fix(Vulkan): read external memory usage (#500)

* fix(Vulkan): read external memory usage
* fix(Electron template): minor performance improvement

Author: giladgd

.github/workflows/build.yml

@@ -133,7 +133,7 @@ jobs:
       - name: Install Cuda 13.0 on Windows (1)
         if: matrix.config.name == 'Windows (1)'
         shell: bash
-        timeout-minutes: 30
+        timeout-minutes: 60
         run: |
           curl -Lo cuda_13.0.0_windows_network.exe https://developer.download.nvidia.com/compute/cuda/13.0.0/network_installers/cuda_13.0.0_windows_network.exe
           

llama/addon/globals/addonLog.h

@@ -2,6 +2,7 @@
 #include "llama.h"
 #include "napi.h"
 
+
 struct addon_logger_log {
     public:
         const int logLevelNumber;

llama/gpuInfo/vulkan-gpu-info.cpp

@@ -42,18 +42,17 @@ static bool enumerateVulkanDevices(size_t* total, size_t* used, size_t* unifiedM
             physicalDevice.getMemoryProperties2(&memProps2);
 
             for (uint32_t i = 0; i < memProps.memoryHeapCount; ++i) {
-                const auto flags = memProps.memoryHeaps[i].flags;
+                const auto heap = memProps2.memoryProperties.memoryHeaps[i];
 
-                if (flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
-                    const auto size = memProps.memoryHeaps[i].size;
-                    totalMem += size;
-                    usedMem += memoryBudgetProperties.heapUsage[i];
+                if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
+                    totalMem += heap.size;
+                    usedMem += memoryBudgetProperties.heapUsage[i] + (heap.size - memoryBudgetProperties.heapBudget[i]);
 
-                    if (flags & vk::MemoryHeapFlagBits::eMultiInstance) {
-                        totalUnifiedMemorySize += size;
+                    if (heap.flags & vk::MemoryHeapFlagBits::eMultiInstance) {
+                        totalUnifiedMemorySize += heap.size;
                     }
 
-                    if (size > 0 && addDeviceNames) {
+                    if (heap.size > 0 && addDeviceNames) {
                         (*deviceNames).push_back(std::string(deviceProps.deviceName.data()));
                     }
 

templates/electron-typescript-react/electron/state/llmState.ts

@@ -307,6 +307,7 @@ export const llmFunctions = {
 
                     try {
                         await chatSession?.preloadPrompt("", {
+                            functions: modelFunctions, // these won't be called, but are used to avoid redundant context shifts
                             signal: promptAbortController?.signal
                         });
                     } catch (err) {