fix: adapt to breaking llama.cpp changes

Author: giladgd

llama/addon/addon.cpp

@@ -151,6 +151,12 @@ class AddonBackendUnloadWorker : public Napi::AsyncWorker {
         }
 };
 
+Napi::Value addonLoadBackends(const Napi::CallbackInfo& info) {
+    ggml_backend_load_all();
+
+    return info.Env().Undefined();
+}
+
 Napi::Value addonInit(const Napi::CallbackInfo& info) {
     if (backendInitialized) {
         Napi::Promise::Deferred deferred = Napi::Promise::Deferred::New(info.Env());
@@ -205,6 +211,7 @@ Napi::Object registerCallback(Napi::Env env, Napi::Object exports) {
         Napi::PropertyDescriptor::Function("getGpuDeviceInfo", getGpuDeviceInfo),
         Napi::PropertyDescriptor::Function("getGpuType", getGpuType),
         Napi::PropertyDescriptor::Function("getSwapInfo", getSwapInfo),
+        Napi::PropertyDescriptor::Function("loadBackends", addonLoadBackends),
         Napi::PropertyDescriptor::Function("init", addonInit),
         Napi::PropertyDescriptor::Function("dispose", addonDispose),
     });
@@ -215,7 +222,6 @@ Napi::Object registerCallback(Napi::Env env, Napi::Object exports) {
     AddonContext::init(exports);
     AddonSampler::init(exports);
 
-    ggml_backend_load_all();
     llama_log_set(addonLlamaCppLogCallback, nullptr);
 
     exports.AddFinalizer(addonFreeLlamaBackend, static_cast<int*>(nullptr));

src/bindings/AddonTypes.ts

@@ -76,6 +76,7 @@ export type BindingModule = {
         free: number
     },
     init(): Promise<void>,
+    loadBackends(): void,
     dispose(): Promise<void>
 };
 

src/bindings/Llama.ts

@@ -126,6 +126,8 @@ export class Llama {
             this._bindings.setLoggerLogLevel(LlamaLogLevelToAddonLogLevel.get(this._logLevel) ?? defaultLogLevel);
         }
 
+        this._bindings.loadBackends();
+
         this._onExit = this._onExit.bind(this);
 
         process.on("exit", this._onExit);
@@ -605,6 +607,8 @@ function getTransformedLogLevel(level: LlamaLogLevel, message: string): LlamaLog
         return LlamaLogLevel.info;
     else if (level === LlamaLogLevel.warn && message.startsWith("ggml_metal_init: skipping kernel_") && message.endsWith("(not supported)"))
         return LlamaLogLevel.log;
+    else if (level === LlamaLogLevel.warn && message.startsWith("ggml_cuda_init: GGML_CUDA_FORCE_") && message.endsWith(" no"))
+        return LlamaLogLevel.log;
 
     return level;
 }

src/bindings/utils/testBindingBinary.ts

@@ -196,6 +196,7 @@ if (process.env.TEST_BINDING_CP === "true" && (process.parentPort != null || pro
         if (message.type === "start") {
             try {
                 const binding: BindingModule = require(message.bindingBinaryPath);
+                binding.loadBackends();
                 await binding.init();
                 binding.getGpuVramInfo();
                 binding.getGpuDeviceInfo();