llama-cpp-pro 0.2.2

package/dist/esm/definitions.js

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=definitions.js.map

package/dist/esm/definitions.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"definitions.js","sourceRoot":"","sources":["../../src/definitions.ts"],"names":[],"mappings":"","sourcesContent":["// Native parameter types that match llama.rn exactly\nexport interface NativeEmbeddingParams {\n  embd_normalize?: number;\n}\n\nexport interface NativeContextParams {\n  model: string;\n  /**\n   * Chat template to override the default one from the model.\n   */\n  chat_template?: string;\n\n  is_model_asset?: boolean;\n  use_progress_callback?: boolean;\n\n  n_ctx?: number;\n  n_batch?: number;\n  n_ubatch?: number;\n\n  n_threads?: number;\n\n  /**\n   * Path to draft model for speculative decoding (mobile optimization)\n   */\n  draft_model?: string;\n  /**\n   * Number of tokens to predict speculatively (default: 3 for mobile)\n   */\n  speculative_samples?: number;\n  /**\n   * Enable mobile-optimized speculative decoding\n   */\n  mobile_speculative?: boolean;\n\n  /**\n   * Number of layers to offload to GPU (iOS, Android, Desktop)\n   */\n  n_gpu_layers?: number;\n  /**\n   * Skip GPU devices (iOS, Android, Desktop force-CPU mode)\n   */\n  no_gpu_devices?: boolean;\n\n  /**\n   * Enable flash attention, only recommended in GPU device (Experimental in llama.cpp)\n   */\n  flash_attn?: boolean;\n\n  /**\n   * KV cache data type for the K (Experimental in llama.cpp)\n   */\n  cache_type_k?: string;\n  /**\n   * KV cache data type for the V (Experimental in llama.cpp)\n   */\n  cache_type_v?: string;\n\n  use_mlock?: boolean;\n  use_mmap?: boolean;\n  vocab_only?: boolean;\n\n  /**\n   * Single LoRA adapter path\n   */\n  lora?: string;\n  /**\n   * Single LoRA adapter scale\n   */\n  lora_scaled?: number;\n  /**\n   * LoRA adapter list\n   */\n  lora_list?: Array<{ path: string; scaled?: number }>;\n\n  rope_freq_base?: number;\n  rope_freq_scale?: number;\n\n  pooling_type?: number;\n\n  /**\n   * Enable context shifting to handle prompts larger than context size\n   */\n  ctx_shift?: boolean;\n\n  /**\n   * Use a unified buffer across the input sequences when computing the attention.\n   * Try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix.\n   */\n  kv_unified?: boolean;\n\n  /**\n   * Use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)\n   */\n  swa_full?: boolean;\n\n  /**\n   * Number of layers to keep MoE weights on CPU\n   */\n  n_cpu_moe?: number;\n\n  // Embedding params\n  embedding?: boolean;\n  embd_normalize?: number;\n}\n\nexport interface NativeCompletionParams {\n  prompt: string;\n  n_threads?: number;\n  /**\n   * Enable Jinja. Default: true if supported by the model\n   */\n  jinja?: boolean;\n  /**\n   * JSON schema for convert to grammar for structured JSON output.\n   * It will be override by grammar if both are set.\n   */\n  json_schema?: string;\n  /**\n   * Set grammar for grammar-based sampling (GBNF format). Default: no grammar\n   * This will override json_schema if both are provided.\n   */\n  grammar?: string;\n  /**\n   * Lazy grammar sampling, trigger by grammar_triggers. Default: false\n   */\n  grammar_lazy?: boolean;\n  /**\n   * Enable thinking if jinja is enabled. Default: true\n   */\n  enable_thinking?: boolean;\n  /**\n   * Force thinking to be open. Default: false\n   */\n  thinking_forced_open?: boolean;\n  /**\n   * Lazy grammar triggers. Default: []\n   */\n  grammar_triggers?: Array<{\n    type: number;\n    value: string;\n    token: number;\n  }>;\n  preserved_tokens?: Array<string>;\n  chat_format?: number;\n  reasoning_format?: string;\n  /**\n   * Path to an image file to process before generating text.\n   * When provided, the image will be processed and added to the context.\n   * Requires multimodal support to be enabled via initMultimodal.\n   */\n  media_paths?: Array<string>;\n  /**\n   * Specify a JSON array of stopping strings.\n   * These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`\n   */\n  stop?: Array<string>;\n  /**\n   * Set the maximum number of tokens to predict when generating text.\n   * **Note:** May exceed the set limit slightly if the last token is a partial multibyte character.\n   * When 0,no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.\n   */\n  n_predict?: number;\n  /**\n   * If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings.\n   * Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings.\n   * Default: `0`\n   */\n  n_probs?: number;\n  /**\n   * Limit the next token selection to the K most probable tokens.  Default: `40`\n   */\n  top_k?: number;\n  /**\n   * Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P. Default: `0.95`\n   */\n  top_p?: number;\n  /**\n   * The minimum probability for a token to be considered, relative to the probability of the most likely token. Default: `0.05`\n   */\n  min_p?: number;\n  /**\n   * Set the chance for token removal via XTC sampler. Default: `0.0`, which is disabled.\n   */\n  xtc_probability?: number;\n  /**\n   * Set a minimum probability threshold for tokens to be removed via XTC sampler. Default: `0.1` (> `0.5` disables XTC)\n   */\n  xtc_threshold?: number;\n  /**\n   * Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.\n   */\n  typical_p?: number;\n  /**\n   * Adjust the randomness of the generated text. Default: `0.8`\n   */\n  temperature?: number;\n  /**\n   * Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.\n   */\n  penalty_last_n?: number;\n  /**\n   * Control the repetition of token sequences in the generated text. Default: `1.0`\n   */\n  penalty_repeat?: number;\n  /**\n   * Repeat alpha frequency penalty. Default: `0.0`, which is disabled.\n   */\n  penalty_freq?: number;\n  /**\n   * Repeat alpha presence penalty. Default: `0.0`, which is disabled.\n   */\n  penalty_present?: number;\n  /**\n   * Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.\n   */\n  mirostat?: number;\n  /**\n   * Set the Mirostat target entropy, parameter tau. Default: `5.0`\n   */\n  mirostat_tau?: number;\n  /**\n   * Set the Mirostat learning rate, parameter eta. Default: `0.1`\n   */\n  mirostat_eta?: number;\n  /**\n   * Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.\n   */\n  dry_multiplier?: number;\n  /**\n   * Set the DRY repetition penalty base value. Default: `1.75`\n   */\n  dry_base?: number;\n  /**\n   * Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`\n   */\n  dry_allowed_length?: number;\n  /**\n   * How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.\n   */\n  dry_penalty_last_n?: number;\n  /**\n   * Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\\n', ':', '\"', '*']`\n   */\n  dry_sequence_breakers?: Array<string>;\n  /**\n   * Top n sigma sampling as described in academic paper \"Top-nσ: Not All Logits Are You Need\" https://arxiv.org/pdf/2411.07641. Default: `-1.0` (Disabled)\n   */\n  top_n_sigma?: number;\n\n  /**\n   * Ignore end of stream token and continue generating. Default: `false`\n   */\n  ignore_eos?: boolean;\n  /**\n   * Modify the likelihood of a token appearing in the generated text completion.\n   * For example, use `\"logit_bias\": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `\"logit_bias\": [[15043,-1.0]]` to decrease its likelihood.\n   * Setting the value to false, `\"logit_bias\": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings,\n   * e.g.`[[\"Hello, World!\",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does.\n   * Default: `[]`\n   */\n  logit_bias?: Array<Array<number>>;\n  /**\n   * Set the random number generator (RNG) seed. Default: `-1`, which is a random seed.\n   */\n  seed?: number;\n\n  /**\n   * Guide tokens for the completion.\n   * Help prevent hallucinations by forcing the TTS to use the correct words.\n   * Default: `[]`\n   */\n  guide_tokens?: Array<number>;\n\n  emit_partial_completion: boolean;\n}\n\nexport interface NativeCompletionTokenProbItem {\n  tok_str: string;\n  prob: number;\n}\n\nexport interface NativeCompletionTokenProb {\n  content: string;\n  probs: Array<NativeCompletionTokenProbItem>;\n}\n\nexport interface NativeCompletionResultTimings {\n  prompt_n: number;\n  prompt_ms: number;\n  prompt_per_token_ms: number;\n  prompt_per_second: number;\n  predicted_n: number;\n  predicted_ms: number;\n  predicted_per_token_ms: number;\n  predicted_per_second: number;\n}\n\nexport interface NativeCompletionResult {\n  /**\n   * Original text (Ignored reasoning_content / tool_calls)\n   */\n  text: string;\n  /**\n   * Reasoning content (parsed for reasoning model)\n   */\n  reasoning_content: string;\n  /**\n   * Tool calls (parsed from response)\n   */\n  tool_calls: Array<{\n    type: 'function';\n    function: {\n      name: string;\n      arguments: string; // JSON string of arguments\n    };\n    id?: string;\n  }>;\n  /**\n   * Content text (Filtered text by reasoning_content / tool_calls)\n   */\n  content: string;\n\n  chat_format: number;\n\n  tokens_predicted: number;\n  tokens_evaluated: number;\n  truncated: boolean;\n  stopped_eos: boolean;\n  stopped_word: string;\n  stopped_limit: number;\n  stopping_word: string;\n  context_full: boolean;\n  interrupted: boolean;\n  tokens_cached: number;\n  timings: NativeCompletionResultTimings;\n\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\n  audio_tokens?: Array<number>;\n}\n\nexport interface NativeTokenizeResult {\n  tokens: Array<number>;\n  /**\n   * Whether the tokenization contains images\n   */\n  has_images: boolean;\n  /**\n   * Bitmap hashes of the images\n   */\n  bitmap_hashes: Array<number>;\n  /**\n   * Chunk positions of the text and images\n   */\n  chunk_pos: Array<number>;\n  /**\n   * Chunk positions of the images\n   */\n  chunk_pos_images: Array<number>;\n}\n\nexport interface NativeEmbeddingResult {\n  embedding: Array<number>;\n}\n\nexport interface NativeLlamaContext {\n  contextId: number;\n  model: {\n    desc: string;\n    size: number;\n    nEmbd: number;\n    nParams: number;\n    chatTemplates: {\n      llamaChat: boolean; // Chat template in llama-chat.cpp\n      minja: {\n        // Chat template supported by minja.hpp\n        default: boolean;\n        defaultCaps: {\n          tools: boolean;\n          toolCalls: boolean;\n          toolResponses: boolean;\n          systemRole: boolean;\n          parallelToolCalls: boolean;\n          toolCallId: boolean;\n        };\n        toolUse: boolean;\n        toolUseCaps: {\n          tools: boolean;\n          toolCalls: boolean;\n          toolResponses: boolean;\n          systemRole: boolean;\n          parallelToolCalls: boolean;\n          toolCallId: boolean;\n        };\n      };\n    };\n    metadata: Object;\n    isChatTemplateSupported: boolean; // Deprecated\n  };\n  /**\n   * Loaded library name for Android\n   */\n  androidLib?: string;\n  gpu: boolean;\n  reasonNoGPU: string;\n}\n\nexport interface NativeSessionLoadResult {\n  tokens_loaded: number;\n  prompt: string;\n}\n\nexport interface NativeLlamaMessagePart {\n  type: 'text';\n  text: string;\n}\n\nexport interface NativeLlamaChatMessage {\n  role: string;\n  content: string | Array<NativeLlamaMessagePart>;\n}\n\nexport interface FormattedChatResult {\n  type: 'jinja' | 'llama-chat';\n  prompt: string;\n  has_media: boolean;\n  media_paths?: Array<string>;\n}\n\nexport interface JinjaFormattedChatResult extends FormattedChatResult {\n  chat_format?: number;\n  grammar?: string;\n  grammar_lazy?: boolean;\n  grammar_triggers?: Array<{\n    type: number;\n    value: string;\n    token: number;\n  }>;\n  thinking_forced_open?: boolean;\n  preserved_tokens?: Array<string>;\n  additional_stops?: Array<string>;\n}\n\nexport interface NativeImageProcessingResult {\n  success: boolean;\n  prompt: string;\n  error?: string;\n}\n\nexport interface NativeRerankParams {\n  normalize?: number;\n}\n\nexport interface NativeRerankResult {\n  score: number;\n  index: number;\n}\n\n// High-level types for the plugin interface\nexport interface LlamaCppMessagePart {\n  type: string;\n  text?: string;\n  image_url?: {\n    url?: string;\n  };\n  input_audio?: {\n    format: string;\n    data?: string;\n    url?: string;\n  };\n}\n\nexport interface LlamaCppOAICompatibleMessage {\n  role: string;\n  content?: string | LlamaCppMessagePart[];\n}\n\nexport interface ToolCall {\n  type: 'function';\n  id?: string;\n  function: {\n    name: string;\n    arguments: string; // JSON string\n  };\n}\n\nexport interface TokenData {\n  token: string;\n  completion_probabilities?: Array<NativeCompletionTokenProb>;\n  // Parsed content from accumulated text\n  content?: string;\n  reasoning_content?: string;\n  tool_calls?: Array<ToolCall>;\n  accumulated_text?: string;\n}\n\nexport interface ContextParams extends Omit<\n  NativeContextParams,\n  'cache_type_k' | 'cache_type_v' | 'pooling_type'\n> {\n  cache_type_k?:\n    | 'f16'\n    | 'f32'\n    | 'bf16'\n    | 'q8_0'\n    | 'q4_0'\n    | 'q4_1'\n    | 'iq4_nl'\n    | 'q5_0'\n    | 'q5_1';\n  cache_type_v?:\n    | 'f16'\n    | 'f32'\n    | 'bf16'\n    | 'q8_0'\n    | 'q4_0'\n    | 'q4_1'\n    | 'iq4_nl'\n    | 'q5_0'\n    | 'q5_1';\n  pooling_type?: 'none' | 'mean' | 'cls' | 'last' | 'rank';\n}\n\nexport interface EmbeddingParams extends NativeEmbeddingParams {}\n\nexport interface RerankParams {\n  normalize?: number;\n}\n\nexport interface RerankResult {\n  score: number;\n  index: number;\n  document?: string;\n}\n\nexport interface CompletionResponseFormat {\n  type: 'text' | 'json_object' | 'json_schema';\n  json_schema?: {\n    strict?: boolean;\n    schema: object;\n  };\n  schema?: object; // for json_object type\n}\n\nexport interface CompletionBaseParams {\n  prompt?: string;\n  messages?: LlamaCppOAICompatibleMessage[];\n  chatTemplate?: string; // deprecated\n  chat_template?: string;\n  jinja?: boolean;\n  tools?: object;\n  parallel_tool_calls?: object;\n  tool_choice?: string;\n  response_format?: CompletionResponseFormat;\n  media_paths?: string[];\n  add_generation_prompt?: boolean;\n  /*\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\n   */\n  now?: string | number;\n  chat_template_kwargs?: Record<string, string>;\n  /**\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\n   * Used for if last assistant message is for prefill purpose\n   */\n  prefill_text?: string;\n}\n\nexport interface CompletionParams extends Omit<\n  NativeCompletionParams,\n  'emit_partial_completion' | 'prompt'\n> {\n  prompt?: string;\n  messages?: LlamaCppOAICompatibleMessage[];\n  chatTemplate?: string; // deprecated\n  chat_template?: string;\n  jinja?: boolean;\n  /**\n   * GBNF grammar for structured output. Takes precedence over json_schema.\n   */\n  grammar?: string;\n  tools?: object;\n  parallel_tool_calls?: object;\n  tool_choice?: string;\n  response_format?: CompletionResponseFormat;\n  media_paths?: string[];\n  add_generation_prompt?: boolean;\n  /*\n   * Timestamp in seconds since epoch to apply to chat template's strftime_now\n   */\n  now?: string | number;\n  chat_template_kwargs?: Record<string, string>;\n  /**\n   * Prefill text to be used for chat parsing (Generation Prompt + Content)\n   * Used for if last assistant message is for prefill purpose\n   */\n  prefill_text?: string;\n}\n\nexport interface BenchResult {\n  modelDesc: string;\n  modelSize: number;\n  modelNParams: number;\n  ppAvg: number;\n  ppStd: number;\n  tgAvg: number;\n  tgStd: number;\n}\n\n// Main plugin interface\nexport interface LlamaCppPlugin {\n  // Core initialization and management\n  toggleNativeLog(options: { enabled: boolean }): Promise<void>;\n  setContextLimit(options: { limit: number }): Promise<void>;\n  modelInfo(options: { path: string; skip?: string[] }): Promise<Object>;\n  initContext(options: { contextId: number; params: NativeContextParams }): Promise<NativeLlamaContext>;\n  releaseContext(options: { contextId: number }): Promise<void>;\n  releaseAllContexts(): Promise<void>;\n\n  // Chat and completion\n  getFormattedChat(options: {\n    contextId: number;\n    messages: string;\n    chatTemplate?: string;\n    params?: {\n      jinja?: boolean;\n      json_schema?: string;\n      tools?: string;\n      parallel_tool_calls?: string;\n      tool_choice?: string;\n      enable_thinking?: boolean;\n      add_generation_prompt?: boolean;\n      now?: string;\n      chat_template_kwargs?: string;\n    };\n  }): Promise<JinjaFormattedChatResult | string>;\n\n  completion(options: {\n    contextId: number;\n    params: NativeCompletionParams;\n  }): Promise<NativeCompletionResult>;\n\n  // Chat-first methods (like llama-cli -sys)\n  chat(options: {\n    contextId: number;\n    messages: LlamaCppOAICompatibleMessage[];\n    system?: string; // Simple system prompt like llama-cli -sys\n    chatTemplate?: string;\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\n  }): Promise<NativeCompletionResult>;\n\n  // Simple chat with system prompt (like llama-cli -sys \"You are a helpful assistant\")\n  chatWithSystem(options: {\n    contextId: number;\n    system: string;\n    message: string;\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\n  }): Promise<NativeCompletionResult>;\n\n  // Simple text generation (like llama-cli -p \"prompt\")\n  generateText(options: {\n    contextId: number;\n    prompt: string;\n    params?: Omit<NativeCompletionParams, 'prompt' | 'messages'>;\n  }): Promise<NativeCompletionResult>;\n\n  stopCompletion(options: { contextId: number }): Promise<void>;\n\n  // Session management\n  loadSession(options: {\n    contextId: number;\n    filepath: string;\n  }): Promise<NativeSessionLoadResult>;\n\n  saveSession(options: {\n    contextId: number;\n    filepath: string;\n    size: number;\n  }): Promise<number>;\n\n  // Tokenization\n  tokenize(options: {\n    contextId: number;\n    text: string;\n    imagePaths?: Array<string>;\n  }): Promise<NativeTokenizeResult>;\n\n  detokenize(options: {\n    contextId: number;\n    tokens: number[];\n  }): Promise<string>;\n\n  // Embeddings and reranking\n  embedding(options: {\n    contextId: number;\n    text: string;\n    params: NativeEmbeddingParams;\n  }): Promise<NativeEmbeddingResult>;\n\n  rerank(options: {\n    contextId: number;\n    query: string;\n    documents: Array<string>;\n    params?: NativeRerankParams;\n  }): Promise<Array<NativeRerankResult>>;\n\n  // Benchmarking\n  bench(options: {\n    contextId: number;\n    pp: number;\n    tg: number;\n    pl: number;\n    nr: number;\n  }): Promise<string>;\n\n  // LoRA adapters\n  applyLoraAdapters(options: {\n    contextId: number;\n    loraAdapters: Array<{ path: string; scaled?: number }>;\n  }): Promise<void>;\n\n  removeLoraAdapters(options: { contextId: number }): Promise<void>;\n\n  getLoadedLoraAdapters(options: {\n    contextId: number;\n  }): Promise<Array<{ path: string; scaled?: number }>>;\n\n  // Multimodal methods\n  initMultimodal(options: {\n    contextId: number;\n    params: {\n      path: string;\n      use_gpu: boolean;\n    };\n  }): Promise<boolean>;\n\n  isMultimodalEnabled(options: {\n    contextId: number;\n  }): Promise<boolean>;\n\n  getMultimodalSupport(options: {\n    contextId: number;\n  }): Promise<{\n    vision: boolean;\n    audio: boolean;\n  }>;\n\n  releaseMultimodal(options: {\n    contextId: number;\n  }): Promise<void>;\n\n  // TTS methods\n  initVocoder(options: {\n    contextId: number;\n    params: {\n      path: string;\n      n_batch?: number;\n    };\n  }): Promise<boolean>;\n\n  isVocoderEnabled(options: { contextId: number }): Promise<boolean>;\n\n  getFormattedAudioCompletion(options: {\n    contextId: number;\n    speakerJsonStr: string;\n    textToSpeak: string;\n  }): Promise<{\n    prompt: string;\n    grammar?: string;\n  }>;\n\n  getAudioCompletionGuideTokens(options: {\n    contextId: number;\n    textToSpeak: string;\n  }): Promise<Array<number>>;\n\n  decodeAudioTokens(options: {\n    contextId: number;\n    tokens: number[];\n  }): Promise<Array<number>>;\n\n  releaseVocoder(options: { contextId: number }): Promise<void>;\n\n  // Model download and management\n  downloadModel(options: {\n    url: string;\n    filename: string;\n  }): Promise<string>;\n\n  getDownloadProgress(options: {\n    url: string;\n  }): Promise<{\n    progress: number;\n    completed: boolean;\n    failed: boolean;\n    errorMessage?: string;\n    localPath?: string;\n    downloadedBytes: number;\n    totalBytes: number;\n  }>;\n\n  cancelDownload(options: {\n    url: string;\n  }): Promise<boolean>;\n\n  getAvailableModels(): Promise<Array<{\n    name: string;\n    path: string;\n    size: number;\n  }>>;\n\n  // Grammar utilities\n  convertJsonSchemaToGrammar(options: {\n    schema: string;\n  }): Promise<string>;\n\n  /**\n   * Start an in-process HTTP server in native code (not the WebView). Binds to loopback by default.\n   * Uses the same GGUF stack as initContext; exposes GET /health, GET /v1/models, POST /v1/chat/completions (OpenAI-shaped).\n   * This is a lightweight cpp-httplib server — not the full upstream llama.cpp `llama-server` executable.\n   */\n  startNativeLlamaServer(options: {\n    modelPath: string;\n    host?: string;\n    port?: number;\n    params?: NativeContextParams;\n  }): Promise<{ running: boolean }>;\n\n  stopNativeLlamaServer(): Promise<void>;\n\n  isNativeLlamaServerRunning(): Promise<{ running: boolean }>;\n\n  // Events\n  addListener(eventName: string, listenerFunc: (data: any) => void): Promise<void>;\n  removeAllListeners(eventName: string): Promise<void>;\n}\n"]}

package/dist/esm/desktop.d.ts

@@ -0,0 +1,37 @@
+import type { NativeContextParams, NativeLlamaContext } from './definitions';
+import { LlamaCppWeb } from './web';
+/**
+ * Capacitor LlamaCpp implementation for Electron desktop.
+ * Core inference (chat, completion, embeddings) uses the native GPU/CPU sidecar.
+ * Multimodal, LoRA, TTS, and benchmarking use the WASM worker (same as PWA).
+ */
+export declare class LlamaCppDesktop extends LlamaCppWeb {
+    private desktopProvider;
+    private sidecarActive;
+    private gpuEnabled;
+    constructor();
+    initContext({ contextId, params, }: {
+        contextId: number;
+        params: NativeContextParams & {
+            embedding?: boolean;
+        };
+    }): Promise<NativeLlamaContext>;
+    setContextLimit(opts: {
+        limit: number;
+    }): Promise<void>;
+    releaseContext({ contextId }: {
+        contextId: number;
+    }): Promise<void>;
+    startNativeLlamaServer(options: {
+        modelPath: string;
+        host?: string;
+        port?: number;
+        params?: NativeContextParams;
+    }): Promise<{
+        running: boolean;
+    }>;
+    isNativeLlamaServerRunning(): Promise<{
+        running: boolean;
+    }>;
+    stopNativeLlamaServer(): Promise<void>;
+}

package/dist/esm/desktop.js

@@ -0,0 +1,133 @@
+import { LlamaCppWeb } from './web';
+import { DesktopProvider } from './isomorphic/provider.desktop';
+import { getDesktopBridge } from './isomorphic/desktop.runtime';
+const MODEL_DESC_DESKTOP = {
+    desc: 'Desktop sidecar model',
+    size: 0,
+    nEmbd: 0,
+    nParams: 0,
+    chatTemplates: {
+        llamaChat: true,
+        minja: {
+            default: true,
+            defaultCaps: {
+                tools: false, toolCalls: false, toolResponses: false,
+                systemRole: true, parallelToolCalls: false, toolCallId: false,
+            },
+            toolUse: false,
+            toolUseCaps: {
+                tools: false, toolCalls: false, toolResponses: false,
+                systemRole: true, parallelToolCalls: false, toolCallId: false,
+            },
+        },
+    },
+    metadata: {},
+    isChatTemplateSupported: true,
+};
+/** Sidecar DELETE route is `/v1/internal/models/([^/]+)` — never use absolute paths as ids. */
+function sidecarModelIdFromPath(modelPath) {
+    const base = modelPath.split(/[/\\]/).pop() || modelPath;
+    return base.replace(/\.gguf$/i, '') || base;
+}
+/**
+ * Capacitor LlamaCpp implementation for Electron desktop.
+ * Core inference (chat, completion, embeddings) uses the native GPU/CPU sidecar.
+ * Multimodal, LoRA, TTS, and benchmarking use the WASM worker (same as PWA).
+ */
+export class LlamaCppDesktop extends LlamaCppWeb {
+    constructor() {
+        super();
+        this.sidecarActive = false;
+        this.gpuEnabled = false;
+        this.desktopProvider = new DesktopProvider();
+        this.provider =
+            this.desktopProvider;
+    }
+    async initContext({ contextId, params, }) {
+        var _a;
+        const modelPath = params.model;
+        const modelId = sidecarModelIdFromPath(modelPath);
+        const bridge = getDesktopBridge();
+        if (bridge === null || bridge === void 0 ? void 0 : bridge.ensureSidecar) {
+            const result = await bridge.ensureSidecar({
+                modelPath,
+                modelId,
+                n_ctx: params.n_ctx,
+                n_gpu_layers: params.n_gpu_layers,
+                n_threads: params.n_threads,
+                embedding: params.embedding,
+            });
+            if ((result === null || result === void 0 ? void 0 : result.ok) && result.port) {
+                this.desktopProvider.setSidecarPort(result.port);
+                await this.desktopProvider.loadModel({
+                    modelId,
+                    modelPath,
+                    n_ctx: params.n_ctx,
+                    n_gpu_layers: params.n_gpu_layers,
+                    n_threads: params.n_threads,
+                    embedding: params.embedding,
+                });
+                this.sidecarActive = true;
+                this.gpuEnabled = !!result.gpuEnabled;
+                this.contextToModel.set(contextId, modelId);
+                return {
+                    contextId,
+                    gpu: this.gpuEnabled,
+                    reasonNoGPU: this.gpuEnabled ? '' : ((_a = result.reasonNoGpu) !== null && _a !== void 0 ? _a : 'CPU-only inference'),
+                    model: MODEL_DESC_DESKTOP,
+                };
+            }
+        }
+        const fallback = await super.initContext({ contextId, params });
+        this.sidecarActive = false;
+        return fallback;
+    }
+    async setContextLimit(opts) {
+        await this.desktopProvider.setContextLimit(opts.limit);
+    }
+    async releaseContext({ contextId }) {
+        const modelId = this.contextToModel.get(contextId);
+        if (modelId && this.sidecarActive) {
+            await this.desktopProvider.unloadModel(modelId);
+        }
+        await super.releaseContext({ contextId });
+    }
+    async startNativeLlamaServer(options) {
+        var _a, _b, _c, _d;
+        const bridge = getDesktopBridge();
+        if (!(bridge === null || bridge === void 0 ? void 0 : bridge.ensureSidecar)) {
+            throw new Error('LlamaCppDesktop: IPC bridge not available — register ipc-handlers in main process');
+        }
+        const result = await bridge.ensureSidecar({
+            modelPath: options.modelPath,
+            modelId: sidecarModelIdFromPath(options.modelPath),
+            host: (_a = options.host) !== null && _a !== void 0 ? _a : '127.0.0.1',
+            port: options.port,
+            n_ctx: (_b = options.params) === null || _b === void 0 ? void 0 : _b.n_ctx,
+            n_gpu_layers: (_c = options.params) === null || _c === void 0 ? void 0 : _c.n_gpu_layers,
+            n_threads: (_d = options.params) === null || _d === void 0 ? void 0 : _d.n_threads,
+        });
+        if ((result === null || result === void 0 ? void 0 : result.ok) && result.port) {
+            this.desktopProvider.setSidecarPort(result.port);
+            this.sidecarActive = true;
+            return { running: true };
+        }
+        return { running: false };
+    }
+    async isNativeLlamaServerRunning() {
+        const bridge = getDesktopBridge();
+        if (bridge === null || bridge === void 0 ? void 0 : bridge.getSidecarStatus) {
+            const st = await bridge.getSidecarStatus();
+            return { running: !!(st === null || st === void 0 ? void 0 : st.running) };
+        }
+        return { running: this.sidecarActive };
+    }
+    async stopNativeLlamaServer() {
+        const bridge = getDesktopBridge();
+        if (bridge === null || bridge === void 0 ? void 0 : bridge.stopSidecar) {
+            await bridge.stopSidecar();
+        }
+        this.sidecarActive = false;
+    }
+}
+//# sourceMappingURL=desktop.js.map

package/dist/esm/desktop.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"desktop.js","sourceRoot":"","sources":["../../src/desktop.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,WAAW,EAAE,MAAM,OAAO,CAAC;AACpC,OAAO,EAAE,eAAe,EAAE,MAAM,+BAA+B,CAAC;AAChE,OAAO,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAGhE,MAAM,kBAAkB,GAAG;IACzB,IAAI,EAAE,uBAAuB;IAC7B,IAAI,EAAE,CAAC;IACP,KAAK,EAAE,CAAC;IACR,OAAO,EAAE,CAAC;IACV,aAAa,EAAE;QACb,SAAS,EAAE,IAAI;QACf,KAAK,EAAE;YACL,OAAO,EAAE,IAAI;YACb,WAAW,EAAE;gBACX,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,KAAK,EAAE,aAAa,EAAE,KAAK;gBACpD,UAAU,EAAE,IAAI,EAAE,iBAAiB,EAAE,KAAK,EAAE,UAAU,EAAE,KAAK;aAC9D;YACD,OAAO,EAAE,KAAK;YACd,WAAW,EAAE;gBACX,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,KAAK,EAAE,aAAa,EAAE,KAAK;gBACpD,UAAU,EAAE,IAAI,EAAE,iBAAiB,EAAE,KAAK,EAAE,UAAU,EAAE,KAAK;aAC9D;SACF;KACF;IACD,QAAQ,EAAE,EAAE;IACZ,uBAAuB,EAAE,IAAI;CAC9B,CAAC;AAEF,+FAA+F;AAC/F,SAAS,sBAAsB,CAAC,SAAiB;IAC/C,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,IAAI,SAAS,CAAC;IACzD,OAAO,IAAI,CAAC,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC,IAAI,IAAI,CAAC;AAC9C,CAAC;AAED;;;;GAIG;AACH,MAAM,OAAO,eAAgB,SAAQ,WAAW;IAK9C;QACE,KAAK,EAAE,CAAC;QAJF,kBAAa,GAAG,KAAK,CAAC;QACtB,eAAU,GAAG,KAAK,CAAC;QAIzB,IAAI,CAAC,eAAe,GAAG,IAAI,eAAe,EAAE,CAAC;QAC5C,IAA6C,CAAC,QAAQ;YACrD,IAAI,CAAC,eAAyC,CAAC;IACnD,CAAC;IAEQ,KAAK,CAAC,WAAW,CAAC,EACzB,SAAS,EACT,MAAM,GAIP;;QACC,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC;QAC/B,MAAM,OAAO,GAAG,sBAAsB,CAAC,SAAS,CAAC,CAAC;QAClD,MAAM,MAAM,GAAG,gBAAgB,EAAE,CAAC;QAElC,IAAI,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,aAAa,EAAE,CAAC;YAC1B,MAAM,MAAM,GAAG,MAAM,MAAM,CAAC,aAAa,CAAC;gBACxC,SAAS;gBACT,OAAO;gBACP,KAAK,EAAE,MAAM,CAAC,KAAK;gBACnB,YAAY,EAAE,MAAM,CAAC,YAAY;gBACjC,SAAS,EAAE,MAAM,CAAC,SAAS;gBAC3B,SAAS,EAAE,MAAM,CAAC,SAAS;aAC5B,CAAC,CAAC;YACH,IAAI,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,EAAE,KAAI,MAAM,CAAC,IAAI,EAAE,CAAC;gBAC9B,IAAI,CAAC,eAAe,CAAC,cAAc,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;gBACjD,MAAM,IAAI,CAAC,eAAe,CAAC,SAAS,CAAC;oBACnC,OAAO;oBACP,SAAS;oBACT,KAAK,EAAE,MAAM,CAAC,KAAK;oBACnB,YAAY,EAAE,MAAM,CAAC,YAAY;oBACjC,SAAS,EAAE,MAAM,CAAC,SAAS;oBAC3B,SAAS,EAAE,MAAM,CAAC,SAAS;iBAC5B,CAAC,CAAC;gBACH,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;gBAC1B,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC;gBACtC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,SAAS,EAAE,OAAO,CAAC,CAAC;gBAC5C,OAAO;oBACL,SAAS;oBACT,GAAG,EAAE,IAAI,CAAC,UAAU;oBACpB,WAAW,EAAE,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,MAAA,MAAM,CAAC,WAAW,mCAAI,oBAAoB,CAAC;oBAChF,KAAK,EAAE,kBAAkB;iBAC1B,CAAC;YACJ,CAAC;QACH,CAAC;QAED,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,WAAW,CAAC,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC,CAAC;QAChE,IAAI,CAAC,aAAa,GAAG,KAAK,CAAC;QAC3B,OAAO,QAAQ,CAAC;IAClB,CAAC;IAED,KAAK,CAAC,eAAe,CAAC,IAAuB;QAC3C,MAAM,IAAI,CAAC,eAAe,CAAC,eAAe,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IACzD,CAAC;IAEQ,KAAK,CAAC,cAAc,CAAC,EAAE,SAAS,EAAyB;QAChE,MAAM,OAAO,GAAG,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACnD,IAAI,OAAO,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;YAClC,MAAM,IAAI,CAAC,eAAe,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC;QAClD,CAAC;QACD,MAAM,KAAK,CAAC,cAAc,CAAC,EAAE,SAAS,EAAE,CAAC,CAAC;IAC5C,CAAC;IAEQ,KAAK,CAAC,sBAAsB,CAAC,OAKrC;;QACC,MAAM,MAAM,GAAG,gBAAgB,EAAE,CAAC;QAClC,IAAI,CAAC,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,aAAa,CAAA,EAAE,CAAC;YAC3B,MAAM,IAAI,KAAK,CAAC,mFAAmF,CAAC,CAAC;QACvG,CAAC;QACD,MAAM,MAAM,GAAG,MAAM,MAAM,CAAC,aAAa,CAAC;YACxC,SAAS,EAAE,OAAO,CAAC,SAAS;YAC5B,OAAO,EAAE,sBAAsB,CAAC,OAAO,CAAC,SAAS,CAAC;YAClD,IAAI,EAAE,MAAA,OAAO,CAAC,IAAI,mCAAI,WAAW;YACjC,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,KAAK,EAAE,MAAA,OAAO,CAAC,MAAM,0CAAE,KAAK;YAC5B,YAAY,EAAE,MAAA,OAAO,CAAC,MAAM,0CAAE,YAAY;YAC1C,SAAS,EAAE,MAAA,OAAO,CAAC,MAAM,0CAAE,SAAS;SACrC,CAAC,CAAC;QACH,IAAI,CAAA,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,EAAE,KAAI,MAAM,CAAC,IAAI,EAAE,CAAC;YAC9B,IAAI,CAAC,eAAe,CAAC,cAAc,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;YACjD,IAAI,CAAC,aAAa,GAAG,IAAI,CAAC;YAC1B,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;QAC3B,CAAC;QACD,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;IAC5B,CAAC;IAEQ,KAAK,CAAC,0BAA0B;QACvC,MAAM,MAAM,GAAG,gBAAgB,EAAE,CAAC;QAClC,IAAI,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,gBAAgB,EAAE,CAAC;YAC7B,MAAM,EAAE,GAAG,MAAM,MAAM,CAAC,gBAAgB,EAAE,CAAC;YAC3C,OAAO,EAAE,OAAO,EAAE,CAAC,CAAC,CAAA,EAAE,aAAF,EAAE,uBAAF,EAAE,CAAE,OAAO,CAAA,EAAE,CAAC;QACpC,CAAC;QACD,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,aAAa,EAAE,CAAC;IACzC,CAAC;IAEQ,KAAK,CAAC,qBAAqB;QAClC,MAAM,MAAM,GAAG,gBAAgB,EAAE,CAAC;QAClC,IAAI,MAAM,aAAN,MAAM,uBAAN,MAAM,CAAE,WAAW,EAAE,CAAC;YACxB,MAAM,MAAM,CAAC,WAAW,EAAE,CAAC;QAC7B,CAAC;QACD,IAAI,CAAC,aAAa,GAAG,KAAK,CAAC;IAC7B,CAAC;CACF","sourcesContent":["import type { NativeContextParams, NativeLlamaContext } from './definitions';\nimport { LlamaCppWeb } from './web';\nimport { DesktopProvider } from './isomorphic/provider.desktop';\nimport { getDesktopBridge } from './isomorphic/desktop.runtime';\nimport { WebProvider } from './isomorphic/provider.web';\n\nconst MODEL_DESC_DESKTOP = {\n  desc: 'Desktop sidecar model',\n  size: 0,\n  nEmbd: 0,\n  nParams: 0,\n  chatTemplates: {\n    llamaChat: true,\n    minja: {\n      default: true,\n      defaultCaps: {\n        tools: false, toolCalls: false, toolResponses: false,\n        systemRole: true, parallelToolCalls: false, toolCallId: false,\n      },\n      toolUse: false,\n      toolUseCaps: {\n        tools: false, toolCalls: false, toolResponses: false,\n        systemRole: true, parallelToolCalls: false, toolCallId: false,\n      },\n    },\n  },\n  metadata: {},\n  isChatTemplateSupported: true,\n};\n\n/** Sidecar DELETE route is `/v1/internal/models/([^/]+)` — never use absolute paths as ids. */\nfunction sidecarModelIdFromPath(modelPath: string): string {\n  const base = modelPath.split(/[/\\\\]/).pop() || modelPath;\n  return base.replace(/\\.gguf$/i, '') || base;\n}\n\n/**\n * Capacitor LlamaCpp implementation for Electron desktop.\n * Core inference (chat, completion, embeddings) uses the native GPU/CPU sidecar.\n * Multimodal, LoRA, TTS, and benchmarking use the WASM worker (same as PWA).\n */\nexport class LlamaCppDesktop extends LlamaCppWeb {\n  private desktopProvider: DesktopProvider;\n  private sidecarActive = false;\n  private gpuEnabled = false;\n\n  constructor() {\n    super();\n    this.desktopProvider = new DesktopProvider();\n    (this as unknown as { provider: WebProvider }).provider =\n      this.desktopProvider as unknown as WebProvider;\n  }\n\n  override async initContext({\n    contextId,\n    params,\n  }: {\n    contextId: number;\n    params: NativeContextParams & { embedding?: boolean };\n  }): Promise<NativeLlamaContext> {\n    const modelPath = params.model;\n    const modelId = sidecarModelIdFromPath(modelPath);\n    const bridge = getDesktopBridge();\n\n    if (bridge?.ensureSidecar) {\n      const result = await bridge.ensureSidecar({\n        modelPath,\n        modelId,\n        n_ctx: params.n_ctx,\n        n_gpu_layers: params.n_gpu_layers,\n        n_threads: params.n_threads,\n        embedding: params.embedding,\n      });\n      if (result?.ok && result.port) {\n        this.desktopProvider.setSidecarPort(result.port);\n        await this.desktopProvider.loadModel({\n          modelId,\n          modelPath,\n          n_ctx: params.n_ctx,\n          n_gpu_layers: params.n_gpu_layers,\n          n_threads: params.n_threads,\n          embedding: params.embedding,\n        });\n        this.sidecarActive = true;\n        this.gpuEnabled = !!result.gpuEnabled;\n        this.contextToModel.set(contextId, modelId);\n        return {\n          contextId,\n          gpu: this.gpuEnabled,\n          reasonNoGPU: this.gpuEnabled ? '' : (result.reasonNoGpu ?? 'CPU-only inference'),\n          model: MODEL_DESC_DESKTOP,\n        };\n      }\n    }\n\n    const fallback = await super.initContext({ contextId, params });\n    this.sidecarActive = false;\n    return fallback;\n  }\n\n  async setContextLimit(opts: { limit: number }): Promise<void> {\n    await this.desktopProvider.setContextLimit(opts.limit);\n  }\n\n  override async releaseContext({ contextId }: { contextId: number }): Promise<void> {\n    const modelId = this.contextToModel.get(contextId);\n    if (modelId && this.sidecarActive) {\n      await this.desktopProvider.unloadModel(modelId);\n    }\n    await super.releaseContext({ contextId });\n  }\n\n  override async startNativeLlamaServer(options: {\n    modelPath: string;\n    host?: string;\n    port?: number;\n    params?: NativeContextParams;\n  }): Promise<{ running: boolean }> {\n    const bridge = getDesktopBridge();\n    if (!bridge?.ensureSidecar) {\n      throw new Error('LlamaCppDesktop: IPC bridge not available — register ipc-handlers in main process');\n    }\n    const result = await bridge.ensureSidecar({\n      modelPath: options.modelPath,\n      modelId: sidecarModelIdFromPath(options.modelPath),\n      host: options.host ?? '127.0.0.1',\n      port: options.port,\n      n_ctx: options.params?.n_ctx,\n      n_gpu_layers: options.params?.n_gpu_layers,\n      n_threads: options.params?.n_threads,\n    });\n    if (result?.ok && result.port) {\n      this.desktopProvider.setSidecarPort(result.port);\n      this.sidecarActive = true;\n      return { running: true };\n    }\n    return { running: false };\n  }\n\n  override async isNativeLlamaServerRunning(): Promise<{ running: boolean }> {\n    const bridge = getDesktopBridge();\n    if (bridge?.getSidecarStatus) {\n      const st = await bridge.getSidecarStatus();\n      return { running: !!st?.running };\n    }\n    return { running: this.sidecarActive };\n  }\n\n  override async stopNativeLlamaServer(): Promise<void> {\n    const bridge = getDesktopBridge();\n    if (bridge?.stopSidecar) {\n      await bridge.stopSidecar();\n    }\n    this.sidecarActive = false;\n  }\n}\n"]}

package/dist/esm/index.d.ts

@@ -0,0 +1,200 @@
+import type { NativeContextParams, NativeLlamaContext, NativeCompletionParams, NativeCompletionTokenProb, NativeCompletionResult, NativeTokenizeResult, NativeEmbeddingResult, NativeSessionLoadResult, NativeEmbeddingParams, NativeRerankParams, NativeRerankResult, NativeCompletionTokenProbItem, NativeCompletionResultTimings, JinjaFormattedChatResult, FormattedChatResult, NativeImageProcessingResult, LlamaCppMessagePart, LlamaCppOAICompatibleMessage, ContextParams, EmbeddingParams, RerankParams, RerankResult, CompletionResponseFormat, CompletionParams, BenchResult, LlamaCppPlugin, ToolCall, TokenData } from './definitions';
+export declare const LLAMACPP_MTMD_DEFAULT_MEDIA_MARKER = "<__media__>";
+declare const LlamaCpp: LlamaCppPlugin;
+export type RNLlamaMessagePart = LlamaCppMessagePart;
+export type RNLlamaOAICompatibleMessage = LlamaCppOAICompatibleMessage;
+export type { NativeContextParams, NativeLlamaContext, NativeCompletionParams, NativeCompletionTokenProb, NativeCompletionResult, NativeTokenizeResult, NativeEmbeddingResult, NativeSessionLoadResult, NativeEmbeddingParams, NativeRerankParams, NativeRerankResult, NativeCompletionTokenProbItem, NativeCompletionResultTimings, FormattedChatResult, JinjaFormattedChatResult, NativeImageProcessingResult, ContextParams, EmbeddingParams, RerankParams, RerankResult, CompletionResponseFormat, CompletionParams, BenchResult, ToolCall, TokenData, };
+export declare const RNLLAMA_MTMD_DEFAULT_MEDIA_MARKER = "<__media__>";
+export declare class LlamaContext {
+    id: number;
+    gpu: boolean;
+    reasonNoGPU: string;
+    model: NativeLlamaContext['model'];
+    constructor({ contextId, gpu, reasonNoGPU, model }: NativeLlamaContext);
+    /**
+     * Load cached prompt & completion state from a file.
+     */
+    loadSession(filepath: string): Promise<NativeSessionLoadResult>;
+    /**
+     * Save current cached prompt & completion state to a file.
+     */
+    saveSession(filepath: string, options?: {
+        tokenSize: number;
+    }): Promise<number>;
+    isLlamaChatSupported(): boolean;
+    isJinjaSupported(): boolean;
+    getFormattedChat(messages: RNLlamaOAICompatibleMessage[], template?: string | null, params?: {
+        jinja?: boolean;
+        response_format?: CompletionResponseFormat;
+        tools?: object;
+        parallel_tool_calls?: object;
+        tool_choice?: string;
+        enable_thinking?: boolean;
+        add_generation_prompt?: boolean;
+        now?: string | number;
+        chat_template_kwargs?: Record<string, string>;
+    }): Promise<FormattedChatResult | JinjaFormattedChatResult>;
+    /**
+     * Generate a completion based on the provided parameters
+     * @param params Completion parameters including prompt or messages
+     * @param callback Optional callback for token-by-token streaming
+     * @returns Promise resolving to the completion result
+     *
+     * Note: For multimodal support, you can include an media_paths parameter.
+     * This will process the images and add them to the context before generating text.
+     * Multimodal support must be enabled via initMultimodal() first.
+     */
+    completion(params: CompletionParams, callback?: (data: TokenData) => void): Promise<NativeCompletionResult>;
+    stopCompletion(): Promise<void>;
+    /**
+     * Tokenize text or text with images
+     * @param text Text to tokenize
+     * @param params.media_paths Array of image paths to tokenize (if multimodal is enabled)
+     * @returns Promise resolving to the tokenize result
+     */
+    tokenize(text: string, { media_paths: mediaPaths, }?: {
+        media_paths?: string[];
+    }): Promise<NativeTokenizeResult>;
+    detokenize(tokens: number[]): Promise<string>;
+    embedding(text: string, params?: EmbeddingParams): Promise<NativeEmbeddingResult>;
+    /**
+     * Rerank documents based on relevance to a query
+     * @param query The query text to rank documents against
+     * @param documents Array of document texts to rank
+     * @param params Optional reranking parameters
+     * @returns Promise resolving to an array of ranking results with scores and indices
+     */
+    rerank(query: string, documents: string[], params?: RerankParams): Promise<RerankResult[]>;
+    bench(pp: number, tg: number, pl: number, nr: number): Promise<BenchResult>;
+    applyLoraAdapters(loraList: Array<{
+        path: string;
+        scaled?: number;
+    }>): Promise<void>;
+    removeLoraAdapters(): Promise<void>;
+    getLoadedLoraAdapters(): Promise<Array<{
+        path: string;
+        scaled?: number;
+    }>>;
+    /**
+     * Initialize multimodal support with a mmproj file
+     * @param params Parameters for multimodal support
+     * @param params.path Path to the multimodal projector file
+     * @param params.use_gpu Whether to use GPU
+     * @returns Promise resolving to true if initialization was successful
+     */
+    initMultimodal({ path, use_gpu: useGpu, }: {
+        path: string;
+        use_gpu?: boolean;
+    }): Promise<boolean>;
+    /**
+     * Check if multimodal support is enabled
+     * @returns Promise resolving to true if multimodal is enabled
+     */
+    isMultimodalEnabled(): Promise<boolean>;
+    /**
+     * Check multimodal support
+     * @returns Promise resolving to an object with vision and audio support
+     */
+    getMultimodalSupport(): Promise<{
+        vision: boolean;
+        audio: boolean;
+    }>;
+    /**
+     * Release multimodal support
+     * @returns Promise resolving to void
+     */
+    releaseMultimodal(): Promise<void>;
+    /**
+     * Initialize TTS support with a vocoder model
+     * @param params Parameters for TTS support
+     * @param params.path Path to the vocoder model
+     * @param params.n_batch Batch size for the vocoder model
+     * @returns Promise resolving to true if initialization was successful
+     */
+    initVocoder({ path, n_batch: nBatch }: {
+        path: string;
+        n_batch?: number;
+    }): Promise<boolean>;
+    /**
+     * Check if TTS support is enabled
+     * @returns Promise resolving to true if TTS is enabled
+     */
+    isVocoderEnabled(): Promise<boolean>;
+    /**
+     * Get a formatted audio completion prompt
+     * @param speakerJsonStr JSON string representing the speaker
+     * @param textToSpeak Text to speak
+     * @returns Promise resolving to the formatted audio completion result with prompt and grammar
+     */
+    getFormattedAudioCompletion(speaker: object | null, textToSpeak: string): Promise<{
+        prompt: string;
+        grammar?: string;
+    }>;
+    /**
+     * Get guide tokens for audio completion
+     * @param textToSpeak Text to speak
+     * @returns Promise resolving to the guide tokens
+     */
+    getAudioCompletionGuideTokens(textToSpeak: string): Promise<Array<number>>;
+    /**
+     * Decode audio tokens
+     * @param tokens Array of audio tokens
+     * @returns Promise resolving to the decoded audio tokens
+     */
+    decodeAudioTokens(tokens: number[]): Promise<Array<number>>;
+    /**
+     * Release TTS support
+     * @returns Promise resolving to void
+     */
+    releaseVocoder(): Promise<void>;
+    release(): Promise<void>;
+}
+export declare function toggleNativeLog(enabled: boolean): Promise<void>;
+export declare function addNativeLogListener(listener: (level: string, text: string) => void): {
+    remove: () => void;
+};
+export declare function setContextLimit(limit: number): Promise<void>;
+export declare function loadLlamaModelInfo(model: string): Promise<Object>;
+export declare function initLlama({ model, is_model_asset: isModelAsset, pooling_type: poolingType, lora, lora_list: loraList, ...rest }: ContextParams, onProgress?: (progress: number) => void): Promise<LlamaContext>;
+export declare function releaseAllLlama(): Promise<void>;
+export declare function downloadModel(url: string, filename: string): Promise<string>;
+export declare function getDownloadProgress(url: string): Promise<{
+    progress: number;
+    completed: boolean;
+    failed: boolean;
+    errorMessage?: string;
+    localPath?: string;
+    downloadedBytes: number;
+    totalBytes: number;
+}>;
+export declare function cancelDownload(url: string): Promise<boolean>;
+export declare function getAvailableModels(): Promise<Array<{
+    name: string;
+    path: string;
+    size: number;
+}>>;
+/**
+ * Convert a JSON schema to GBNF grammar format
+ * @param schema JSON schema object
+ * @returns Promise resolving to GBNF grammar string
+ */
+export declare function convertJsonSchemaToGrammar(schema: object): Promise<string>;
+export declare const BuildInfo: {
+    number: string;
+    commit: string;
+};
+export { LlamaCpp };
+export * from './isomorphic/provider.interface';
+export * from './isomorphic/errors';
+export * from './isomorphic/provider.factory';
+export * from './isomorphic/provider.native';
+export * from './isomorphic/provider.web';
+export * from './isomorphic/provider.desktop';
+export * from './isomorphic/desktop.runtime';
+export { LlamaCppDesktop } from './desktop';
+export * from './isomorphic/model.admission';
+export * from './isomorphic/model.scheduler';
+export * from './isomorphic/wasmMemoryPolicy';
+export * from './isomorphic/wasmMemoryCalibration';
+export * from './storage/manifest';
+export * from './storage/opfs.store';