react-native-litert-lm 0.2.0 → 0.2.2

package/src/memoryTracker.ts

@@ -0,0 +1,268 @@
+/**
+ * Memory tracking utilities for LiteRT-LM using real native memory metrics.
+ *
+ * Records real memory usage from OS-level APIs via `getMemoryUsage()`,
+ * and stores snapshots in a native-backed ArrayBuffer allocated via
+ * `NitroModules.createNativeArrayBuffer()` (v0.34+) for zero-copy interop.
+ *
+ * @example
+ * ```typescript
+ * import { createMemoryTracker } from 'react-native-litert-lm';
+ *
+ * const tracker = createMemoryTracker(100);
+ *
+ * // Record a real snapshot (typically called internally after inference)
+ * tracker.record({
+ *   timestamp: Date.now(),
+ *   nativeHeapBytes: usage.nativeHeapBytes,
+ *   residentBytes: usage.residentBytes,
+ *   availableMemoryBytes: usage.availableMemoryBytes,
+ * });
+ *
+ * console.log(`Peak RSS: ${tracker.getPeakMemory()} bytes`);
+ * ```
+ */
+
+import { NitroModules } from "react-native-nitro-modules";
+
+/**
+ * A single memory usage snapshot with real data from OS APIs.
+ */
+export interface MemorySnapshot {
+  /** Unix timestamp in milliseconds */
+  timestamp: number;
+  /** Native heap allocated bytes (Debug.getNativeHeapAllocatedSize on Android, task_info on iOS) */
+  nativeHeapBytes: number;
+  /** Process resident set size (RSS) in bytes */
+  residentBytes: number;
+  /** Available system memory in bytes */
+  availableMemoryBytes: number;
+}
+
+/** Number of Float64 fields per snapshot */
+const FIELDS_PER_SNAPSHOT = 4;
+/** Bytes per Float64 value */
+const BYTES_PER_FIELD = Float64Array.BYTES_PER_ELEMENT; // 8
+
+/**
+ * Memory tracker that stores snapshots in a native-backed ArrayBuffer.
+ *
+ * Uses `NitroModules.createNativeArrayBuffer()` to allocate the backing
+ * buffer in native (C++) memory, ensuring zero-copy interop with native
+ * methods and keeping memory tracking data off the JS heap.
+ */
+export interface MemoryTracker {
+  /**
+   * Record a new memory snapshot.
+   * @param snapshot The memory usage data to record
+   * @returns true if recorded, false if buffer is full
+   */
+  record(snapshot: MemorySnapshot): boolean;
+
+  /**
+   * Get all recorded snapshots as structured objects.
+   */
+  getSnapshots(): MemorySnapshot[];
+
+  /**
+   * Get the number of recorded snapshots.
+   */
+  getSnapshotCount(): number;
+
+  /**
+   * Get the maximum number of snapshots this tracker can hold.
+   */
+  getCapacity(): number;
+
+  /**
+   * Get the peak resident set size across all snapshots.
+   */
+  getPeakMemory(): number;
+
+  /**
+   * Get the latest memory snapshot, or undefined if none recorded.
+   */
+  getLatestSnapshot(): MemorySnapshot | undefined;
+
+  /**
+   * Get the underlying native ArrayBuffer.
+   * This buffer is allocated via `NitroModules.createNativeArrayBuffer()`
+   * and lives in native memory, enabling zero-copy transfer to native methods.
+   */
+  getNativeBuffer(): ArrayBuffer;
+
+  /**
+   * Get the Float64Array view over the native buffer.
+   */
+  getView(): Float64Array;
+
+  /**
+   * Reset the tracker, clearing all recorded snapshots.
+   * The native buffer is preserved (not reallocated).
+   */
+  reset(): void;
+
+  /**
+   * Get a summary of memory usage statistics.
+   */
+  getSummary(): MemoryTrackerSummary;
+}
+
+/**
+ * Summary statistics from the memory tracker.
+ */
+export interface MemoryTrackerSummary {
+  /** Number of snapshots recorded */
+  snapshotCount: number;
+  /** Peak resident set size in bytes */
+  peakResidentBytes: number;
+  /** Average resident set size in bytes */
+  averageResidentBytes: number;
+  /** Latest resident set size in bytes */
+  currentResidentBytes: number;
+  /** Peak native heap allocated in bytes */
+  peakNativeHeapBytes: number;
+  /** Latest native heap allocated in bytes */
+  currentNativeHeapBytes: number;
+  /** RSS delta from first to last snapshot in bytes */
+  residentDeltaBytes: number;
+  /** Size of the native tracking buffer itself in bytes */
+  trackerBufferSizeBytes: number;
+}
+
+/**
+ * Create a new memory tracker backed by a native ArrayBuffer.
+ *
+ * @param maxSnapshots Maximum number of snapshots to store (default: 256)
+ * @returns A MemoryTracker instance
+ */
+export function createMemoryTracker(maxSnapshots: number = 256): MemoryTracker {
+  const bufferSize = maxSnapshots * FIELDS_PER_SNAPSHOT * BYTES_PER_FIELD;
+
+  // Use NitroModules.createNativeArrayBuffer for native-backed allocation.
+  const nativeBuffer = NitroModules.createNativeArrayBuffer(bufferSize);
+  const view = new Float64Array(nativeBuffer);
+
+  let currentIndex = 0;
+
+  return {
+    record(snapshot: MemorySnapshot): boolean {
+      if (currentIndex >= maxSnapshots) {
+        return false;
+      }
+
+      const offset = currentIndex * FIELDS_PER_SNAPSHOT;
+      view[offset] = snapshot.timestamp;
+      view[offset + 1] = snapshot.nativeHeapBytes;
+      view[offset + 2] = snapshot.residentBytes;
+      view[offset + 3] = snapshot.availableMemoryBytes;
+      currentIndex++;
+
+      return true;
+    },
+
+    getSnapshots(): MemorySnapshot[] {
+      const snapshots: MemorySnapshot[] = [];
+      for (let i = 0; i < currentIndex; i++) {
+        const offset = i * FIELDS_PER_SNAPSHOT;
+        snapshots.push({
+          timestamp: view[offset]!,
+          nativeHeapBytes: view[offset + 1]!,
+          residentBytes: view[offset + 2]!,
+          availableMemoryBytes: view[offset + 3]!,
+        });
+      }
+      return snapshots;
+    },
+
+    getSnapshotCount(): number {
+      return currentIndex;
+    },
+
+    getCapacity(): number {
+      return maxSnapshots;
+    },
+
+    getPeakMemory(): number {
+      let peak = 0;
+      for (let i = 0; i < currentIndex; i++) {
+        const rss = view[i * FIELDS_PER_SNAPSHOT + 2]!;
+        if (rss > peak) {
+          peak = rss;
+        }
+      }
+      return peak;
+    },
+
+    getLatestSnapshot(): MemorySnapshot | undefined {
+      if (currentIndex === 0) return undefined;
+      const offset = (currentIndex - 1) * FIELDS_PER_SNAPSHOT;
+      return {
+        timestamp: view[offset]!,
+        nativeHeapBytes: view[offset + 1]!,
+        residentBytes: view[offset + 2]!,
+        availableMemoryBytes: view[offset + 3]!,
+      };
+    },
+
+    getNativeBuffer(): ArrayBuffer {
+      return nativeBuffer;
+    },
+
+    getView(): Float64Array {
+      return view;
+    },
+
+    reset(): void {
+      view.fill(0);
+      currentIndex = 0;
+    },
+
+    getSummary(): MemoryTrackerSummary {
+      let peakRss = 0;
+      let peakHeap = 0;
+      let sumRss = 0;
+      let firstRss = 0;
+      let lastRss = 0;
+      let lastHeap = 0;
+
+      for (let i = 0; i < currentIndex; i++) {
+        const offset = i * FIELDS_PER_SNAPSHOT;
+        const heap = view[offset + 1]!;
+        const rss = view[offset + 2]!;
+
+        if (rss > peakRss) peakRss = rss;
+        if (heap > peakHeap) peakHeap = heap;
+        sumRss += rss;
+        if (i === 0) firstRss = rss;
+        if (i === currentIndex - 1) {
+          lastRss = rss;
+          lastHeap = heap;
+        }
+      }
+
+      return {
+        snapshotCount: currentIndex,
+        peakResidentBytes: peakRss,
+        averageResidentBytes: currentIndex > 0 ? sumRss / currentIndex : 0,
+        currentResidentBytes: lastRss,
+        peakNativeHeapBytes: peakHeap,
+        currentNativeHeapBytes: lastHeap,
+        residentDeltaBytes: lastRss - firstRss,
+        trackerBufferSizeBytes: bufferSize,
+      };
+    },
+  };
+}
+
+/**
+ * Create a native ArrayBuffer for efficient data transfer.
+ *
+ * A convenience wrapper around `NitroModules.createNativeArrayBuffer()`.
+ *
+ * @param size Size in bytes
+ * @returns A native-backed ArrayBuffer
+ */
+export function createNativeBuffer(size: number): ArrayBuffer {
+  return NitroModules.createNativeArrayBuffer(size);
+}

package/src/modelFactory.ts

@@ -0,0 +1,120 @@
+import { NitroModules } from "react-native-nitro-modules";
+import { LiteRTLM, LLMConfig } from "./specs/LiteRTLM.nitro";
+import { createMemoryTracker, MemoryTracker } from "./memoryTracker";
+
+/**
+ * Creates a new LiteRT-LM inference engine instance.
+ *
+ * Optionally creates a native-backed memory tracker using
+ * `NitroModules.createNativeArrayBuffer()` (v0.34+) for efficient
+ * zero-copy memory usage tracking.
+ *
+ * @param options.enableMemoryTracking Enable automatic memory tracking (default: false)
+ * @param options.maxMemorySnapshots Maximum number of memory snapshots to store (default: 256)
+ */
+export function createLLM(options?: {
+  enableMemoryTracking?: boolean;
+  maxMemorySnapshots?: number;
+}): LiteRTLM & { memoryTracker?: MemoryTracker } {
+  const native = NitroModules.createHybridObject<LiteRTLM>("LiteRTLM");
+
+  const enableTracking = options?.enableMemoryTracking ?? false;
+  const tracker = enableTracking
+    ? createMemoryTracker(options?.maxMemorySnapshots ?? 256)
+    : undefined;
+
+  /**
+   * Record a real memory snapshot using OS-level APIs via getMemoryUsage().
+   */
+  const recordMemorySnapshot = () => {
+    if (!tracker) return;
+    try {
+      const usage = native.getMemoryUsage();
+      tracker.record({
+        timestamp: Date.now(),
+        nativeHeapBytes: usage.nativeHeapBytes,
+        residentBytes: usage.residentBytes,
+        availableMemoryBytes: usage.availableMemoryBytes,
+      });
+    } catch {
+      // Ignore errors during memory tracking - it's non-critical
+    }
+  };
+
+  return {
+    ...native,
+    memoryTracker: tracker,
+    loadModel: async (pathOrUrl: string, config?: LLMConfig) => {
+      let modelPath = pathOrUrl;
+
+      // Check if it's a URL
+      if (pathOrUrl.startsWith("http://") || pathOrUrl.startsWith("https://")) {
+        // Extract filename from URL
+        const fileName = pathOrUrl.split("/").pop();
+        if (!fileName) {
+          throw new Error(`Invalid model URL: ${pathOrUrl}`);
+        }
+
+        console.log(`Checking model at ${pathOrUrl}...`);
+        modelPath = await native.downloadModel(
+          pathOrUrl,
+          fileName,
+          (progress) => {
+            console.log(`Download progress: ${progress}`);
+          },
+        );
+        console.log(`Model downloaded to: ${modelPath}`);
+      }
+
+      const result = await native.loadModel(modelPath, config);
+
+      // Record initial memory snapshot after model load
+      if (tracker) {
+        tracker.reset();
+        recordMemorySnapshot();
+      }
+
+      return result;
+    },
+    sendMessage: async (...args: Parameters<typeof native.sendMessage>) => {
+      const result = await native.sendMessage(...args);
+      recordMemorySnapshot();
+      return result;
+    },
+    sendMessageAsync: (...args: Parameters<typeof native.sendMessageAsync>) => {
+      const [message, onToken] = args;
+      native.sendMessageAsync(message, (token, done) => {
+        onToken(token, done);
+        if (done) {
+          recordMemorySnapshot();
+        }
+      });
+    },
+    sendMessageWithImage: async (
+      ...args: Parameters<typeof native.sendMessageWithImage>
+    ) => {
+      const result = await native.sendMessageWithImage(...args);
+      recordMemorySnapshot();
+      return result;
+    },
+    sendMessageWithAudio: async (
+      ...args: Parameters<typeof native.sendMessageWithAudio>
+    ) => {
+      const result = await native.sendMessageWithAudio(...args);
+      recordMemorySnapshot();
+      return result;
+    },
+    getHistory: native.getHistory.bind(native),
+    resetConversation: () => {
+      native.resetConversation();
+      // KV cache is cleared on reset, record the drop
+      recordMemorySnapshot();
+    },
+    isReady: native.isReady.bind(native),
+    getStats: native.getStats.bind(native),
+    getMemoryUsage: native.getMemoryUsage.bind(native),
+    close: native.close.bind(native),
+    downloadModel: native.downloadModel.bind(native),
+    deleteModel: native.deleteModel.bind(native),
+  };
+}

package/src/specs/LiteRTLM.nitro.ts

@@ -21,6 +21,13 @@ export type Role = "user" | "model" | "system";
  * Configuration options for loading an LLM.
  */
 export interface LLMConfig {
+  /**
+   * System prompt to set the model's behavior.
+   * This is prepended to the conversation to guide model responses.
+   * @example "You are a helpful coding assistant."
+   */
+  systemPrompt?: string;
+
   /**
    * Primary compute backend for text generation.
    * - 'cpu': CPU inference (slower but always available)
@@ -92,6 +99,21 @@ export interface GenerationStats {
   tokensPerSecond: number;
 }
 
+/**
+ * Real memory usage statistics from the native runtime.
+ * Measured from OS-level APIs, not estimated.
+ */
+export interface MemoryUsage {
+  /** Native heap allocated bytes (Debug.getNativeHeapAllocatedSize on Android, malloc_size on iOS) */
+  nativeHeapBytes: number;
+  /** Total process resident set size (RSS) in bytes */
+  residentBytes: number;
+  /** Available system memory in bytes */
+  availableMemoryBytes: number;
+  /** Whether the system considers memory low */
+  isLowMemory: boolean;
+}
+
 /**
  * LiteRT-LM: High-performance LLM inference engine.
  * Supports Gemma 3n, Phi-4, Qwen, and other .litertlm models.
@@ -138,6 +160,25 @@ export interface LiteRTLM extends HybridObject<{
    */
   sendMessageWithImage(message: string, imagePath: string): Promise<string>;
 
+  /**
+   * Download a model file from a URL.
+   * @param url URL to download from.
+   * @param fileName Filename to save as (in app's files directory).
+   * @param onProgress Callback for download progress (0.0 - 1.0).
+   * @returns Absolute path to the downloaded file.
+   */
+  downloadModel(
+    url: string,
+    fileName: string,
+    onProgress?: (progress: number) => void,
+  ): Promise<string>;
+
+  /**
+   * Delete a downloaded model file.
+   * @param fileName Filename to delete (in app's files directory).
+   */
+  deleteModel(fileName: string): Promise<void>;
+
   /**
    * Send a text message with audio (multimodal).
    * @param message User message text.
@@ -178,6 +219,12 @@ export interface LiteRTLM extends HybridObject<{
    */
   getStats(): GenerationStats;
 
+  /**
+   * Get real memory usage from the native runtime.
+   * Uses OS-level APIs to report actual memory consumption.
+   */
+  getMemoryUsage(): MemoryUsage;
+
   /**
    * Release all native resources.
    * Call this when done with the LLM instance.

package/src/templates.ts

@@ -0,0 +1,105 @@
+/**
+ * Prompt template utilities for different LLM families.
+ *
+ * LiteRT-LM's Conversation API may handle templates internally for some models,
+ * but these utilities give developers explicit control for custom workflows
+ * or when using models with different template formats.
+ *
+ * @example
+ * ```typescript
+ * import { applyGemmaTemplate, ChatMessage } from 'react-native-litert-lm';
+ *
+ * const history: ChatMessage[] = [
+ *   { role: 'user', content: 'What is React Native?' },
+ *   { role: 'model', content: 'React Native is a framework for building...' },
+ *   { role: 'user', content: 'How do I use hooks?' }
+ * ];
+ *
+ * const prompt = applyGemmaTemplate(history, 'You are a helpful coding assistant.');
+ * ```
+ */
+
+/**
+ * A message in a conversation.
+ */
+export type ChatMessage = {
+  role: "user" | "model" | "system";
+  content: string;
+};
+
+/**
+ * Apply Gemma chat template (Gemma 2, Gemma 3, Gemma 3n).
+ *
+ * @param history Array of previous messages
+ * @param systemPrompt Optional system prompt
+ * @returns Formatted prompt string
+ */
+export function applyGemmaTemplate(
+  history: ChatMessage[],
+  systemPrompt?: string,
+): string {
+  let result = "";
+
+  if (systemPrompt) {
+    result += `<start_of_turn>system\n${systemPrompt}<end_of_turn>\n`;
+  }
+
+  for (const m of history) {
+    result += `<start_of_turn>${m.role}\n${m.content}<end_of_turn>\n`;
+  }
+
+  result += "<start_of_turn>model\n";
+  return result;
+}
+
+/**
+ * Apply Phi chat template (Phi-3, Phi-4).
+ *
+ * @param history Array of previous messages
+ * @param systemPrompt Optional system prompt
+ * @returns Formatted prompt string
+ */
+export function applyPhiTemplate(
+  history: ChatMessage[],
+  systemPrompt?: string,
+): string {
+  let result = "";
+
+  if (systemPrompt) {
+    result += `<|system|>\n${systemPrompt}<|end|>\n`;
+  }
+
+  for (const m of history) {
+    const role = m.role === "model" ? "assistant" : m.role;
+    result += `<|${role}|>\n${m.content}<|end|>\n`;
+  }
+
+  result += "<|assistant|>\n";
+  return result;
+}
+
+/**
+ * Apply Llama 3 chat template.
+ *
+ * @param history Array of previous messages
+ * @param systemPrompt Optional system prompt
+ * @returns Formatted prompt string
+ */
+export function applyLlamaTemplate(
+  history: ChatMessage[],
+  systemPrompt?: string,
+): string {
+  let result = "<|begin_of_text|>";
+
+  if (systemPrompt) {
+    result += `<|start_header_id|>system<|end_header_id|>\n\n${systemPrompt}<|eot_id|>`;
+  }
+
+  for (const m of history) {
+    const role = m.role === "model" ? "assistant" : m.role;
+    result += `<|start_header_id|>${role}<|end_header_id|>\n\n${m.content}<|eot_id|>`;
+  }
+
+  result += "<|start_header_id|>assistant<|end_header_id|>\n\n";
+  return result;
+}