fix: future-proof llama.cpp backend loading

Author: giladgd

llama/addon/addon.cpp

@@ -152,7 +152,17 @@ class AddonBackendUnloadWorker : public Napi::AsyncWorker {
 };
 
 Napi::Value addonLoadBackends(const Napi::CallbackInfo& info) {
-    ggml_backend_load_all();
+    const bool forceLoadLibraries = info.Length() == 0
+        ? false
+        : info[0].IsBoolean()
+            ? info[0].As<Napi::Boolean>().Value()
+            : false;
+
+    ggml_backend_reg_count();
+
+    if (forceLoadLibraries) {
+        ggml_backend_load_all();
+    }
 
     return info.Env().Undefined();
 }

src/bindings/AddonTypes.ts

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

src/bindings/Llama.ts

@@ -127,6 +127,9 @@ export class Llama {
         }
 
         this._bindings.loadBackends();
+        const loadedGpu = bindings.getGpuType();
+        if (loadedGpu == null || (loadedGpu === false && gpu !== false))
+            this._bindings.loadBackends(true);
 
         this._onExit = this._onExit.bind(this);
 

src/bindings/utils/testBindingBinary.ts

@@ -196,7 +196,12 @@ if (process.env.TEST_BINDING_CP === "true" && (process.parentPort != null || pro
         if (message.type === "start") {
             try {
                 const binding: BindingModule = require(message.bindingBinaryPath);
+
                 binding.loadBackends();
+                const loadedGpu = binding.getGpuType();
+                if (loadedGpu == null || (loadedGpu === false && message.gpu !== false))
+                    binding.loadBackends(true);
+
                 await binding.init();
                 binding.getGpuVramInfo();
                 binding.getGpuDeviceInfo();