mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2024-12-26 00:29:21 +01:00
ggml : add max buffer sizes to opencl and metal backends (llama/5181)
This commit is contained in:
parent
cb9de61659
commit
44e517f074
12
ggml-metal.m
12
ggml-metal.m
@ -2398,6 +2398,16 @@ GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backen
|
||||
UNUSED(buft);
|
||||
}
|
||||
|
||||
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
|
||||
id<MTLDevice> device = ggml_backend_metal_get_device();
|
||||
size_t max_size = device.maxBufferLength;
|
||||
ggml_backend_metal_free_device();
|
||||
|
||||
return max_size;
|
||||
|
||||
UNUSED(buft);
|
||||
}
|
||||
|
||||
GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
|
||||
return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
|
||||
|
||||
@ -2416,7 +2426,7 @@ GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
|
||||
/* .get_name = */ ggml_backend_metal_buffer_type_get_name,
|
||||
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
|
||||
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
|
||||
/* .get_max_size = */ NULL, // TODO: return device.maxBufferLength
|
||||
/* .get_max_size = */ ggml_backend_metal_buffer_type_get_max_size,
|
||||
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
|
||||
/* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
|
||||
/* .is_host = */ ggml_backend_metal_buffer_type_is_host,
|
||||
|
@ -2125,6 +2125,15 @@ static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_
|
||||
GGML_UNUSED(buffer_type);
|
||||
}
|
||||
|
||||
static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
|
||||
static size_t max_size = -1;
|
||||
if (max_size == (size_t)-1) {
|
||||
ggml_cl_init();
|
||||
clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_size, NULL);
|
||||
}
|
||||
return max_size;
|
||||
}
|
||||
|
||||
static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
|
||||
//return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
|
||||
return ggml_backend_is_cpu(backend);
|
||||
@ -2136,7 +2145,7 @@ static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
|
||||
/* .get_name = */ ggml_backend_opencl_buffer_type_name,
|
||||
/* .alloc_buffer = */ ggml_backend_opencl_buffer_type_alloc_buffer,
|
||||
/* .get_alignment = */ ggml_backend_opencl_buffer_type_get_alignment,
|
||||
/* .get_max_size = */ NULL, // TODO: return from device info
|
||||
/* .get_max_size = */ ggml_backend_opencl_buffer_type_get_max_size,
|
||||
/* .get_alloc_size = */ NULL,
|
||||
/* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
|
||||
/* .is_host = */ NULL,
|
||||
|
Loading…
Reference in New Issue
Block a user