From d8c43a6258e84fd03d08fcdc88a30ad2b7d5336c Mon Sep 17 00:00:00 2001 From: alanshurafa <138934986+alanshurafa@users.noreply.github.com> Date: Fri, 6 Nov 2026 23:19:03 -0500 Subject: [PATCH 1/2] vulkan: create the device that was registered, not the index's ggml-vulkan cached device_indices at instance init and re-enumerated physical devices later to create the device. When the loader's order changes in between (seen on Windows with an iGPU and a dGPU), the same index picks the other GPU while the logs still name the first. Keep the handles from the first enumeration and use them everywhere. Also add the change as patches/ggml/0004 so sync-ggml.sh keeps it. Co-Authored-By: Claude Opus --- ggml/UPSTREAM | 1 + ggml/src/ggml-vulkan/ggml-vulkan-types.h | 3 + ggml/src/ggml-vulkan/ggml-vulkan.cpp | 43 +++--- ...-vulkan-keep-physical-device-handles.patch | 125 ++++++++++++++++++ 4 files changed, 148 insertions(+), 24 deletions(-) create mode 100644 patches/ggml/0004-vulkan-keep-physical-device-handles.patch diff --git a/ggml/UPSTREAM b/ggml/UPSTREAM index 923e32df..1c998cca 100644 --- a/ggml/UPSTREAM +++ b/ggml/UPSTREAM @@ -4,6 +4,7 @@ patches: patches/ggml/0001-fix-threadpool-oversubscription.patch patches/ggml/0002-backend-reg-filter.patch patches/ggml/0003-metal-mapped-buffer-null.patch + patches/ggml/0004-vulkan-keep-physical-device-handles.patch This directory is generated from the upstream ggml tree at the SHA above, minus .github/, with the listed downstream patches applied in order. Do not edit it by diff --git a/ggml/src/ggml-vulkan/ggml-vulkan-types.h b/ggml/src/ggml-vulkan/ggml-vulkan-types.h index 4158ff77..8251e0a9 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan-types.h +++ b/ggml/src/ggml-vulkan/ggml-vulkan-types.h @@ -1345,6 +1345,9 @@ struct vk_instance_t { PFN_vkCmdInsertDebugUtilsLabelEXT pfn_vkCmdInsertDebugUtilsLabelEXT = {}; std::vector device_indices; + // physical device handles from the enumeration that produced device_indices; + // the loader's order can change later (e.g. on Windows), the handles do not + std::vector physical_devices; std::vector device_supports_membudget; vk_device devices[GGML_VK_MAX_DEVICES]; }; diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 5587dfc2..ebc1c24a 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -3999,16 +3999,14 @@ vk_device ggml_vk_get_device(size_t idx) { device->memory_logger = std::unique_ptr(new vk_memory_logger()); - size_t dev_num = vk_instance.device_indices[idx]; - - std::vector physical_devices = vk_instance.instance.enumeratePhysicalDevices(); - - if (dev_num >= physical_devices.size()) { - std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl; + if (idx >= vk_instance.physical_devices.size()) { + std::cerr << "ggml_vulkan: Device with index " << idx << " does not exist." << std::endl; throw std::runtime_error("Device not found"); } - device->physical_device = physical_devices[dev_num]; + // use the handle from instance init rather than re-enumerating: the loader's + // order can change in between, and device_indices would then pick another GPU + device->physical_device = vk_instance.physical_devices[idx]; const std::vector ext_props = device->physical_device.enumerateDeviceExtensionProperties(); device->architecture = get_device_architecture(device->physical_device); @@ -4933,15 +4931,9 @@ static void ggml_vk_print_gpu_info(size_t idx) { size_t dev_num = vk_instance.device_indices[idx]; VK_LOG_DEBUG("ggml_vk_print_gpu_info(" << dev_num << ")"); GGML_ASSERT(vk_instance_initialized); + GGML_ASSERT(idx < vk_instance.physical_devices.size()); - std::vector devices = vk_instance.instance.enumeratePhysicalDevices(); - - if (dev_num >= devices.size()) { - std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl; - throw std::runtime_error("Device not found"); - } - - vk::PhysicalDevice physical_device = devices[dev_num]; + vk::PhysicalDevice physical_device = vk_instance.physical_devices[idx]; std::vector ext_props = physical_device.enumerateDeviceExtensionProperties(); bool fp16_storage = false; @@ -5410,10 +5402,15 @@ void ggml_vk_instance_init() { return; } } + + for (size_t idx : vk_instance.device_indices) { + vk_instance.physical_devices.push_back(devices[idx]); + } + GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size()); for (size_t i = 0; i < vk_instance.device_indices.size(); i++) { - vk::PhysicalDevice vkdev = devices[vk_instance.device_indices[i]]; + vk::PhysicalDevice vkdev = vk_instance.physical_devices[i]; std::vector extensionprops = vkdev.enumerateDeviceExtensionProperties(); bool membudget_supported = false; @@ -12903,11 +12900,10 @@ int ggml_vk_get_device_count() { void ggml_vk_get_device_description(int device, char * description, size_t description_size) { ggml_vk_instance_init(); - - std::vector devices = vk_instance.instance.enumeratePhysicalDevices(); + GGML_ASSERT(device >= 0 && device < (int) vk_instance.physical_devices.size()); vk::PhysicalDeviceProperties props; - devices[device].getProperties(&props); + vk_instance.physical_devices[device].getProperties(&props); snprintf(description, description_size, "%s", props.deviceName.data()); } @@ -15181,15 +15177,14 @@ int ggml_backend_vk_get_device_count() { void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) { GGML_ASSERT(device < (int) vk_instance.device_indices.size()); - int dev_idx = vk_instance.device_indices[device]; - ggml_vk_get_device_description(dev_idx, description, description_size); + ggml_vk_get_device_description(device, description, description_size); } void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) { GGML_ASSERT(device < (int) vk_instance.device_indices.size()); GGML_ASSERT(device < (int) vk_instance.device_supports_membudget.size()); - vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]]; + vk::PhysicalDevice vkdev = vk_instance.physical_devices[device]; vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops; vk::PhysicalDeviceMemoryProperties2 memprops = {}; const bool membudget_supported = vk_instance.device_supports_membudget[device]; @@ -15221,7 +15216,7 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total static vk::PhysicalDeviceType ggml_backend_vk_get_device_type(int device_idx) { GGML_ASSERT(device_idx >= 0 && device_idx < (int) vk_instance.device_indices.size()); - vk::PhysicalDevice device = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device_idx]]; + vk::PhysicalDevice device = vk_instance.physical_devices[device_idx]; vk::PhysicalDeviceProperties2 props = {}; device.getProperties2(&props); @@ -15232,7 +15227,7 @@ static vk::PhysicalDeviceType ggml_backend_vk_get_device_type(int device_idx) { static std::string ggml_backend_vk_get_device_pci_id(int device_idx) { GGML_ASSERT(device_idx >= 0 && device_idx < (int) vk_instance.device_indices.size()); - vk::PhysicalDevice device = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device_idx]]; + vk::PhysicalDevice device = vk_instance.physical_devices[device_idx]; const std::vector ext_props = device.enumerateDeviceExtensionProperties(); diff --git a/patches/ggml/0004-vulkan-keep-physical-device-handles.patch b/patches/ggml/0004-vulkan-keep-physical-device-handles.patch new file mode 100644 index 00000000..8d008b90 --- /dev/null +++ b/patches/ggml/0004-vulkan-keep-physical-device-handles.patch @@ -0,0 +1,125 @@ +diff --git a/src/ggml-vulkan/ggml-vulkan-types.h b/src/ggml-vulkan/ggml-vulkan-types.h +index 4158ff77..8251e0a9 100644 +--- a/src/ggml-vulkan/ggml-vulkan-types.h ++++ b/src/ggml-vulkan/ggml-vulkan-types.h +@@ -1345,6 +1345,9 @@ struct vk_instance_t { + PFN_vkCmdInsertDebugUtilsLabelEXT pfn_vkCmdInsertDebugUtilsLabelEXT = {}; + + std::vector device_indices; ++ // physical device handles from the enumeration that produced device_indices; ++ // the loader's order can change later (e.g. on Windows), the handles do not ++ std::vector physical_devices; + std::vector device_supports_membudget; + vk_device devices[GGML_VK_MAX_DEVICES]; + }; +diff --git a/src/ggml-vulkan/ggml-vulkan.cpp b/src/ggml-vulkan/ggml-vulkan.cpp +index 5587dfc2..ebc1c24a 100644 +--- a/src/ggml-vulkan/ggml-vulkan.cpp ++++ b/src/ggml-vulkan/ggml-vulkan.cpp +@@ -3999,16 +3999,14 @@ vk_device ggml_vk_get_device(size_t idx) { + + device->memory_logger = std::unique_ptr(new vk_memory_logger()); + +- size_t dev_num = vk_instance.device_indices[idx]; +- +- std::vector physical_devices = vk_instance.instance.enumeratePhysicalDevices(); +- +- if (dev_num >= physical_devices.size()) { +- std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl; ++ if (idx >= vk_instance.physical_devices.size()) { ++ std::cerr << "ggml_vulkan: Device with index " << idx << " does not exist." << std::endl; + throw std::runtime_error("Device not found"); + } + +- device->physical_device = physical_devices[dev_num]; ++ // use the handle from instance init rather than re-enumerating: the loader's ++ // order can change in between, and device_indices would then pick another GPU ++ device->physical_device = vk_instance.physical_devices[idx]; + const std::vector ext_props = device->physical_device.enumerateDeviceExtensionProperties(); + + device->architecture = get_device_architecture(device->physical_device); +@@ -4933,15 +4931,9 @@ static void ggml_vk_print_gpu_info(size_t idx) { + size_t dev_num = vk_instance.device_indices[idx]; + VK_LOG_DEBUG("ggml_vk_print_gpu_info(" << dev_num << ")"); + GGML_ASSERT(vk_instance_initialized); ++ GGML_ASSERT(idx < vk_instance.physical_devices.size()); + +- std::vector devices = vk_instance.instance.enumeratePhysicalDevices(); +- +- if (dev_num >= devices.size()) { +- std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl; +- throw std::runtime_error("Device not found"); +- } +- +- vk::PhysicalDevice physical_device = devices[dev_num]; ++ vk::PhysicalDevice physical_device = vk_instance.physical_devices[idx]; + std::vector ext_props = physical_device.enumerateDeviceExtensionProperties(); + + bool fp16_storage = false; +@@ -5410,10 +5402,15 @@ void ggml_vk_instance_init() { + return; + } + } ++ ++ for (size_t idx : vk_instance.device_indices) { ++ vk_instance.physical_devices.push_back(devices[idx]); ++ } ++ + GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size()); + + for (size_t i = 0; i < vk_instance.device_indices.size(); i++) { +- vk::PhysicalDevice vkdev = devices[vk_instance.device_indices[i]]; ++ vk::PhysicalDevice vkdev = vk_instance.physical_devices[i]; + std::vector extensionprops = vkdev.enumerateDeviceExtensionProperties(); + + bool membudget_supported = false; +@@ -12903,11 +12900,10 @@ int ggml_vk_get_device_count() { + + void ggml_vk_get_device_description(int device, char * description, size_t description_size) { + ggml_vk_instance_init(); +- +- std::vector devices = vk_instance.instance.enumeratePhysicalDevices(); ++ GGML_ASSERT(device >= 0 && device < (int) vk_instance.physical_devices.size()); + + vk::PhysicalDeviceProperties props; +- devices[device].getProperties(&props); ++ vk_instance.physical_devices[device].getProperties(&props); + + snprintf(description, description_size, "%s", props.deviceName.data()); + } +@@ -15181,15 +15177,14 @@ int ggml_backend_vk_get_device_count() { + + void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) { + GGML_ASSERT(device < (int) vk_instance.device_indices.size()); +- int dev_idx = vk_instance.device_indices[device]; +- ggml_vk_get_device_description(dev_idx, description, description_size); ++ ggml_vk_get_device_description(device, description, description_size); + } + + void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) { + GGML_ASSERT(device < (int) vk_instance.device_indices.size()); + GGML_ASSERT(device < (int) vk_instance.device_supports_membudget.size()); + +- vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]]; ++ vk::PhysicalDevice vkdev = vk_instance.physical_devices[device]; + vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops; + vk::PhysicalDeviceMemoryProperties2 memprops = {}; + const bool membudget_supported = vk_instance.device_supports_membudget[device]; +@@ -15221,7 +15216,7 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total + static vk::PhysicalDeviceType ggml_backend_vk_get_device_type(int device_idx) { + GGML_ASSERT(device_idx >= 0 && device_idx < (int) vk_instance.device_indices.size()); + +- vk::PhysicalDevice device = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device_idx]]; ++ vk::PhysicalDevice device = vk_instance.physical_devices[device_idx]; + + vk::PhysicalDeviceProperties2 props = {}; + device.getProperties2(&props); +@@ -15232,7 +15227,7 @@ static vk::PhysicalDeviceType ggml_backend_vk_get_device_type(int device_idx) { + static std::string ggml_backend_vk_get_device_pci_id(int device_idx) { + GGML_ASSERT(device_idx >= 0 && device_idx < (int) vk_instance.device_indices.size()); + +- vk::PhysicalDevice device = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device_idx]]; ++ vk::PhysicalDevice device = vk_instance.physical_devices[device_idx]; + + const std::vector ext_props = device.enumerateDeviceExtensionProperties(); + From db1fbb6c9a9ce63ee2be36b30765791e1ec12c3a Mon Sep 17 00:00:00 2001 From: alanshurafa <138934986+alanshurafa@users.noreply.github.com> Date: Sat, 7 Nov 2026 02:45:48 -0500 Subject: [PATCH 2/2] Renumber Vulkan patch to 0005 so it does not collide with #246's 0004 Co-Authored-By: Claude Opus 5.5 --- ggml/UPSTREAM | 2 +- ...les.patch => 0005-vulkan-keep-physical-device-handles.patch} | 0 2 files changed, 1 insertion(+), 1 deletion(-) rename patches/ggml/{0004-vulkan-keep-physical-device-handles.patch => 0005-vulkan-keep-physical-device-handles.patch} (100%) diff --git a/ggml/UPSTREAM b/ggml/UPSTREAM index 1c998cca..45e8a03a 100644 --- a/ggml/UPSTREAM +++ b/ggml/UPSTREAM @@ -4,7 +4,7 @@ patches: patches/ggml/0001-fix-threadpool-oversubscription.patch patches/ggml/0002-backend-reg-filter.patch patches/ggml/0003-metal-mapped-buffer-null.patch - patches/ggml/0004-vulkan-keep-physical-device-handles.patch + patches/ggml/0005-vulkan-keep-physical-device-handles.patch This directory is generated from the upstream ggml tree at the SHA above, minus .github/, with the listed downstream patches applied in order. Do not edit it by diff --git a/patches/ggml/0004-vulkan-keep-physical-device-handles.patch b/patches/ggml/0005-vulkan-keep-physical-device-handles.patch similarity index 100% rename from patches/ggml/0004-vulkan-keep-physical-device-handles.patch rename to patches/ggml/0005-vulkan-keep-physical-device-handles.patch