fix: prevent loading Vulkan if the device is unsupported

Author: giladgd

llama/addon/addon.cpp

@@ -237,6 +237,7 @@ Napi::Object registerCallback(Napi::Env env, Napi::Object exports) {
         Napi::PropertyDescriptor::Function("getGpuVramInfo", getGpuVramInfo),
         Napi::PropertyDescriptor::Function("getGpuDeviceInfo", getGpuDeviceInfo),
         Napi::PropertyDescriptor::Function("getGpuType", getGpuType),
+        Napi::PropertyDescriptor::Function("ensureGpuDeviceIsSupported", ensureGpuDeviceIsSupported),
         Napi::PropertyDescriptor::Function("getSwapInfo", getSwapInfo),
         Napi::PropertyDescriptor::Function("getMemoryInfo", getMemoryInfo),
         Napi::PropertyDescriptor::Function("loadBackends", addonLoadBackends),

llama/addon/globals/getGpuInfo.cpp

@@ -128,3 +128,14 @@ Napi::Value getGpuType(const Napi::CallbackInfo& info) {
 
     return info.Env().Undefined();
 }
+
+Napi::Value ensureGpuDeviceIsSupported(const Napi::CallbackInfo& info) {
+#ifdef GPU_INFO_USE_VULKAN
+    if (!checkIsVulkanEnvSupported()) {
+        Napi::Error::New(info.Env(), "Vulkan device is not supported").ThrowAsJavaScriptException();
+        return info.Env().Undefined();
+    }
+#endif
+
+    return info.Env().Undefined();
+}

llama/addon/globals/getGpuInfo.h

@@ -7,4 +7,5 @@
 Napi::Value getGpuVramInfo(const Napi::CallbackInfo& info);
 Napi::Value getGpuDeviceInfo(const Napi::CallbackInfo& info);
 std::pair<ggml_backend_dev_t, std::string> getGpuDevice();
-Napi::Value getGpuType(const Napi::CallbackInfo& info);
+Napi::Value getGpuType(const Napi::CallbackInfo& info);
+Napi::Value ensureGpuDeviceIsSupported(const Napi::CallbackInfo& info);

llama/gpuInfo/vulkan-gpu-info.cpp

@@ -80,3 +80,13 @@ static bool enumerateVulkanDevices(size_t* total, size_t* used, size_t* unifiedM
 bool gpuInfoGetTotalVulkanDevicesInfo(size_t* total, size_t* used, size_t* unifiedMemorySize, gpuInfoVulkanWarningLogCallback_t warningLogCallback) {
     return enumerateVulkanDevices(total, used, unifiedMemorySize, false, nullptr, warningLogCallback);
 }
+
+bool checkIsVulkanEnvSupported() {
+    VkPhysicalDeviceVulkan11Features vk11_features;
+
+    if (!vk11_features.storageBuffer16BitAccess) {
+        return false;
+    }
+
+    return true;
+}

llama/gpuInfo/vulkan-gpu-info.h

@@ -6,3 +6,4 @@
 typedef void (*gpuInfoVulkanWarningLogCallback_t)(const char* message);
 
 bool gpuInfoGetTotalVulkanDevicesInfo(size_t* total, size_t* used, size_t* unifiedMemorySize, gpuInfoVulkanWarningLogCallback_t warningLogCallback);
+bool checkIsVulkanEnvSupported();

src/bindings/AddonTypes.ts

@@ -72,6 +72,7 @@ export type BindingModule = {
         deviceNames: string[]
     },
     getGpuType(): "cuda" | "vulkan" | "metal" | false | undefined,
+    ensureGpuDeviceIsSupported(): void,
     getSwapInfo(): {
         total: number,
         maxSize: number,

src/bindings/Llama.ts

@@ -108,6 +108,8 @@ export class Llama {
         if (loadedGpu == null || (loadedGpu === false && buildGpu !== false))
             bindings.loadBackends(path.dirname(bindingPath));
 
+        bindings.ensureGpuDeviceIsSupported();
+
         this._gpu = bindings.getGpuType() ?? false;
         this._supportsGpuOffloading = bindings.getSupportsGpuOffloading();
         this._supportsMmap = bindings.getSupportsMmap();

src/bindings/utils/testBindingBinary.ts

@@ -211,6 +211,8 @@ if (process.env.TEST_BINDING_CP === "true" && (process.parentPort != null || pro
                 if (gpuType !== message.gpu)
                     throw new Error(`Binary GPU type mismatch. Expected: ${message.gpu}, got: ${gpuType}`);
 
+                binding.ensureGpuDeviceIsSupported();
+
                 sendMessage({type: "done"});
             } catch (err) {
                 console.error(err);