cactus-react-native 0.1.1 → 0.1.3

package/src/lm.ts

@@ -0,0 +1,49 @@
+import { initLlama, LlamaContext } from './index'
+import type {
+  ContextParams,
+  CompletionParams,
+  CactusOAICompatibleMessage,
+  NativeCompletionResult,
+  EmbeddingParams,
+  NativeEmbeddingResult,
+} from './index'
+
+export class CactusLM {
+  private context: LlamaContext
+
+  private constructor(context: LlamaContext) {
+    this.context = context
+  }
+
+  static async init(
+    params: ContextParams,
+    onProgress?: (progress: number) => void,
+  ): Promise<CactusLM> {
+    const context = await initLlama(params, onProgress)
+    return new CactusLM(context)
+  }
+
+  async completion(
+    messages: CactusOAICompatibleMessage[],
+    params: CompletionParams = {},
+    callback?: (data: any) => void,
+  ): Promise<NativeCompletionResult> {
+    return this.context.completion({ messages, ...params }, callback)
+  }
+
+  async embedding(
+    text: string,
+    params?: EmbeddingParams,
+  ): Promise<NativeEmbeddingResult> {
+    return this.context.embedding(text, params)
+  }
+
+  async rewind(): Promise<void> {
+    // @ts-ignore
+    return this.context?.rewind()
+  }
+
+  async release(): Promise<void> {
+    return this.context.release()
+  }
+} 

package/src/tts.ts

@@ -0,0 +1,45 @@
+import {
+  LlamaContext,
+  initVocoder,
+  getFormattedAudioCompletion,
+  decodeAudioTokens,
+  releaseVocoder,
+} from './index'
+import type { NativeAudioDecodeResult } from './index'
+
+export class CactusTTS {
+  private context: LlamaContext
+
+  private constructor(context: LlamaContext) {
+    this.context = context
+  }
+
+  static async init(
+    context: LlamaContext,
+    vocoderModelPath: string,
+  ): Promise<CactusTTS> {
+    await initVocoder(context.id, vocoderModelPath)
+    return new CactusTTS(context)
+  }
+
+  async generate(
+    textToSpeak: string,
+    speakerJsonStr: string,
+  ): Promise<NativeAudioDecodeResult> {
+    const { formatted_prompt } = await getFormattedAudioCompletion(
+      this.context.id,
+      speakerJsonStr,
+      textToSpeak,
+    )
+    // This part is simplified. In a real scenario, the tokens from 
+    // the main model would be generated and passed to decodeAudioTokens.
+    // For now, we are assuming a direct path which may not be fully functional
+    // without the main model's token output for TTS.
+    const tokens = (await this.context.tokenize(formatted_prompt)).tokens
+    return decodeAudioTokens(this.context.id, tokens)
+  }
+
+  async release(): Promise<void> {
+    return releaseVocoder(this.context.id)
+  }
+} 

package/src/vlm.ts

@@ -0,0 +1,70 @@
+import {
+  initLlama,
+  initMultimodal,
+  multimodalCompletion,
+  LlamaContext,
+} from './index'
+import type {
+  ContextParams,
+  CompletionParams,
+  CactusOAICompatibleMessage,
+  NativeCompletionResult,
+} from './index'
+
+export type VLMContextParams = ContextParams & {
+  mmproj: string
+}
+
+export type VLMCompletionParams = Omit<CompletionParams, 'prompt'> & {
+  images?: string[]
+}
+
+export class CactusVLM {
+  private context: LlamaContext
+
+  private constructor(context: LlamaContext) {
+    this.context = context
+  }
+
+  static async init(
+    params: VLMContextParams,
+    onProgress?: (progress: number) => void,
+  ): Promise<CactusVLM> {
+    const context = await initLlama(params, onProgress)
+
+    // Explicitly disable GPU for the multimodal projector for stability.
+    await initMultimodal(context.id, params.mmproj, false)
+
+    return new CactusVLM(context)
+  }
+
+  async completion(
+    messages: CactusOAICompatibleMessage[],
+    params: VLMCompletionParams = {},
+    callback?: (data: any) => void,
+  ): Promise<NativeCompletionResult> {
+    if (params.images && params.images.length > 0) {
+      const formattedPrompt = await this.context.getFormattedChat(messages)
+      const prompt =
+        typeof formattedPrompt === 'string'
+          ? formattedPrompt
+          : formattedPrompt.prompt
+      return multimodalCompletion(
+        this.context.id,
+        prompt,
+        params.images,
+        { ...params, prompt, emit_partial_completion: !!callback },
+      )
+    }
+    return this.context.completion({ messages, ...params }, callback)
+  }
+
+  async rewind(): Promise<void> {
+    // @ts-ignore
+    return this.context?.rewind()
+  }
+
+  async release(): Promise<void> {
+    return this.context.release()
+  }
+} 

package/lib/commonjs/NativeCactus.js

@@ -1,10 +0,0 @@
-"use strict";
-
-Object.defineProperty(exports, "__esModule", {
-  value: true
-});
-exports.default = void 0;
-var _reactNative = require("react-native");
-// New TTS/Audio types
-var _default = exports.default = _reactNative.TurboModuleRegistry.get('Cactus');
-//# sourceMappingURL=NativeCactus.js.map

package/lib/commonjs/chat.js

@@ -1,37 +0,0 @@
-"use strict";
-
-Object.defineProperty(exports, "__esModule", {
-  value: true
-});
-exports.formatChat = formatChat;
-function formatChat(messages) {
-  const chat = [];
-  messages.forEach(currMsg => {
-    const role = currMsg.role || '';
-    let content = '';
-    if ('content' in currMsg) {
-      if (typeof currMsg.content === 'string') {
-        ;
-        ({
-          content
-        } = currMsg);
-      } else if (Array.isArray(currMsg.content)) {
-        currMsg.content.forEach(part => {
-          if ('text' in part) {
-            content += `${content ? '\n' : ''}${part.text}`;
-          }
-        });
-      } else {
-        throw new TypeError("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
-      }
-    } else {
-      throw new Error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
-    }
-    chat.push({
-      role,
-      content
-    });
-  });
-  return chat;
-}
-//# sourceMappingURL=chat.js.map