[BUG]: Vulkan Backend Often Falls Back to CPU Due to `vulkaninfo --summary` Timeout

[LSXAxeller]

Description

When using the Vulkan backend on my machine, it frequently falls back to using the CPU for inference. This seems to be happening because we're not giving vulkaninfo enough time to complete when trying to detect the GPU.

Currently, we're using TimeSpan.FromSeconds(1) as the timeout for the vulkaninfo --summary command in GetVulkanSummary(). However, on my machine, running this command takes about 1.92 seconds on average. This means that most of the time, the command times out before it can return the necessary information, leading to the GPU not being detected and the fallback to CPU inference.

I did a quick test using a batch script to measure the execution time of vulkaninfo --summary (see code snippets below). This confirmed that the command consistently takes longer than 1 second to complete on my system.

To address this, I simply increased the timeout to 10 seconds (TimeSpan.FromSeconds(10)). This seems to have resolved the issue, as the GPU is now consistently detected and used for inference.

I would appreciate if we got this changed in next release.

Code Snippets:

Current GetVulkanSummary() function:

private static string? GetVulkanSummary()
        {
            // Note: on Linux, this requires `vulkan-tools` to be installed. (`sudo apt install vulkan-tools`)
            try
            {
                // Start a process to read vulkan info
                Process process = new()
                {
                    StartInfo = new()
                    {
                        FileName = "vulkaninfo",
                        Arguments = "--summary",
                        RedirectStandardOutput = true,
                        RedirectStandardError = true,
                        UseShellExecute = false,
                        CreateNoWindow = true
                    }
                };
                var (exitCode, output, error, ok) = process.SafeRun(TimeSpan.FromSeconds(1)); // Timeout is 1 second which needs to be increased

                if (!ok)
                    return null;

                // Return the output
                return output;
            }
            catch
            {
                // Return null if we failed to get the Vulkan version
                return null;
            }
        }

Batch script for measuring vulkaninfo execution time:

@echo off
set start=%time%

vulkaninfo --summary

set end=%time%

echo Start Time: %start%
echo End Time: %end%

REM Calculate the difference in time
set /a diff_seconds=((1%end:~0,2%-1%start:~0,2%)*3600 + (1%end:~3,2%-1%start:~3,2%)*60 + (1%end:~6,2%-1%start:~6,2%)) %% 86400
set /a diff_milliseconds=((1%end:~9,2%-1%start:~9,2%) + (1%end:~6,2%-1%start:~6,2%) * 100 + (1%end:~3,2%-1%start:~3,2%)*6000 + (1%end:~0,2%-1%start:~0,2%)*360000) %% 8640000

echo.
echo Execution time: %diff_seconds%.%diff_milliseconds:~-3% seconds

pause

Example output from the batch script:

==========
VULKANINFO
==========

Vulkan Instance Version: 1.3.250


Instance Extensions: count = 13
-------------------------------
VK_EXT_debug_report                    : extension revision 10
VK_EXT_debug_utils                     : extension revision 2
VK_EXT_swapchain_colorspace            : extension revision 4
VK_KHR_device_group_creation           : extension revision 1
VK_KHR_external_fence_capabilities     : extension revision 1
VK_KHR_external_memory_capabilities    : extension revision 1
VK_KHR_external_semaphore_capabilities : extension revision 1
VK_KHR_get_physical_device_properties2 : extension revision 2
VK_KHR_get_surface_capabilities2       : extension revision 1
VK_KHR_portability_enumeration         : extension revision 1
VK_KHR_surface                         : extension revision 25
VK_KHR_win32_surface                   : extension revision 6
VK_LUNARG_direct_driver_loading        : extension revision 1

Instance Layers: count = 16
---------------------------
VK_LAYER_AMD_switchable_graphics  AMD switchable graphics layer                                                                                     1.3.260  version 1
VK_LAYER_EOS_Overlay              Vulkan overlay layer for Epic Online Services                                                                     1.2.136  version 1
VK_LAYER_EOS_Overlay              Vulkan overlay layer for Epic Online Services
                            1.2.136  version 1
VK_LAYER_KHRONOS_profiles         Khronos Profiles layer
                            1.3.290  version 1
VK_LAYER_KHRONOS_shader_object    Khronos Shader object layer
                            1.3.290  version 1
VK_LAYER_KHRONOS_synchronization2 Khronos Synchronization2 layer
                            1.3.290  version 1
VK_LAYER_KHRONOS_validation       Khronos Validation Layer
                            1.3.290  version 1
VK_LAYER_LUNARG_api_dump          LunarG API dump layer
                            1.3.290  version 2
VK_LAYER_LUNARG_crash_diagnostic  Crash Diagnostic Layer is a crash/hang debugging tool that helps determines GPU progress in a Vulkan application. 1.3.290  version 1
VK_LAYER_LUNARG_gfxreconstruct    GFXReconstruct Capture Layer Version 1.0.5
                            1.3.290  version 4194309
VK_LAYER_LUNARG_monitor           Execution Monitoring Layer
                            1.3.290  version 1
VK_LAYER_LUNARG_screenshot        LunarG image capture layer
                            1.3.290  version 1
VK_LAYER_OBS_HOOK                 Open Broadcaster Software hook
                            1.3.216  version 1
VK_LAYER_RTSS                     RTSS overlay hook bootstrap
                            1.3.224  version 1
VK_LAYER_VALVE_steam_fossilize    Steam Pipeline Caching Layer
                            1.3.207  version 1
VK_LAYER_VALVE_steam_overlay      Steam Overlay Layer
                            1.3.207  version 1

Devices:
========
GPU0:
        apiVersion         = 1.3.260
        driverVersion      = 2.0.279
        vendorID           = 0x1002
        deviceID           = 0x67df
        deviceType         = PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
        deviceName         = Radeon RX 580 Series
        driverID           = DRIVER_ID_AMD_PROPRIETARY
        driverName         = AMD proprietary driver
        driverInfo         = 24.1.1 (AMD proprietary shader compiler)
        conformanceVersion = 1.3.3.1
        deviceUUID         = 00000000-0100-0000-0000-000000000000
        driverUUID         = 414d442d-5749-4e2d-4452-560000000000

Start Time: 15:34:33.59
End Time: 15:34:34.51

Execution time: 1.92 seconds
Press any key to continue . . .

Reproduction Steps

• Install Vulkan backend
• Enable backend

	NativeLibraryConfig.All
            .WithCuda(false)
            .WithVulkan(true)
            .WithAutoFallback();

• Load any model

Environment & Configuration

• Operating system: Windows 11
• .NET runtime version: 8.0
• LLamaSharp version: v0.16.0
• CPU & GPU device: I5-11400F & Radeon RX 580 4GB

Known Workarounds

No response

[martindevans]

Thanks for the detailed investigation! If you'd like to open a PR with this change I'll be to happy to review and merge it :)

[LSXPrime]

Thanks for the detailed investigation! If you'd like to open a PR with this change I'll be to happy to review and merge it :)

Thanks! I've just opened a pull request with the proposed change: #931 . Please take a look when you have a chance.