upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-04 09:00:48 -05:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Introduce driver version check for Windows Intel GPU to mitigate crashing (#25192)

Browse Source 
* Removed crash guard for Intel

Crash fixed from driver 32.0.101.8860

* Added driver version check for windows

* Change to convert from driverVersion rather than string

* No need to use signed

* Refactor

* allow GPU other than Xe2+

* adjusted function body position
This commit is contained in:
Masato Nakasaka
2026-07-31 23:26:37 +09:00
committed by GitHub
parent db7d8b24b5
commit eb41d503ba
1 changed files with 23 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -1934,6 +1934,7 @@ struct ggml_vk_garbage_collector {
static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_context subctx);
static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested = nullptr);
static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor);
static bool vk_memory_logger_enabled = false;
@@ -5552,8 +5553,11 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
// Intel Arc B390 was observed segfaulting with this shader.
if (device->subgroup_basic && device->subgroup_shuffle && device->vendor_id != VK_VENDOR_ID_INTEL) {
// Intel Windows driver older than 32.0.101.8860 will crash when using fwht kernels on Xe2+ GPUS so we gate that here
const bool can_use_fwht = device->driver_id != vk::DriverId::eIntelProprietaryWindows ||
device->architecture != vk_device_architecture::INTEL_XE2 ||
(device->architecture == vk_device_architecture::INTEL_XE2 && ggml_vk_intel_windows_driver_equals_or_newer_than(device->properties.driverVersion, 101, 8860));
if (can_use_fwht && device->subgroup_basic && device->subgroup_shuffle) {
int idx = 0;
for (uint32_t n : {64, 128, 256, 512}) {
if (device->subgroup_size <= n) {
@@ -5561,8 +5565,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
}
++idx;
}
} else if (device->driver_id != vk::DriverId::eIntelProprietaryWindows) {
// Disabled on Intel Windows due to a driver bug: https://github.com/ggml-org/llama.cpp/pull/23964#issuecomment-4598226147
} else if (can_use_fwht) {
int idx = 0;
for (uint32_t n : {64, 128, 256, 512}) {
const uint32_t block_size = std::min(device->subgroup_size, n);
@@ -18466,6 +18469,22 @@ static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev)
}
}
static bool ggml_vk_intel_windows_driver_equals_or_newer_than(uint32_t driver_version, uint32_t threshold_major, uint32_t threshold_minor) {
#if defined(_WIN32)
// Intel Windows encodes xxx.yyyy as [31:14].[13:0].
const uint32_t major = driver_version >> 14;
const uint32_t minor = driver_version & 0x3fff;
return major > threshold_major || (major == threshold_major && minor >= threshold_minor);
#else
GGML_UNUSED(driver_version);
GGML_UNUSED(threshold_major);
GGML_UNUSED(threshold_minor);
return true;
#endif
}
// checks
#ifdef GGML_VULKAN_CHECK_RESULTS