mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

vulkan: add noncontiguous GLU support (#21081)

Browse Source 
* vulkan: add noncontiguous GLU support

* fix compile issue
This commit is contained in:
Ruben Ortlam
2026-03-28 08:44:56 +01:00
committed by GitHub
parent 1f5d15e665
commit 0eb4764182
3 changed files with 53 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+27 -12
View File
@@ -1112,6 +1112,16 @@ struct vk_op_glu_push_constants {
uint32_t mode; // 0: default, 1: swapped, 2: split
float alpha; // for swiglu_oai
float limit;
uint32_t nb01;
uint32_t nb02;
uint32_t nb03;
uint32_t ne01;
uint32_t ne02;
uint32_t nb11;
uint32_t nb12;
uint32_t nb13;
uint32_t ne11;
uint32_t ne12;
};
struct vk_op_unary_push_constants {
@@ -5044,7 +5054,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
} else {
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
}
vk::DeviceCreateInfo device_create_info;
vk::DeviceCreateInfo device_create_info{};
std::vector<const char *> device_extensions;
vk::PhysicalDeviceFeatures device_features = device->physical_device.getFeatures();
@@ -5413,12 +5423,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
#endif
device->name = GGML_VK_NAME + std::to_string(idx);
device_create_info = {
vk::DeviceCreateFlags(),
device_queue_create_infos,
{},
device_extensions
};
device_create_info
.setFlags(vk::DeviceCreateFlags())
.setQueueCreateInfos(device_queue_create_infos)
.setPEnabledExtensionNames(device_extensions);
device_create_info.setPNext(&device_features2);
device->device = device->physical_device.createDevice(device_create_info);
@@ -11048,8 +11056,6 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
const float alpha = op_params_f[2];
const float limit = op_params_f[3];
GGML_ASSERT(ggml_is_contiguous(src0));
if (!split) {
GGML_ASSERT(src0->ne[0] / 2 == dst->ne[0]);
} else {
@@ -11067,7 +11073,17 @@ static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const
(uint32_t)dst->ne[0],
mode,
alpha,
limit
limit,
(uint32_t)(src0->nb[1] / src0->nb[0]),
(uint32_t)(src0->nb[2] / src0->nb[0]),
(uint32_t)(src0->nb[3] / src0->nb[0]),
(uint32_t)src0->ne[1],
(uint32_t)src0->ne[2],
(uint32_t)(dst->nb[1] / dst->nb[0]),
(uint32_t)(dst->nb[2] / dst->nb[0]),
(uint32_t)(dst->nb[3] / dst->nb[0]),
(uint32_t)dst->ne[1],
(uint32_t)dst->ne[2]
});
}
@@ -15217,8 +15233,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_GLU_OP_SWIGLU_OAI:
case GGML_GLU_OP_GEGLU_ERF:
case GGML_GLU_OP_GEGLU_QUICK:
return ggml_is_contiguous(op->src[0]) &&
(op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
(op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
(op->src[0]->type == op->type);
default: