStore.getState();

    // Build model params
    const params = {
      model: modelPath,
      use_mlock: true,
      n_ctx: settings.contextLength,
      n_threads: settings.nThreads,
      n_batch: settings.nBatch,
      n_gpu_layers: settings.enableGpu ? settings.gpuLayers : 0,
      flash_attn: settings.flashAttn,
      cache_type_k: settings.cacheType,
      cache_type_v: settings.cacheType,
    };

    // Initialize context
    this.context = await initContextWithFallback(params, settings.contextLength, params.n_gpu_layers);

    // Load multimodal (vision) if mmproj provided
    if (mmProjPath) {
      await this.initializeMultimodal(mmProjPath);
    }
  }

  async generateResponse(
    messages: Message[],
    onStream?: (token: string) => void,
    onComplete?: (fullResponse: string) => void
  ): Promise<string> {
    if (!this.context) throw new Error('No model loaded');

    const oaiMessages = this.convertToOAIMessages(messages);
    let fullResponse = '';

    await this.context.completion(
      {
        messages: oaiMessages,
        temperature: settings.temperature,
        top_p: settings.topP,
        max_tokens: settings.maxTokens,
        repeat_penalty: settings.repeatPenalty,
      },
      (data) => {
        if (!data.token) return;
        fullResponse += data.token;
        onStream?.(data.token);
      }
    );

    onComplete?.(fullResponse);
    return fullResponse;
  }
}

Vision Inference (Multimodal)

Vision models require a mmproj file (multimodal projector) containing the CLIP vision encoder weights.

async initializeMultimodal(mmProjPath: string): Promise<boolean> {
  if (!this.context) return false;

  const devInfo = useAppStore.getState().deviceInfo;
  const mem = devInfo?.totalMemory ?? 0;

  // Disable CLIP GPU on low-RAM devices (≤4GB)
  const useGpuForClip = Platform.OS === 'ios' && !devInfo?.isEmulator && mem > 4 * BYTES_PER_GB;

  const { initialized, support } = await initMultimodal(
    this.context,
    mmProjPath,
    useGpuForClip
  );

  this.multimodalInitialized = initialized;
  this.multimodalSupport = support;
  return initialized;
}

Tool Calling Detection

llama.rn introspects the model’s jinja chat template to detect tool calling support.

private detectToolCallingSupport(): void {
  if (!this.context) {
    this.toolCallingSupported = false;
    return;
  }

  const model = (this.context as any)?.model;
  const jinja = model?.chatTemplates?.jinja;

  this.toolCallingSupported = !!(
    jinja?.defaultCaps?.toolCalls ||
    jinja?.toolUse ||
    jinja?.toolUseCaps?.toolCalls
  );
}

GPU Acceleration

Android: OpenCL (Adreno GPUs)
iOS: Metal (Apple GPUs)

On devices with ≤4GB RAM, GPU layers are forced to 0 to prevent abort() crashes during Metal/OpenCL buffer allocation (see src/services/llmHelpers.ts:getGpuLayersForDevice).

whisper.rn (Speech Transcription)

Upstream: github.com/mybigday/whisper.rn
Wraps: whisper.cpp

Features

Real-time audio recording and transcription
Partial results (word-by-word streaming)
Multiple model sizes: Tiny (39MB), Base (74MB), Small (244MB)
No network required (100% on-device)

TypeScript API

import { whisperContext } from 'whisper.rn';

// Load model
const ctx = await whisperContext.initContext({
  filePath: '/path/to/ggml-tiny.bin',
});

// Transcribe audio file
const result = await ctx.transcribe({
  audioPath: '/path/to/audio.wav',
  language: 'en',
  maxLen: 1,
  tokenTimestamps: true,
});

console.log(result.result); // "Hello, world!"

LocalDreamModule (Android Stable Diffusion)

Path: android/app/src/main/java/ai/offgridmobile/localdream/LocalDreamModule.kt:32
Wraps: local-dream C++ library (MNN/QNN backends)

Architecture

LocalDreamModule spawns a subprocess running an HTTP server on localhost:18081. The TypeScript layer sends POST requests to /generate, and the server responds with Server-Sent Events (SSE) for progress updates.

TypeScript
    ↓ POST /generate
HTTP Server (localhost:18081)
    ↓
local-dream C++ (subprocess)
    ↓ MNN or QNN
CPU or NPU inference

Backend Selection

@ReactMethod
fun loadModel(params: ReadableMap, promise: Promise) {
  val modelPath = params.getString("modelPath") ?: throw IllegalArgumentException("modelPath is required")
  val rawModelDir = File(modelPath)
  val requestedBackend = params.getString("backend") ?: "auto"

  // Resolve model directory for CPU (MNN) and/or NPU (QNN)
  val cpuModelDir = resolveModelDir(rawModelDir, isCpu = true)  // Looks for unet.mnn
  val qnnModelDir = resolveModelDir(rawModelDir, isCpu = false) // Looks for unet.bin
  val npuSupported = isNpuSupportedInternal()

  val (backend, modelDir) = when (requestedBackend.lowercase()) {
    "mnn", "cpu" -> if (cpuModelDir != null) "mnn" to cpuModelDir else null
    "qnn", "npu" -> if (qnnModelDir != null) "qnn" to qnnModelDir else null
    "auto" -> {
      when {
        qnnModelDir != null && npuSupported -> "qnn" to qnnModelDir
        cpuModelDir != null -> "mnn" to cpuModelDir
        qnnModelDir != null -> "qnn" to qnnModelDir
        else -> null
      }
    }
    else -> null
  } ?: throw IllegalArgumentException("Model files not found")

  // Start server with selected backend
  tryStartServer(modelPath, modelDir, backend, isCpu = backend == "mnn")
}

Image Generation Flow

@ReactMethod
fun generateImage(params: ReadableMap, promise: Promise) {
  val body = JSONObject().apply {
    put("prompt", params.getString("prompt") ?: "")
    put("negative_prompt", params.getString("negativePrompt") ?: "")
    put("steps", params.getInt("steps") ?: 20)
    put("cfg", params.getDouble("guidanceScale") ?: 7.5)
    put("seed", params.getInt("seed") ?: (Math.random() * 2147483647).toInt())
    put("width", params.getInt("width") ?: 512)
    put("height", params.getInt("height") ?: 512)
    put("scheduler", "dpm")
    put("show_diffusion_process", true)
    put("show_diffusion_stride", params.getInt("previewInterval") ?: 2)
  }

  val url = URL("http://127.0.0.1:$SERVER_PORT/generate")
  val connection = (url.openConnection() as HttpURLConnection).apply {
    requestMethod = "POST"
    doOutput = true
    setRequestProperty("Content-Type", "application/json")
    setRequestProperty("Accept", "text/event-stream")
  }

  // Send request
  OutputStreamWriter(connection.outputStream).use { it.write(body.toString()) }

  // Parse SSE stream
  BufferedReader(InputStreamReader(connection.inputStream)).use { reader ->
    var line: String?
    while (reader.readLine().also { line = it } != null) {
      if (line!!.startsWith("data: ")) {
        val data = JSONObject(line!!.substring(6))
        when (data.optString("type")) {
          "progress" -> {
            val step = data.getInt("step")
            val totalSteps = data.getInt("total_steps")
            val previewBase64 = data.optString("image", "")

            // Emit progress event
            sendEvent("LocalDreamProgress", Arguments.createMap().apply {
              putInt("step", step)
              putInt("totalSteps", totalSteps)
            })

            // Save preview image if present
            if (previewBase64.isNotEmpty()) {
              saveRgbToPng(previewBase64, width, height, previewPath)
            }
          }
          "complete" -> {
            val imageBase64 = data.getString("image")
            saveRgbToPng(imageBase64, width, height, finalPath)
            promise.resolve(Arguments.createMap().apply {
              putString("imagePath", finalPath)
            })
          }
        }
      }
    }
  }
}

RGB → PNG Conversion

local-dream returns images as base64-encoded raw RGB bytes. LocalDreamModule decodes and converts to PNG:

internal fun saveRgbToPng(base64Rgb: String, width: Int, height: Int, outputPath: String) {
  val rgbBytes = Base64.decode(base64Rgb, Base64.DEFAULT)
  val expectedSize = width * height * 3

  if (rgbBytes.size != expectedSize) {
    throw IllegalArgumentException(
      "RGB data size ${rgbBytes.size} doesn't match expected $expectedSize (${width}x${height}x3)"
    )
  }

  val bitmap = Bitmap.createBitmap(width, height, Bitmap.Config.ARGB_8888)
  val pixels = IntArray(width * height)

  for (i in 0 until width * height) {
    val idx = i * 3
    val r = rgbBytes[idx].toInt() and 0xFF
    val g = rgbBytes[idx + 1].toInt() and 0xFF
    val b = rgbBytes[idx + 2].toInt() and 0xFF
    pixels[i] = (0xFF shl 24) or (r shl 16) or (g shl 8) or b
  }

  bitmap.setPixels(pixels, 0, width, 0, 0, width, height)

  File(outputPath).parentFile?.mkdirs()
  FileOutputStream(outputPath).use { out ->
    bitmap.compress(Bitmap.CompressFormat.PNG, 100, out)
  }
  bitmap.recycle()
}

CoreMLDiffusionModule (iOS Stable Diffusion)

Path: ios/CoreMLDiffusionModule.swift:11
Wraps: Apple’s ml-stable-diffusion Swift library

Features

Neural Engine (ANE) + CPU acceleration
Supports SD 1.5, SD 2.1, and SDXL models
DPM-Solver multistep scheduler
Palettized (6-bit) and full-precision (fp16) models

Model Detection (SD vs SDXL)

SDXL models use a different text encoder (TextEncoder2.mlmodelc). CoreMLDiffusionModule auto-detects and loads the correct pipeline:

private func isXLModel(at url: URL) -> Bool {
  let te2 = url.appendingPathComponent("TextEncoder2.mlmodelc")
  return FileManager.default.fileExists(atPath: te2.path)
}

@objc func loadModel(_ params: NSDictionary,
                     resolver resolve: @escaping RCTPromiseResolveBlock,
                     rejecter reject: @escaping RCTPromiseRejectBlock) {
  guard let modelPath = params["modelPath"] as? String else {
    reject("ERR_INVALID_PARAMS", "modelPath is required", nil)
    return
  }

  let url = URL(fileURLWithPath: modelPath)
  let config = MLModelConfiguration()
  config.computeUnits = .cpuAndNeuralEngine

  let pipe: StableDiffusionPipelineProtocol

  if isXLModel(at: url) {
    pipe = try StableDiffusionXLPipeline(
      resourcesAt: url,
      configuration: config,
      reduceMemory: true
    )
  } else {
    pipe = try StableDiffusionPipeline(
      resourcesAt: url,
      controlNet: [],
      configuration: config,
      reduceMemory: true
    )
  }

  try pipe.loadResources()
  self.pipeline = pipe
  resolve(true)
}

Image Generation

@objc func generateImage(_ params: NSDictionary,
                         resolver resolve: @escaping RCTPromiseResolveBlock,
                         rejecter reject: @escaping RCTPromiseRejectBlock) {
  guard let pipe = pipeline else {
    reject("ERR_NO_MODEL", "No model loaded", nil)
    return
  }

  let prompt = params["prompt"] as? String ?? ""
  let negativePrompt = params["negativePrompt"] as? String ?? ""
  let steps = params["steps"] as? Int ?? 20
  let guidanceScale = params["guidanceScale"] as? Double ?? 7.5
  let seed = params["seed"] as? UInt32 ?? UInt32.random(in: 0..<UInt32.max)

  var config = PipelineConfiguration(prompt: prompt)
  config.negativePrompt = negativePrompt
  config.stepCount = steps
  config.guidanceScale = Float(guidanceScale)
  config.seed = seed

  let images = try pipe.generateImages(configuration: config) { progress in
    // Emit progress event
    self.sendEvent(withName: "LocalDreamProgress", body: [
      "step": progress.step,
      "totalSteps": progress.stepCount,
      "progress": Double(progress.step) / Double(progress.stepCount)
    ])
    return !self.cancelRequested // Continue if not cancelled
  }

  guard let cgImage = images.compactMap({ $0 }).first else {
    reject("ERR_NO_IMAGE", "Pipeline produced no image", nil)
    return
  }

  // Save to documents directory
  let imageId = UUID().uuidString
  let docsDir = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask).first!
  let generatedDir = docsDir.appendingPathComponent("generated_images")
  try FileManager.default.createDirectory(at: generatedDir, withIntermediateDirectories: true)

  let imagePath = generatedDir.appendingPathComponent("\(imageId).png")
  let uiImage = UIImage(cgImage: cgImage)
  let pngData = uiImage.pngData()!
  try pngData.write(to: imagePath)

  resolve([
    "id": imageId,
    "imagePath": imagePath.path,
    "width": cgImage.width,
    "height": cgImage.height,
    "seed": seed
  ])
}

PDFExtractorModule

Android (PdfiumAndroid)

Path: android/app/src/main/java/ai/offgridmobile/pdf/PDFExtractorModule.kt:11

@ReactMethod
fun extractText(filePath: String, maxChars: Double, promise: Promise) {
  Thread {
    val file = File(filePath)
    if (!file.exists()) {
      promise.reject("PDF_ERROR", "File not found: $filePath")
      return@Thread
    }

    val limit = maxChars.toInt()
    val core = PdfiumCore(reactApplicationContext)
    val fd = ParcelFileDescriptor.open(file, ParcelFileDescriptor.MODE_READ_ONLY)
    val doc = core.newDocument(fd)
    val pageCount = doc.getPageCount()
    val sb = StringBuilder()

    for (i in 0 until pageCount) {
      val page = doc.openPage(i)
      val textPage = page.openTextPage()
      val charCount = textPage.textPageCountChars()

      if (charCount > 0) {
        val text = textPage.textPageGetText(0, charCount)
        if (text != null) sb.append(text).append("\n\n")
      }

      textPage.close()
      page.close()

      if (sb.length >= limit) {
        sb.setLength(limit)
        sb.append("\n\n... [Extracted ${i + 1} of $pageCount pages]")
        break
      }
    }

    doc.close()
    fd.close()
    promise.resolve(sb.toString())
  }.start()
}

iOS (PDFKit)

Path: ios/OffgridMobile/PDFExtractor/PDFExtractorModule.swift:4

@objc
func extractText(_ filePath: String, maxChars: Double,
                 resolver resolve: @escaping RCTPromiseResolveBlock,
                 rejecter reject: @escaping RCTPromiseRejectBlock) {
  DispatchQueue.global(qos: .userInitiated).async {
    guard let url = URL(string: filePath) ?? URL(fileURLWithPath: filePath) as URL?,
          let document = PDFDocument(url: url) else {
      reject("PDF_ERROR", "Could not open PDF file", nil)
      return
    }

    let limit = Int(maxChars)
    var fullText = ""

    for pageIndex in 0..<document.pageCount {
      if let page = document.page(at: pageIndex), let pageText = page.string {
        fullText += pageText
        if pageIndex < document.pageCount - 1 {
          fullText += "\n\n"
        }
      }

      if fullText.count >= limit {
        fullText = String(fullText.prefix(limit))
        fullText += "\n\n... [Extracted \(pageIndex + 1) of \(document.pageCount) pages]"
        break
      }
    }

    resolve(fullText)
  }
}

DownloadManagerModule (Android)

Path: android/app/src/main/java/ai/offgridmobile/download/DownloadManagerModule.kt:17

Features

Background downloads via Android’s native DownloadManager
Progress polling (500ms intervals)
Persistent download tracking via SharedPreferences
Automatic cleanup of completed/stale downloads
System notifications for download progress

Key Methods

@ReactMethod
fun startDownload(params: ReadableMap, promise: Promise) {
  val url = params.getString("url") ?: throw IllegalArgumentException("URL is required")
  val fileName = params.getString("fileName") ?: throw IllegalArgumentException("fileName is required")
  val title = params.getString("title") ?: fileName
  val modelId = params.getString("modelId") ?: ""

  val request = DownloadManager.Request(Uri.parse(url))
    .setTitle(title)
    .setNotificationVisibility(DownloadManager.Request.VISIBILITY_VISIBLE_NOTIFY_COMPLETED)
    .setDestinationInExternalFilesDir(
      reactApplicationContext,
      Environment.DIRECTORY_DOWNLOADS,
      fileName
    )
    .setAllowedOverMetered(true)
    .setAllowedOverRoaming(true)

  val downloadId = downloadManager.enqueue(request)

  // Persist download info to SharedPreferences
  val downloadInfo = JSONObject().apply {
    put("downloadId", downloadId)
    put("url", url)
    put("fileName", fileName)
    put("modelId", modelId)
    put("title", title)
    put("status", "pending")
    put("startedAt", System.currentTimeMillis())
  }
  persistDownload(downloadId, downloadInfo)

  promise.resolve(Arguments.createMap().apply {
    putDouble("downloadId", downloadId.toDouble())
    putString("fileName", fileName)
    putString("modelId", modelId)
  })
}

Race Condition Fix: On slow emulators, download completion notification could arrive before React Native received the DownloadComplete event. Fixed by tracking completedEventSent separately:

if (status == "completed" && !completedEventSent) {
  sendEvent("DownloadComplete", eventParams)
  updateDownloadStatus(downloadId, "completed", localUri)
}

Architecture

Platform Details

Performance

Overview

llama.rn (Text + Vision Inference)

Platform Support

TypeScript API

Vision Inference (Multimodal)

Tool Calling Detection

GPU Acceleration

whisper.rn (Speech Transcription)

Features

TypeScript API

LocalDreamModule (Android Stable Diffusion)

Architecture

Backend Selection

Image Generation Flow

RGB → PNG Conversion

CoreMLDiffusionModule (iOS Stable Diffusion)

Features

Model Detection (SD vs SDXL)

Image Generation

PDFExtractorModule

Android (PdfiumAndroid)

iOS (PDFKit)

DownloadManagerModule (Android)

Features

Key Methods

Build docs developers (and LLMs) love

Architecture

Platform Details

Performance

​Overview

​llama.rn (Text + Vision Inference)

​Platform Support

​TypeScript API

​Vision Inference (Multimodal)

​Tool Calling Detection

​GPU Acceleration

​whisper.rn (Speech Transcription)

​Features

​TypeScript API

​LocalDreamModule (Android Stable Diffusion)

​Architecture

​Backend Selection

​Image Generation Flow

​RGB → PNG Conversion

​CoreMLDiffusionModule (iOS Stable Diffusion)

​Features

​Model Detection (SD vs SDXL)

​Image Generation

​PDFExtractorModule

​Android (PdfiumAndroid)

​iOS (PDFKit)

​DownloadManagerModule (Android)

​Features

​Key Methods

Build docs developers (and LLMs) love

Overview

llama.rn (Text + Vision Inference)

Platform Support

TypeScript API

Vision Inference (Multimodal)

Tool Calling Detection

GPU Acceleration

whisper.rn (Speech Transcription)

Features

TypeScript API

LocalDreamModule (Android Stable Diffusion)

Architecture

Backend Selection

Image Generation Flow

RGB → PNG Conversion

CoreMLDiffusionModule (iOS Stable Diffusion)

Features

Model Detection (SD vs SDXL)

Image Generation

PDFExtractorModule

Android (PdfiumAndroid)

iOS (PDFKit)

DownloadManagerModule (Android)

Features

Key Methods