vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : refactor model loader with backend registry (#10026)

Browse source
This commit is contained in:
Diego Devesa 2024-10-30 02:01:23 +01:00 • committed by GitHub
parent 8f275a7c45
commit c5b0f4b5d9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
18 changed files with 1903 additions and 2019 deletions
Download patch file Download diff file Expand all files Collapse all files

26
ggml/src/ggml-vulkan.cpp
View file

@ -6247,13 +6247,8 @@ static void ggml_vk_get_device_description(int device, char * description, size_
// device backend
static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
return ctx->name.c_str();
}
static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
return buffer->buft->iface.get_name == ggml_backend_vk_buffer_type_name;
}
static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
@ -6317,7 +6312,6 @@ static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t v
}
static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
/* .get_name = */ ggml_backend_vk_buffer_get_name,
/* .free_buffer = */ ggml_backend_vk_buffer_free_buffer,
/* .get_base = */ ggml_backend_vk_buffer_get_base,
/* .init_tensor = */ ggml_backend_vk_buffer_init_tensor,
@ -6413,7 +6407,6 @@ static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_vk_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_vk_host_buffer_free_buffer;
return buffer;
@ -6646,7 +6639,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
static ggml_backend_i ggml_backend_vk_interface = {
/* .get_name = */ ggml_backend_vk_name,
/* .free = */ ggml_backend_vk_free,
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
@ -6656,9 +6648,6 @@ static ggml_backend_i ggml_backend_vk_interface = {
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_vk_graph_compute,
/* .supports_op = */ NULL,
/* .supports_buft = */ NULL,
/* .offload_op = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
@ -6717,7 +6706,7 @@ void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total
//////////////////////////
struct ggml_backend_vk_device_context {
int device;
size_t device;
std::string name;
std::string description;
};
@ -6749,7 +6738,7 @@ static ggml_backend_buffer_type_t ggml_backend_vk_device_get_host_buffer_type(gg
static enum ggml_backend_dev_type ggml_backend_vk_device_get_type(ggml_backend_dev_t dev) {
UNUSED(dev);
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
return GGML_BACKEND_DEVICE_TYPE_GPU;
}
static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
@ -6758,9 +6747,10 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
props->type = ggml_backend_vk_device_get_type(dev);
ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* async */ false,
/* host_buffer */ true,
/* events */ false,
/* .async = */ false,
/* .host_buffer = */ true,
/* .buffer_from_host_ptr = */ false,
/* .events = */ false,
};
}
@ -6949,7 +6939,7 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
for (size_t i = 0; i < ggml_backend_vk_get_device_count(); i++) {
for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
char desc[256];
ggml_backend_vk_get_device_description(i, desc, sizeof(desc));