@fugood/llama.node 1.0.0-beta.2 → 1.0.0-beta.3

package/package.json

@@ -1,15 +1,17 @@
 {
   "name": "@fugood/llama.node",
   "access": "public",
-  "version": "1.0.0-beta.2",
+  "version": "1.0.0-beta.3",
   "description": "An another Node binding of llama.cpp",
   "main": "lib/index.js",
   "scripts": {
-    "postinstall": "node scripts/build.js",
+    "bootstrap": "npm install --omit=optional",
+    "postinstall": "node scripts/check.js",
     "pretest": "node scripts/download-test-models.js",
     "test": "jest",
-    "build": "tsc",
-    "prepack": "npm run build",
+    "build": "npx cmake-js build",
+    "build-js": "tsc",
+    "prepack": "npm run build-js",
     "prebuild-native": "node scripts/generate_win_dynamic_load.js 6",
     "build-native": "cmake-js compile",
     "clean": "rimraf build",
@@ -47,7 +49,7 @@
     ]
   },
   "files": [
-    "scripts/build.js",
+    "scripts/check.js",
     "scripts/llama.cpp.patch",
     "src/*.{cc,c,h,hpp}",
     "src/DecodeAudioTokenWorker.cpp",
@@ -68,19 +70,19 @@
     "CMakeLists.txt"
   ],
   "optionalDependencies": {
-    "@fugood/node-llama-linux-x64": "1.0.0-beta.2",
-    "@fugood/node-llama-linux-x64-vulkan": "1.0.0-beta.2",
-    "@fugood/node-llama-linux-x64-cuda": "1.0.0-beta.2",
-    "@fugood/node-llama-linux-arm64": "1.0.0-beta.2",
-    "@fugood/node-llama-linux-arm64-vulkan": "1.0.0-beta.2",
-    "@fugood/node-llama-linux-arm64-cuda": "1.0.0-beta.2",
-    "@fugood/node-llama-win32-x64": "1.0.0-beta.2",
-    "@fugood/node-llama-win32-x64-vulkan": "1.0.0-beta.2",
-    "@fugood/node-llama-win32-x64-cuda": "1.0.0-beta.2",
-    "@fugood/node-llama-win32-arm64": "1.0.0-beta.2",
-    "@fugood/node-llama-win32-arm64-vulkan": "1.0.0-beta.2",
-    "@fugood/node-llama-darwin-x64": "1.0.0-beta.2",
-    "@fugood/node-llama-darwin-arm64": "1.0.0-beta.2"
+    "@fugood/node-llama-linux-x64": "1.0.0-beta.3",
+    "@fugood/node-llama-linux-x64-vulkan": "1.0.0-beta.3",
+    "@fugood/node-llama-linux-x64-cuda": "1.0.0-beta.3",
+    "@fugood/node-llama-linux-arm64": "1.0.0-beta.3",
+    "@fugood/node-llama-linux-arm64-vulkan": "1.0.0-beta.3",
+    "@fugood/node-llama-linux-arm64-cuda": "1.0.0-beta.3",
+    "@fugood/node-llama-win32-x64": "1.0.0-beta.3",
+    "@fugood/node-llama-win32-x64-vulkan": "1.0.0-beta.3",
+    "@fugood/node-llama-win32-x64-cuda": "1.0.0-beta.3",
+    "@fugood/node-llama-win32-arm64": "1.0.0-beta.3",
+    "@fugood/node-llama-win32-arm64-vulkan": "1.0.0-beta.3",
+    "@fugood/node-llama-darwin-x64": "1.0.0-beta.3",
+    "@fugood/node-llama-darwin-arm64": "1.0.0-beta.3"
   },
   "devDependencies": {
     "@babel/preset-env": "^7.24.4",
@@ -92,14 +94,12 @@
     "cmake-js": "^7.3.0",
     "husky": "^9.0.11",
     "jest": "^29.7.0",
+    "node-addon-api": "^8.0.0",
     "release-it": "^17.7.0",
     "rimraf": "^6.0.1",
     "typescript": "^5.4.5",
     "wait-for-expect": "^3.0.2"
   },
-  "dependencies": {
-    "node-addon-api": "^8.0.0"
-  },
   "jest": {
     "testEnvironment": "node",
     "moduleFileExtensions": [

package/scripts/{build.js → check.js}

@@ -1,31 +1,18 @@
-const fs = require('fs');
 const path = require('path');
 
 const validAccelerators = process.platform === 'darwin' ? [] : ['vulkan', 'cuda'];
 
-let isBuildFromSource = process.env.npm_config_build_from_source === 'true';
+const accelerator = process.env.npm_config_accelerator || '';
 
-let accelerator = process.env.npm_config_accelerator || '';
-
-const checkPaths = [
-  path.resolve(
-    __dirname,
-    `../node-llama-${process.platform}-${process.arch}${accelerator ? `-${accelerator}` : ''}`
-  ),
-  path.resolve(__dirname, `../build/Release/index.node`),
-];
-
-if (!isBuildFromSource && !checkPaths.some(path => fs.existsSync(path))) {
-  console.warn('Not found prebuild package, please build from source');
-  isBuildFromSource = true;
+if (process.env.npm_config_build_from_source) {
+  console.log('Build from source is enabled');
+} else {
+  process.exit(0);
 }
 
 if (accelerator && !validAccelerators.includes(accelerator)) {
-  throw new Error(`Invalid accelerator: ${accelerator}`);
-}
-
-if (isBuildFromSource) {
-  console.log('Build from source is enabled');
+  console.error(`Invalid accelerator: ${accelerator}`);
+  process.exit(1);
 }
 
 let BuildSystem;

package/src/llama.cpp/common/chat.cpp

@@ -1,8 +1,6 @@
 #include "chat.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "minja/chat-template.hpp"
-#include "minja/minja.hpp"
 
 #include <optional>
 
@@ -15,14 +13,6 @@ static std::string format_time(const std::chrono::system_clock::time_point & now
     return res;
 }
 
-typedef minja::chat_template common_chat_template;
-
-struct common_chat_templates {
-    bool has_explicit_template; // Model had builtin template or template overridde was specified.
-    std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
-    std::unique_ptr<common_chat_template> template_tool_use;
-};
-
 struct templates_params {
     json messages;
     json tools;

package/src/llama.cpp/common/chat.h

@@ -6,8 +6,16 @@
 #include <chrono>
 #include <string>
 #include <vector>
+#include "minja/chat-template.hpp"
+#include "minja/minja.hpp"
 
-struct common_chat_templates;
+typedef minja::chat_template common_chat_template;
+
+struct common_chat_templates {
+    bool has_explicit_template; // Model had builtin template or template overridde was specified.
+    std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
+    std::unique_ptr<common_chat_template> template_tool_use;
+};
 
 struct common_chat_tool_call {
     std::string name;

package/src/llama.cpp/common/common.cpp

@@ -1081,6 +1081,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
 
+    mparams.vocab_only      = params.vocab_only;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;

package/src/llama.cpp/common/common.h

@@ -219,6 +219,7 @@ enum common_reasoning_format {
 };
 
 struct common_params {
+    bool vocab_only               = false;
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)

package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt

@@ -90,7 +90,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         message(STATUS "ARM detected")
 
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
-            message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
+            list(APPEND ARCH_FLAGS /arch:armv8.7)
         else()
             check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
             if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")