happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: fix GPU deduplication logic. 

As reported in https://github.com/ggml-org/llama.cpp/issues/19221, the
(same uuid, same driver) logic is problematic for windows+intel igpu.

Let's just avoid filtering for MoltenVK which is apple-specific, and
keep the logic the  same as before 88d23ad5 - just dedup based on UUID.

Verified that MacOS + 4xVega still reports 4 GPUs with this version.
This commit is contained in:
Oleksandr Kuvshynov 2026-01-30 21:12:05 -05:00
parent ec6c7421e4
commit 9676add968
1 changed files with 8 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -5538,18 +5538,19 @@ static void ggml_vk_instance_init() {
// Check if there are two physical devices corresponding to the same GPU
// This handles the case where the same GPU appears with different drivers (e.g., RADV + AMDVLK on Linux),
// see https://github.com/ggml-org/llama.cpp/pull/7582 for original deduplication.
// However, for MoltenVK on macOS, multiple GPUs on the same card may report the same UUID,
// see https://github.com/KhronosGroup/MoltenVK/issues/2683. Until this is fixed, we'll only deduplicate
// when drivers differ (same driver + same UUID = likely different GPUs)
// MoltenVK on macOS may report the same UUID for distinct GPUs on multi-GPU cards,
// see https://github.com/KhronosGroup/MoltenVK/issues/2683. Skip dedup for MoltenVK.
auto old_device = std::find_if(
vk_instance.device_indices.begin(),
vk_instance.device_indices.end(),
[&devices, &new_id, &new_driver](const size_t k){
if (new_driver.driverID == vk::DriverId::eMoltenvk) {
return false;
}
vk::PhysicalDeviceProperties2 old_props;
vk::PhysicalDeviceDriverProperties old_driver;
vk::PhysicalDeviceIDProperties old_id;
old_props.pNext = &old_driver;
old_driver.pNext = &old_id;
old_props.pNext = &old_id;
devices[k].getProperties2(&old_props);
bool same_uuid = std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID));
@ -5558,10 +5559,7 @@ static void ggml_vk_instance_init() {
std::equal(std::begin(old_id.deviceLUID), std::end(old_id.deviceLUID), std::begin(new_id.deviceLUID))
);
// Only deduplicate if same UUID AND different drivers
// (same driver + same UUID on MoltenVK = likely different GPUs on multi-GPU card)
bool different_driver = (old_driver.driverID != new_driver.driverID);
return same_uuid && different_driver;
return same_uuid;
}
);
if (old_device == vk_instance.device_indices.end()) {