@fugood/llama.node 1.4.13 → 1.4.15

package/lib/binding.ts

@@ -112,7 +112,7 @@ export type CompletionResponseFormat = {
 export type LlamaCompletionOptions = {
   messages?: ChatMessage[]
   jinja?: boolean
-  reasoning_format?: string
+  reasoning_format?: 'none' | 'auto' | 'deepseek'
   chat_template?: string
   response_format?: CompletionResponseFormat
   tools?: Tool[]
@@ -200,6 +200,13 @@ export type LlamaParallelCompletionOptions = LlamaCompletionOptions & {
    */
   save_state_path?: string
 
+  /**
+   * File path to save prompt-only state to after prompt processing.
+   * Useful for fast prompt reuse (especially for recurrent/hybrid models).
+   * Example: `'/path/to/prompt_state.bin'` or `'file:///path/to/prompt_state.bin'`
+   */
+  save_prompt_state_path?: string
+
   /**
    * Number of tokens to load when loading state.
    * If not specified or <= 0, all tokens from the state file will be loaded.
@@ -363,6 +370,8 @@ export type ModelInfo = {
   nEmbd: number
   nParams: number
   size: number
+  is_recurrent: boolean
+  is_hybrid: boolean
   chatTemplates: {
     llamaChat: boolean
     minja: {
@@ -475,6 +484,7 @@ export interface LlamaContext {
       parallel_tool_calls?: boolean
       tool_choice?: string
       enable_thinking?: boolean
+      reasoning_format?: 'none' | 'auto' | 'deepseek'
       add_generation_prompt?: boolean
       now?: string | number
       chat_template_kwargs?: Record<string, string>
@@ -505,9 +515,20 @@ export interface LlamaContext {
   /**
    * Initialize multimodal support with a mmproj file
    * @param options Object containing path and optional use_gpu flag
+   * @param options.path Path to the multimodal projector model file (mmproj)
+   * @param options.use_gpu Whether to use GPU for multimodal processing (default: true)
+   * @param options.image_min_tokens Minimum number of tokens for image input (for dynamic resolution models)
+   * @param options.image_max_tokens Maximum number of tokens for image input (for dynamic resolution models).
+   *                                  Lower values reduce memory usage and improve speed for high-resolution images.
+   *                                  Recommended: 256-512 for faster inference, up to 4096 for maximum detail.
    * @returns boolean indicating if initialization was successful
    */
-  initMultimodal(options: { path: string; use_gpu?: boolean }): boolean
+  initMultimodal(options: {
+    path: string
+    use_gpu?: boolean
+    image_min_tokens?: number
+    image_max_tokens?: number
+  }): boolean
 
   /**
    * Check if multimodal support is enabled

package/lib/index.js

@@ -87,7 +87,7 @@ class LlamaContextWrapper {
         return !!this.ctx.getModelInfo().chatTemplates.llamaChat;
     }
     getFormattedChat(messages, template, params) {
-        var _a, _b;
+        var _a, _b, _c;
         const { messages: chat, has_media, media_paths } = (0, utils_1.formatMediaChat)(messages);
         const useJinja = this.isJinjaSupported() && ((_a = params === null || params === void 0 ? void 0 : params.jinja) !== null && _a !== void 0 ? _a : true);
         let tmpl;
@@ -100,6 +100,7 @@ class LlamaContextWrapper {
             parallel_tool_calls: params === null || params === void 0 ? void 0 : params.parallel_tool_calls,
             tool_choice: params === null || params === void 0 ? void 0 : params.tool_choice,
             enable_thinking: (_b = params === null || params === void 0 ? void 0 : params.enable_thinking) !== null && _b !== void 0 ? _b : true,
+            reasoning_format: (_c = params === null || params === void 0 ? void 0 : params.reasoning_format) !== null && _c !== void 0 ? _c : 'none',
             add_generation_prompt: params === null || params === void 0 ? void 0 : params.add_generation_prompt,
             now: params === null || params === void 0 ? void 0 : params.now,
             chat_template_kwargs: (params === null || params === void 0 ? void 0 : params.chat_template_kwargs)

package/lib/index.ts

@@ -118,6 +118,7 @@ class LlamaContextWrapper {
       parallel_tool_calls?: boolean
       tool_choice?: string
       enable_thinking?: boolean
+      reasoning_format?: 'none' | 'auto' | 'deepseek'
       add_generation_prompt?: boolean
       now?: string | number
       chat_template_kwargs?: Record<string, string>
@@ -136,6 +137,7 @@ class LlamaContextWrapper {
       parallel_tool_calls: params?.parallel_tool_calls,
       tool_choice: params?.tool_choice,
       enable_thinking: params?.enable_thinking ?? true,
+      reasoning_format: params?.reasoning_format ?? 'none',
       add_generation_prompt: params?.add_generation_prompt,
       now: params?.now,
       chat_template_kwargs: params?.chat_template_kwargs
@@ -252,7 +254,12 @@ class LlamaContextWrapper {
     return this.ctx.getLoadedLoraAdapters()
   }
 
-  initMultimodal(options: { path: string; use_gpu?: boolean }): boolean {
+  initMultimodal(options: {
+    path: string
+    use_gpu?: boolean
+    image_min_tokens?: number
+    image_max_tokens?: number
+  }): boolean {
     return this.ctx.initMultimodal(options)
   }
 

package/lib/parallel.ts

@@ -1,10 +1,10 @@
 // Parallel decoding API implementation for llama.node
 import type {
   LlamaContext,
-  LlamaCompletionOptions,
   LlamaCompletionToken,
   RerankParams,
   ParallelStatus,
+  LlamaParallelCompletionOptions,
 } from './binding'
 import { formatMediaChat } from './utils'
 
@@ -68,7 +68,7 @@ export class LlamaParallelAPI {
    * @returns Object with requestId, promise for result, and stop function
    */
   async completion(
-    options: LlamaCompletionOptions,
+    options: LlamaParallelCompletionOptions,
     onToken?: (requestId: number, data: LlamaCompletionToken) => void,
   ): Promise<{
     requestId: number

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@fugood/llama.node",
   "access": "public",
-  "version": "1.4.13",
+  "version": "1.4.15",
   "description": "An another Node binding of llama.cpp",
   "main": "lib/index.js",
   "scripts": {
@@ -72,20 +72,20 @@
     "CMakeLists.txt"
   ],
   "optionalDependencies": {
-    "@fugood/node-llama-darwin-arm64": "1.4.13",
-    "@fugood/node-llama-darwin-x64": "1.4.13",
-    "@fugood/node-llama-linux-arm64": "1.4.13",
-    "@fugood/node-llama-linux-arm64-cuda": "1.4.13",
-    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.13",
-    "@fugood/node-llama-linux-arm64-vulkan": "1.4.13",
-    "@fugood/node-llama-linux-x64": "1.4.13",
-    "@fugood/node-llama-linux-x64-cuda": "1.4.13",
-    "@fugood/node-llama-linux-x64-vulkan": "1.4.13",
-    "@fugood/node-llama-win32-arm64": "1.4.13",
-    "@fugood/node-llama-win32-arm64-vulkan": "1.4.13",
-    "@fugood/node-llama-win32-x64": "1.4.13",
-    "@fugood/node-llama-win32-x64-cuda": "1.4.13",
-    "@fugood/node-llama-win32-x64-vulkan": "1.4.13"
+    "@fugood/node-llama-darwin-arm64": "1.4.15",
+    "@fugood/node-llama-darwin-x64": "1.4.15",
+    "@fugood/node-llama-linux-arm64": "1.4.15",
+    "@fugood/node-llama-linux-arm64-cuda": "1.4.15",
+    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.15",
+    "@fugood/node-llama-linux-arm64-vulkan": "1.4.15",
+    "@fugood/node-llama-linux-x64": "1.4.15",
+    "@fugood/node-llama-linux-x64-cuda": "1.4.15",
+    "@fugood/node-llama-linux-x64-vulkan": "1.4.15",
+    "@fugood/node-llama-win32-arm64": "1.4.15",
+    "@fugood/node-llama-win32-arm64-vulkan": "1.4.15",
+    "@fugood/node-llama-win32-x64": "1.4.15",
+    "@fugood/node-llama-win32-x64-cuda": "1.4.15",
+    "@fugood/node-llama-win32-x64-vulkan": "1.4.15"
   },
   "devDependencies": {
     "@babel/preset-env": "^7.24.4",

package/scripts/llama.cpp.patch

@@ -1,8 +1,8 @@
 diff --git a/src/llama.cpp/common/CMakeLists.txt b/src/llama.cpp/common/CMakeLists.txt
-index f7b99159e..fa37fed19 100644
+index 723973ed7..e4b2c6537 100644
 --- a/src/llama.cpp/common/CMakeLists.txt
 +++ b/src/llama.cpp/common/CMakeLists.txt
-@@ -154,8 +154,14 @@ if (LLAMA_LLGUIDANCE)
+@@ -146,4 +146,11 @@ if (LLAMA_LLGUIDANCE)
      set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
  endif ()
  
@@ -13,11 +13,8 @@ index f7b99159e..fa37fed19 100644
 +else()
 +    set(LLAMA_COMMON_WIN_LIBS "")
 +endif()
- 
++
 +target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} ${LLAMA_COMMON_WIN_LIBS} PUBLIC llama Threads::Threads)
- 
- #
- # copy the license files
 diff --git a/src/llama.cpp/common/chat-peg-parser.cpp b/src/llama.cpp/common/chat-peg-parser.cpp
 index 1bcba9cd8..b7cd68734 100644
 --- a/src/llama.cpp/common/chat-peg-parser.cpp
@@ -32,7 +29,7 @@ index 1bcba9cd8..b7cd68734 100644
  static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
      int count = 0;
 diff --git a/src/llama.cpp/common/chat.cpp b/src/llama.cpp/common/chat.cpp
-index 22e527bab..c3d0affca 100644
+index d531388bc..e6712b368 100644
 --- a/src/llama.cpp/common/chat.cpp
 +++ b/src/llama.cpp/common/chat.cpp
 @@ -7,9 +7,6 @@
@@ -62,7 +59,7 @@ index 22e527bab..c3d0affca 100644
  struct templates_params {
      json messages;
      json tools;
-@@ -752,7 +739,7 @@ static std::string apply(
+@@ -753,7 +740,7 @@ static std::string apply(
          tmpl_inputs.extra_context.merge_patch(*additional_context);
      }
      // TODO: add flag to control date/time, if only for testing purposes.
@@ -72,7 +69,7 @@ index 22e527bab..c3d0affca 100644
      minja::chat_template_options tmpl_opts;
      // To avoid double BOS / EOS tokens, we're manually removing begining / trailing tokens
 diff --git a/src/llama.cpp/common/chat.h b/src/llama.cpp/common/chat.h
-index 8bd4a325f..333b3301f 100644
+index 454085e90..e01390cf9 100644
 --- a/src/llama.cpp/common/chat.h
 +++ b/src/llama.cpp/common/chat.h
 @@ -10,7 +10,18 @@
@@ -96,7 +93,7 @@ index 8bd4a325f..333b3301f 100644
  struct common_chat_tool_call {
      std::string name;
 diff --git a/src/llama.cpp/common/common.cpp b/src/llama.cpp/common/common.cpp
-index 41b2b6833..fe9ba05aa 100644
+index 744f0b4ee..04fcebb9e 100644
 --- a/src/llama.cpp/common/common.cpp
 +++ b/src/llama.cpp/common/common.cpp
 @@ -1361,6 +1361,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
@@ -108,10 +105,10 @@ index 41b2b6833..fe9ba05aa 100644
      mparams.main_gpu        = params.main_gpu;
      mparams.split_mode      = params.split_mode;
 diff --git a/src/llama.cpp/common/common.h b/src/llama.cpp/common/common.h
-index d6fd0d37a..477209ce5 100644
+index e60087dea..c21797cd8 100644
 --- a/src/llama.cpp/common/common.h
 +++ b/src/llama.cpp/common/common.h
-@@ -310,6 +310,7 @@ struct lr_opt {
+@@ -311,6 +311,7 @@ struct lr_opt {
  struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
  
  struct common_params {

package/src/LlamaContext.cpp

@@ -595,6 +595,8 @@ Napi::Value LlamaContext::GetModelInfo(const Napi::CallbackInfo &info) {
   details.Set("nEmbd", llama_model_n_embd(model));
   details.Set("nParams", llama_model_n_params(model));
   details.Set("size", llama_model_size(model));
+  details.Set("is_recurrent", llama_model_is_recurrent(model));
+  details.Set("is_hybrid", llama_model_is_hybrid(model));
 
   Napi::Object chatTemplates = Napi::Object::New(info.Env());
   chatTemplates.Set("llamaChat", _rn_ctx->validateModelChatTemplate(false, nullptr));
@@ -703,6 +705,7 @@ Napi::Value LlamaContext::GetFormattedChat(const Napi::CallbackInfo &info) {
         get_option<bool>(params, "parallel_tool_calls", false);
     auto tool_choice = get_option<std::string>(params, "tool_choice", "");
     auto enable_thinking = get_option<bool>(params, "enable_thinking", false);
+    auto reasoning_format = get_option<std::string>(params, "reasoning_format", "none");
     auto add_generation_prompt = get_option<bool>(params, "add_generation_prompt", true);
     auto now_str = get_option<std::string>(params, "now", "");
 
@@ -721,7 +724,7 @@ Napi::Value LlamaContext::GetFormattedChat(const Napi::CallbackInfo &info) {
     try {
       chatParams = _rn_ctx->getFormattedChatWithJinja(
           messages, chat_template, json_schema_str, tools_str,
-          parallel_tool_calls, tool_choice, enable_thinking,
+          parallel_tool_calls, tool_choice, enable_thinking, reasoning_format,
           add_generation_prompt, now_str, chat_template_kwargs);
     } catch (const nlohmann::json_abi_v3_12_0::detail::parse_error& e) {
       Napi::Error::New(env, e.what()).ThrowAsJavaScriptException();
@@ -962,7 +965,7 @@ Napi::Value LlamaContext::Completion(const Napi::CallbackInfo &info) {
       try {
         chatParams = _rn_ctx->getFormattedChatWithJinja(
             json_stringify(messages), chat_template,
-            json_schema_str, tools_str, parallel_tool_calls, tool_choice, enable_thinking,
+            json_schema_str, tools_str, parallel_tool_calls, tool_choice, enable_thinking, reasoning_format,
             add_generation_prompt, now_str, chat_template_kwargs);
       } catch (const std::exception &e) {
         Napi::Error::New(env, e.what()).ThrowAsJavaScriptException();
@@ -1330,7 +1333,7 @@ extern "C" void cleanup_logging() {
 }
 
 
-// initMultimodal(options: { path: string, use_gpu?: boolean }): boolean
+// initMultimodal(options: { path: string, use_gpu?: boolean, image_min_tokens?: number, image_max_tokens?: number }): boolean
 Napi::Value LlamaContext::InitMultimodal(const Napi::CallbackInfo &info) {
   Napi::Env env = info.Env();
 
@@ -1342,6 +1345,15 @@ Napi::Value LlamaContext::InitMultimodal(const Napi::CallbackInfo &info) {
   auto options = info[0].As<Napi::Object>();
   auto mmproj_path = options.Get("path").ToString().Utf8Value();
   auto use_gpu = options.Get("use_gpu").ToBoolean().Value();
+  int image_min_tokens = -1;
+  int image_max_tokens = -1;
+
+  if (options.Has("image_min_tokens") && options.Get("image_min_tokens").IsNumber()) {
+    image_min_tokens = options.Get("image_min_tokens").ToNumber().Int32Value();
+  }
+  if (options.Has("image_max_tokens") && options.Get("image_max_tokens").IsNumber()) {
+    image_max_tokens = options.Get("image_max_tokens").ToNumber().Int32Value();
+  }
 
   if (mmproj_path.empty()) {
     Napi::TypeError::New(env, "mmproj path is required")
@@ -1357,7 +1369,7 @@ Napi::Value LlamaContext::InitMultimodal(const Napi::CallbackInfo &info) {
 
   // Disable ctx_shift before initializing multimodal
   _rn_ctx->params.ctx_shift = false;
-  bool result = _rn_ctx->initMultimodal(mmproj_path, use_gpu);
+  bool result = _rn_ctx->initMultimodal(mmproj_path, use_gpu, image_min_tokens, image_max_tokens);
   if (!result) {
     Napi::Error::New(env, "Failed to initialize multimodal context")
         .ThrowAsJavaScriptException();

package/src/llama.cpp/CMakeLists.txt

@@ -111,11 +111,16 @@ option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
 
 # 3rd party libs
-option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
-option(LLAMA_HTTPLIB    "llama: if libcurl is disabled, use httplib to download model from an URL" ON)
-option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" OFF)
+option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
+option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
+# deprecated
+option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
+if (LLAMA_CURL)
+    message(WARNING "LLAMA_CURL option is deprecated and will be ignored")
+endif()
+
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
@@ -182,6 +187,9 @@ if (NOT MSVC)
     endif()
 endif()
 
+include("cmake/license.cmake")
+license_add_file("llama.cpp" "LICENSE")
+
 #
 # 3rd-party
 #
@@ -209,11 +217,6 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #
 
-if (NOT LLAMA_BUILD_COMMON)
-    message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL")
-    set(LLAMA_CURL OFF)
-endif()
-
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
     if (LLAMA_HTTPLIB)
@@ -235,6 +238,19 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
     add_subdirectory(tools)
 endif()
 
+# Automatically add all files from the 'licenses' directory
+file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
+
+foreach(FILE_PATH ${EXTRA_LICENSES})
+    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
+    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
+    license_add_file("${NAME}" "${FILE_PATH}")
+endforeach()
+
+if (LLAMA_BUILD_COMMON)
+    license_generate(common)
+endif()
+
 #
 # install
 #

package/src/llama.cpp/common/CMakeLists.txt

@@ -60,6 +60,8 @@ add_library(${TARGET} STATIC
     common.h
     console.cpp
     console.h
+    debug.cpp
+    debug.h
     download.cpp
     download.h
     http.h
@@ -95,17 +97,7 @@ endif()
 # TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
 set(LLAMA_COMMON_EXTRA_LIBS build_info)
 
-if (LLAMA_CURL)
-    # Use curl to download model url
-    find_package(CURL)
-    if (NOT CURL_FOUND)
-        message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF")
-    endif()
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
-    include_directories(${CURL_INCLUDE_DIRS})
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
-elseif (LLAMA_HTTPLIB)
-    # otherwise, use cpp-httplib
+if (LLAMA_HTTPLIB)
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
 endif()
@@ -162,26 +154,3 @@ else()
 endif()
 
 target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} ${LLAMA_COMMON_WIN_LIBS} PUBLIC llama Threads::Threads)
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
-    message(STATUS "Running inside GitHub Actions - copying license files")
-
-    # Copy all files from licenses/ to build/bin/
-    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
-    foreach(LICENSE_FILE ${LICENSE_FILES})
-        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
-        add_custom_command(
-            POST_BUILD
-            TARGET ${TARGET}
-            COMMAND ${CMAKE_COMMAND} -E copy_if_different
-                "${LICENSE_FILE}"
-                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
-            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
-        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
-    endforeach()
-endif()