ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

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

44
ggml/src/ggml-metal.m
View file

@ -3247,12 +3247,6 @@ static enum ggml_status ggml_metal_graph_compute(
// backend interface
static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
UNUSED(buffer);
}
static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
@ -3307,7 +3301,6 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
}
static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .get_name = */ ggml_backend_metal_buffer_get_name,
/* .free_buffer = */ ggml_backend_metal_buffer_free_buffer,
/* .get_base = */ ggml_backend_metal_buffer_get_base,
/* .init_tensor = */ NULL,
@ -3432,6 +3425,29 @@ ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
return &ggml_backend_buffer_type_metal;
}
static const char * ggml_backend_metal_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
return "Metal_Mapped";
UNUSED(buft);
}
static ggml_backend_buffer_type_t ggml_backend_metal_buffer_from_ptr_type(void) {
static struct ggml_backend_buffer_type ggml_backend_buffer_from_ptr_type_metal = {
/* .iface = */ {
/* .get_name = */ ggml_backend_metal_buffer_from_ptr_type_get_name,
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_metal_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_metal_buffer_type_is_host,
},
/* .device = */ &g_ggml_backend_metal_device,
/* .context = */ NULL,
};
return &ggml_backend_buffer_from_ptr_type_metal;
}
// TODO: obsoleted by ggml_backend_metal_device_buffer_from_ptr
ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
@ -3508,7 +3524,7 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
}
}
return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
return ggml_backend_buffer_init(ggml_backend_metal_buffer_from_ptr_type(), ggml_backend_metal_buffer_i, ctx, size);
}
// backend
@ -3529,12 +3545,6 @@ static void ggml_backend_metal_free(ggml_backend_t backend) {
free(backend);
}
static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_metal_buffer_type();
UNUSED(backend);
}
static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
return ggml_metal_graph_compute(backend, cgraph);
}
@ -3601,7 +3611,6 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_name = */ ggml_backend_metal_name,
/* .free = */ ggml_backend_metal_free,
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
@ -3611,9 +3620,6 @@ static struct ggml_backend_i ggml_backend_metal_i = {
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_metal_graph_compute,
/* .supports_op = */ NULL,
/* .supports_buft = */ NULL,
/* .offload_op = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
@ -3708,7 +3714,7 @@ static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t
}
static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
return GGML_BACKEND_DEVICE_TYPE_GPU;
GGML_UNUSED(dev);
}