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

ggml-vulkan: clamp subgroup size to 64 on Qualcomm to prevent instability

This commit is contained in:
Gong-Mi 2026-01-23 21:04:37 +08:00
parent 322605cb62
commit 8e98e723eb
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -4703,6 +4703,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);
device->subgroup_size = subgroup_props.subgroupSize;
if (device->vendor_id == VK_VENDOR_ID_QUALCOMM && device->subgroup_size > 64) {
VK_LOG_INFO("ggml_vulkan: clamping subgroup size to 64 for Qualcomm");
device->subgroup_size = 64;
}
device->subgroup_size_log2 = uint32_t(log2f(float(device->subgroup_size)));
device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
if (sm_builtins) {