@fugood/llama.node 1.4.12 → 1.4.14

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>

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

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": "1.4.12",
+  "version": "1.4.14",
   "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": "1.4.12",
-    "@fugood/node-llama-darwin-x64": "1.4.12",
-    "@fugood/node-llama-linux-arm64": "1.4.12",
-    "@fugood/node-llama-linux-arm64-cuda": "1.4.12",
-    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.12",
-    "@fugood/node-llama-linux-arm64-vulkan": "1.4.12",
-    "@fugood/node-llama-linux-x64": "1.4.12",
-    "@fugood/node-llama-linux-x64-cuda": "1.4.12",
-    "@fugood/node-llama-linux-x64-vulkan": "1.4.12",
-    "@fugood/node-llama-win32-arm64": "1.4.12",
-    "@fugood/node-llama-win32-arm64-vulkan": "1.4.12",
-    "@fugood/node-llama-win32-x64": "1.4.12",
-    "@fugood/node-llama-win32-x64-cuda": "1.4.12",
-    "@fugood/node-llama-win32-x64-vulkan": "1.4.12"
+    "@fugood/node-llama-darwin-arm64": "1.4.14",
+    "@fugood/node-llama-darwin-x64": "1.4.14",
+    "@fugood/node-llama-linux-arm64": "1.4.14",
+    "@fugood/node-llama-linux-arm64-cuda": "1.4.14",
+    "@fugood/node-llama-linux-arm64-snapdragon": "1.4.14",
+    "@fugood/node-llama-linux-arm64-vulkan": "1.4.14",
+    "@fugood/node-llama-linux-x64": "1.4.14",
+    "@fugood/node-llama-linux-x64-cuda": "1.4.14",
+    "@fugood/node-llama-linux-x64-vulkan": "1.4.14",
+    "@fugood/node-llama-win32-arm64": "1.4.14",
+    "@fugood/node-llama-win32-arm64-vulkan": "1.4.14",
+    "@fugood/node-llama-win32-x64": "1.4.14",
+    "@fugood/node-llama-win32-x64-cuda": "1.4.14",
+    "@fugood/node-llama-win32-x64-vulkan": "1.4.14"
   },
   "devDependencies": {
     "@babel/preset-env": "^7.24.4",

package/scripts/llama.cpp.patch

@@ -32,7 +32,7 @@ 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 b98ab21ce..2f782837a 100644
+index 22e527bab..c3d0affca 100644
 --- a/src/llama.cpp/common/chat.cpp
 +++ b/src/llama.cpp/common/chat.cpp
 @@ -7,9 +7,6 @@
@@ -96,10 +96,10 @@ 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 79c475612..cf189f8bc 100644
+index 744f0b4ee..04fcebb9e 100644
 --- a/src/llama.cpp/common/common.cpp
 +++ b/src/llama.cpp/common/common.cpp
-@@ -1342,6 +1342,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
+@@ -1361,6 +1361,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
          mparams.devices = params.devices.data();
      }
  
@@ -108,10 +108,10 @@ index 79c475612..cf189f8bc 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 f8bc686b6..555ba044a 100644
+index 7794c0268..5b77ae0c3 100644
 --- a/src/llama.cpp/common/common.h
 +++ b/src/llama.cpp/common/common.h
-@@ -307,6 +307,7 @@ struct lr_opt {
+@@ -310,6 +310,7 @@ struct lr_opt {
  struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
  
  struct common_params {
@@ -133,10 +133,10 @@ index 7622d0bf4..d2edcfddb 100644
              check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
              if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
 diff --git a/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
-index 13b96d61f..5fa163442 100644
+index 365a24b49..83bf4ee62 100644
 --- a/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
 +++ b/src/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp
-@@ -2680,9 +2680,24 @@ static const char * ggml_backend_hexagon_device_get_description(ggml_backend_dev
+@@ -2798,9 +2798,24 @@ static const char * ggml_backend_hexagon_device_get_description(ggml_backend_dev
      GGML_UNUSED(dev);
  }
  
@@ -163,7 +163,7 @@ index 13b96d61f..5fa163442 100644
      *total = *free;
  
      GGML_UNUSED(dev);
-@@ -2879,10 +2894,17 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
+@@ -3010,10 +3025,17 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
          }
      }
  
@@ -185,7 +185,7 @@ index 13b96d61f..5fa163442 100644
  
      GGML_LOG_INFO("ggml-hex: Hexagon Arch version v%d\n", opt_arch);
  
-@@ -2895,6 +2917,8 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
+@@ -3026,6 +3048,8 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
          } catch (const std::exception & exc) {
              GGML_LOG_ERROR("ggml-hex: failed to create device/session %zu\n", i);
              devices[i].context = nullptr;

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();