@fugood/llama.node 0.3.7 → 0.3.9

package/README.md

@@ -4,9 +4,23 @@
 [![NPM Version](https://img.shields.io/npm/v/%40fugood%2Fllama.node)](https://www.npmjs.com/package/@fugood/llama.node)
 ![NPM Downloads](https://img.shields.io/npm/dw/%40fugood%2Fllama.node)
 
-Node binding of [llama.cpp](https://github.com/ggerganov/llama.cpp).
+An another Node binding of [llama.cpp](https://github.com/ggerganov/llama.cpp) to make same API with [llama.rn](https://github.com/mybigday/llama.rn) as much as possible.
 
-[llama.cpp](https://github.com/ggerganov/llama.cpp): Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
+- [llama.cpp](https://github.com/ggerganov/llama.cpp): Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
+- [llama.rn](https://github.com/mybigday/llama.rn): React Native binding of llama.cpp
+
+## Platform Support
+
+- macOS
+  - arm64: CPU and Metal GPU acceleration
+  - x86_64: CPU only
+- Windows (x86_64 and arm64)
+  - CPU
+  - GPU acceleration via Vulkan
+- Linux (x86_64 and arm64)
+  - CPU
+  - GPU acceleration via Vulkan
+  - GPU acceleration via CUDA
 
 ## Installation
 
@@ -49,6 +63,7 @@ console.log('Result:', text)
 
 - [x] `default`: General usage, not support GPU except macOS (Metal)
 - [x] `vulkan`: Support GPU Vulkan (Windows/Linux), but some scenario might unstable
+- [x] `cuda`: Support GPU CUDA (Linux), but only for limited capability (x86_64: 8.9, arm64: 8.7)
 
 ## License
 

package/lib/binding.ts

@@ -37,6 +37,9 @@ export type LlamaModelOptions = {
   use_mlock?: boolean
   use_mmap?: boolean
   vocab_only?: boolean
+  lora?: string
+  lora_scaled?: number
+  lora_list?: { path: string; scaled: number }[]
 }
 
 export type LlamaCompletionOptions = {
@@ -111,6 +114,11 @@ export interface LlamaContext {
   saveSession(path: string): Promise<void>
   loadSession(path: string): Promise<void>
   release(): Promise<void>
+  applyLoraAdapters(adapters: { path: string; scaled: number }[]): void
+  removeLoraAdapters(adapters: { path: string }[]): void
+  getLoadedLoraAdapters(): { path: string; scaled: number }[]
+  // static
+  loadModelInfo(path: string, skip: string[]): Promise<Object>
 }
 
 export interface Module {

package/lib/index.js

@@ -23,7 +23,7 @@ var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, ge
     });
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.loadModel = void 0;
+exports.loadLlamaModelInfo = exports.initLlama = exports.loadModel = void 0;
 const binding_1 = require("./binding");
 __exportStar(require("./binding"), exports);
 const mods = {};
@@ -34,3 +34,18 @@ const loadModel = (options) => __awaiter(void 0, void 0, void 0, function* () {
     return new mods[variant].LlamaContext(options);
 });
 exports.loadModel = loadModel;
+exports.initLlama = binding_1.loadModule;
+const modelInfoSkip = [
+    // Large fields
+    'tokenizer.ggml.tokens',
+    'tokenizer.ggml.token_type',
+    'tokenizer.ggml.merges',
+    'tokenizer.ggml.scores',
+];
+const loadLlamaModelInfo = (path) => __awaiter(void 0, void 0, void 0, function* () {
+    var _a;
+    const variant = 'default';
+    (_a = mods[variant]) !== null && _a !== void 0 ? _a : (mods[variant] = yield (0, binding_1.loadModule)(variant));
+    return mods[variant].LlamaContext.loadModelInfo(path, modelInfoSkip);
+});
+exports.loadLlamaModelInfo = loadLlamaModelInfo;

package/lib/index.ts

@@ -14,3 +14,19 @@ export const loadModel = async (options: LlamaModelOptionsExtended): Promise<Lla
   mods[variant] ??= await loadModule(options.lib_variant)
   return new mods[variant].LlamaContext(options)
 }
+
+export const initLlama = loadModule
+
+const modelInfoSkip = [
+  // Large fields
+  'tokenizer.ggml.tokens',
+  'tokenizer.ggml.token_type',
+  'tokenizer.ggml.merges',
+  'tokenizer.ggml.scores',
+]
+
+export const loadLlamaModelInfo = async (path: string): Promise<Object> => {
+  const variant = 'default'
+  mods[variant] ??= await loadModule(variant)
+  return mods[variant].LlamaContext.loadModelInfo(path, modelInfoSkip)
+}

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@fugood/llama.node",
   "access": "public",
-  "version": "0.3.7",
+  "version": "0.3.9",
   "description": "Llama.cpp for Node.js",
   "main": "lib/index.js",
   "scripts": {

package/src/EmbeddingWorker.cpp

@@ -9,10 +9,11 @@ void EmbeddingWorker::Execute() {
   llama_kv_cache_clear(_sess->context());
   auto tokens = ::common_tokenize(_sess->context(), _text, true);
   // add SEP if not present
-  if (tokens.empty() || tokens.back() != llama_token_sep(_sess->model())) {
-    tokens.push_back(llama_token_sep(_sess->model()));
+  auto vocab = llama_model_get_vocab(_sess->model());
+  if (tokens.empty() || tokens.back() != llama_vocab_sep(vocab)) {
+    tokens.push_back(llama_vocab_sep(vocab));
   }
-  const int n_embd = llama_n_embd(_sess->model());
+  const int n_embd = llama_model_n_embd(_sess->model());
   do {
     auto ctx = _sess->context();
     int ret =

package/src/LlamaCompletionWorker.cpp

@@ -59,7 +59,9 @@ void LlamaCompletionWorker::Execute() {
   size_t n_cur = 0;
   size_t n_input = 0;
   const auto model = _sess->model();
-  const bool add_bos = llama_add_bos_token(model);
+  auto vocab = llama_model_get_vocab(model);
+
+  const bool add_bos = llama_vocab_get_add_bos(vocab);
   auto ctx = _sess->context();
 
   auto sparams = llama_sampler_chain_default_params();
@@ -130,7 +132,7 @@ void LlamaCompletionWorker::Execute() {
       });
     }
     // is it an end of generation?
-    if (llama_token_is_eog(model, new_token_id)) {
+    if (llama_vocab_is_eog(vocab, new_token_id)) {
       break;
     }
     // check for stop words

package/src/LlamaContext.cpp

@@ -1,4 +1,6 @@
 #include "ggml.h"
+#include "gguf.h"
+#include "llama-impl.h"
 #include "LlamaContext.h"
 #include "DetokenizeWorker.h"
 #include "DisposeWorker.h"
@@ -8,6 +10,56 @@
 #include "SaveSessionWorker.h"
 #include "TokenizeWorker.h"
 
+// loadModelInfo(path: string): object
+Napi::Value LlamaContext::ModelInfo(const Napi::CallbackInfo& info) {
+  Napi::Env env = info.Env();
+  struct gguf_init_params params = {
+    /*.no_alloc = */ false,
+    /*.ctx      = */ NULL,
+  };
+  std::string path = info[0].ToString().Utf8Value();
+  
+  // Convert Napi::Array to vector<string>
+  std::vector<std::string> skip;
+  if (info.Length() > 1 && info[1].IsArray()) {
+    Napi::Array skipArray = info[1].As<Napi::Array>();
+    for (uint32_t i = 0; i < skipArray.Length(); i++) {
+      skip.push_back(skipArray.Get(i).ToString().Utf8Value());
+    }
+  }
+
+  struct gguf_context * ctx = gguf_init_from_file(path.c_str(), params);
+
+  Napi::Object metadata = Napi::Object::New(env);
+  if (std::find(skip.begin(), skip.end(), "version") == skip.end()) {
+    metadata.Set("version", Napi::Number::New(env, gguf_get_version(ctx)));
+  }
+  if (std::find(skip.begin(), skip.end(), "alignment") == skip.end()) {
+    metadata.Set("alignment", Napi::Number::New(env, gguf_get_alignment(ctx)));
+  }
+  if (std::find(skip.begin(), skip.end(), "data_offset") == skip.end()) {
+    metadata.Set("data_offset", Napi::Number::New(env, gguf_get_data_offset(ctx)));
+  }
+
+  // kv
+  {
+    const int n_kv = gguf_get_n_kv(ctx);
+
+    for (int i = 0; i < n_kv; ++i) {
+      const char * key = gguf_get_key(ctx, i);
+      if (std::find(skip.begin(), skip.end(), key) != skip.end()) {
+        continue;
+      }
+      const std::string value = gguf_kv_to_str(ctx, i);
+      metadata.Set(key, Napi::String::New(env, value.c_str()));
+    }
+  }
+
+  gguf_free(ctx);
+
+  return metadata;
+}
+
 std::vector<common_chat_msg> get_messages(Napi::Array messages) {
   std::vector<common_chat_msg> chat;
   for (size_t i = 0; i < messages.Length(); i++) {
@@ -51,8 +103,20 @@ void LlamaContext::Init(Napi::Env env, Napi::Object &exports) {
        InstanceMethod<&LlamaContext::LoadSession>(
            "loadSession",
            static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::ApplyLoraAdapters>(
+           "applyLoraAdapters",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::RemoveLoraAdapters>(
+           "removeLoraAdapters",
+           static_cast<napi_property_attributes>(napi_enumerable)),
+       InstanceMethod<&LlamaContext::GetLoadedLoraAdapters>(
+           "getLoadedLoraAdapters",
+           static_cast<napi_property_attributes>(napi_enumerable)),
        InstanceMethod<&LlamaContext::Release>(
-           "release", static_cast<napi_property_attributes>(napi_enumerable))});
+           "release", static_cast<napi_property_attributes>(napi_enumerable)),
+       StaticMethod<&LlamaContext::ModelInfo>(
+           "loadModelInfo",
+           static_cast<napi_property_attributes>(napi_enumerable))});
   Napi::FunctionReference *constructor = new Napi::FunctionReference();
   *constructor = Napi::Persistent(func);
 #if NAPI_VERSION > 5
@@ -140,14 +204,56 @@ LlamaContext::LlamaContext(const Napi::CallbackInfo &info)
   llama_backend_init();
   llama_numa_init(params.numa);
 
-  auto result = common_init_from_params(params);
+  auto sess = std::make_shared<LlamaSession>(params);
 
-  if (result.model == nullptr || result.context == nullptr) {
+  if (sess->model() == nullptr || sess->context() == nullptr) {
     Napi::TypeError::New(env, "Failed to load model")
         .ThrowAsJavaScriptException();
   }
 
-  _sess = std::make_shared<LlamaSession>(result.model, result.context, params);
+  auto ctx = sess->context();
+  auto model = sess->model();
+
+  std::vector<common_adapter_lora_info> lora;
+  auto lora_path = get_option<std::string>(options, "lora", "");
+  auto lora_scaled = get_option<float>(options, "lora_scaled", 1.0f);
+  if (lora_path != "") {
+    common_adapter_lora_info la;
+    la.path = lora_path;
+    la.scale = lora_scaled;
+    la.ptr = llama_adapter_lora_init(model, lora_path.c_str());
+    if (la.ptr == nullptr) {
+      Napi::TypeError::New(env, "Failed to load lora adapter")
+          .ThrowAsJavaScriptException();
+    }
+    lora.push_back(la);
+  }
+
+  if (options.Has("lora_list") && options.Get("lora_list").IsArray()) {
+    auto lora_list = options.Get("lora_list").As<Napi::Array>();
+    if (lora_list != nullptr) {
+      int lora_list_size = lora_list.Length();
+      for (int i = 0; i < lora_list_size; i++) {
+        auto lora_adapter = lora_list.Get(i).As<Napi::Object>();
+        auto path = lora_adapter.Get("path").ToString();
+        if (path != nullptr) {
+          common_adapter_lora_info la;
+          la.path = path;
+          la.scale = lora_adapter.Get("scaled").ToNumber().FloatValue();
+          la.ptr = llama_adapter_lora_init(model, path.Utf8Value().c_str());
+          if (la.ptr == nullptr) {
+            Napi::TypeError::New(env, "Failed to load lora adapter")
+                .ThrowAsJavaScriptException();
+          }
+          lora.push_back(la);
+        }
+      }
+    }
+  }
+  common_set_adapter_lora(ctx, lora);
+  _lora = lora;
+
+  _sess = sess;
   _info = common_params_get_system_info(params);
 }
 
@@ -162,8 +268,8 @@ bool validateModelChatTemplate(const struct llama_model * model) {
     int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
     if (res >= 0) {
         llama_chat_message chat[] = {{"user", "test"}};
-        std::string tmpl = std::string(model_template.data(), model_template.size());
-        int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
+        const char * tmpl = llama_model_chat_template(model);
+        int32_t chat_res = llama_chat_apply_template(tmpl, chat, 1, true, nullptr, 0);
         return chat_res > 0;
     }
     return res > 0;
@@ -187,6 +293,7 @@ Napi::Value LlamaContext::GetModelInfo(const Napi::CallbackInfo &info) {
   }
   Napi::Object details = Napi::Object::New(info.Env());
   details.Set("desc", desc);
+  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("isChatTemplateSupported", validateModelChatTemplate(model));
@@ -396,6 +503,49 @@ Napi::Value LlamaContext::LoadSession(const Napi::CallbackInfo &info) {
   return worker->Promise();
 }
 
+// applyLoraAdapters(lora_adapters: [{ path: string, scaled: number }]): void
+void LlamaContext::ApplyLoraAdapters(const Napi::CallbackInfo &info) {
+  Napi::Env env = info.Env();
+  std::vector<common_adapter_lora_info> lora;
+  auto lora_adapters = info[0].As<Napi::Array>();
+  for (size_t i = 0; i < lora_adapters.Length(); i++) {
+    auto lora_adapter = lora_adapters.Get(i).As<Napi::Object>();
+    auto path = lora_adapter.Get("path").ToString().Utf8Value();
+    auto scaled = lora_adapter.Get("scaled").ToNumber().FloatValue();
+    common_adapter_lora_info la;
+    la.path = path;
+    la.scale = scaled;
+    la.ptr = llama_adapter_lora_init(_sess->model(), path.c_str());
+    if (la.ptr == nullptr) {
+      Napi::TypeError::New(env, "Failed to load lora adapter")
+          .ThrowAsJavaScriptException();
+    }
+    lora.push_back(la);
+  }
+  common_set_adapter_lora(_sess->context(), lora);
+  _lora = lora;
+}
+
+// removeLoraAdapters(): void
+void LlamaContext::RemoveLoraAdapters(const Napi::CallbackInfo &info) {
+  _lora.clear();
+  common_set_adapter_lora(_sess->context(), _lora);
+}
+
+// getLoadedLoraAdapters(): Promise<{ count, lora_adapters: [{ path: string,
+// scaled: number }] }>
+Napi::Value LlamaContext::GetLoadedLoraAdapters(const Napi::CallbackInfo &info) {
+  Napi::Env env = info.Env();
+  Napi::Array lora_adapters = Napi::Array::New(env, _lora.size());
+  for (size_t i = 0; i < _lora.size(); i++) {
+    Napi::Object lora_adapter = Napi::Object::New(env);
+    lora_adapter.Set("path", _lora[i].path);
+    lora_adapter.Set("scaled", _lora[i].scale);
+    lora_adapters.Set(i, lora_adapter);
+  }
+  return lora_adapters;
+}
+
 // release(): Promise<void>
 Napi::Value LlamaContext::Release(const Napi::CallbackInfo &info) {
   auto env = info.Env();

package/src/LlamaContext.h

@@ -5,6 +5,7 @@ class LlamaCompletionWorker;
 class LlamaContext : public Napi::ObjectWrap<LlamaContext> {
 public:
   LlamaContext(const Napi::CallbackInfo &info);
+  static Napi::Value ModelInfo(const Napi::CallbackInfo& info);
   static void Init(Napi::Env env, Napi::Object &exports);
 
 private:
@@ -18,10 +19,14 @@ private:
   Napi::Value Embedding(const Napi::CallbackInfo &info);
   Napi::Value SaveSession(const Napi::CallbackInfo &info);
   Napi::Value LoadSession(const Napi::CallbackInfo &info);
+  void ApplyLoraAdapters(const Napi::CallbackInfo &info);
+  void RemoveLoraAdapters(const Napi::CallbackInfo &info);
+  Napi::Value GetLoadedLoraAdapters(const Napi::CallbackInfo &info);
   Napi::Value Release(const Napi::CallbackInfo &info);
 
   std::string _info;
   Napi::Object _meta;
   LlamaSessionPtr _sess = nullptr;
+  std::vector<common_adapter_lora_info> _lora;
   LlamaCompletionWorker *_wip = nullptr;
 };

package/src/common.hpp

@@ -11,8 +11,6 @@
 #include <tuple>
 #include <vector>
 
-typedef std::unique_ptr<llama_model, decltype(&llama_free_model)> LlamaCppModel;
-typedef std::unique_ptr<llama_context, decltype(&llama_free)> LlamaCppContext;
 typedef std::unique_ptr<common_sampler, decltype(&common_sampler_free)>
     LlamaCppSampling;
 typedef std::unique_ptr<llama_batch, decltype(&llama_batch_free)> LlamaCppBatch;
@@ -47,17 +45,17 @@ constexpr T get_option(const Napi::Object &options, const std::string &name,
 
 class LlamaSession {
 public:
-  LlamaSession(llama_model *model, llama_context *ctx, common_params params)
-      : model_(LlamaCppModel(model, llama_free_model)),
-        ctx_(LlamaCppContext(ctx, llama_free)), params_(params) {
+  LlamaSession(common_params params)
+      : params_(params) {
+    llama_init_ = common_init_from_params(params);
     tokens_.reserve(params.n_ctx);
   }
 
   ~LlamaSession() { dispose(); }
 
-  inline llama_context *context() { return ctx_.get(); }
+  inline llama_context *context() { return llama_init_.context.get(); }
 
-  inline llama_model *model() { return model_.get(); }
+  inline llama_model *model() { return llama_init_.model.get(); }
 
   inline std::vector<llama_token> *tokens_ptr() { return &tokens_; }
 
@@ -72,13 +70,10 @@ public:
   void dispose() {
     std::lock_guard<std::mutex> lock(mutex);
     tokens_.clear();
-    ctx_.reset();
-    model_.reset();
   }
 
 private:
-  LlamaCppModel model_;
-  LlamaCppContext ctx_;
+  common_init_result llama_init_;
   const common_params params_;
   std::vector<llama_token> tokens_{};
   std::mutex mutex;

package/src/llama.cpp/.github/workflows/build.yml

@@ -60,8 +60,7 @@ jobs:
             -DLLAMA_CURL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DGGML_RPC=ON \
-            -DBUILD_SHARED_LIBS=OFF
+            -DGGML_RPC=ON
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -88,6 +87,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -123,8 +123,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_CURL=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
-            -DBUILD_SHARED_LIBS=OFF
+            -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -151,6 +150,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -181,7 +181,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -219,6 +219,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
           zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
 
       - name: Upload artifacts
@@ -236,7 +237,7 @@ jobs:
     strategy:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
+        build_type: [Debug]
 
     steps:
       - name: Clone
@@ -651,23 +652,23 @@ jobs:
       matrix:
         include:
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'msvc-arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
@@ -798,6 +799,7 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
+          Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -914,7 +916,7 @@ jobs:
         shell: cmd
         run: |
           call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
-          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DGGML_RPC=ON
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
@@ -1239,7 +1241,7 @@ jobs:
 
       - name: Create release
         id: create_release
-        uses: anzz1/action-create-release@v1
+        uses: ggml-org/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with: