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

kompute : use llama_backend_init/llama_backend_free to manage device

Browse source
This commit is contained in:
Jared Van Bortel 2024-01-27 11:55:32 -05:00
parent 050d450297
commit 530462550d
2 changed files with 11 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-kompute.cpp
View file

@ -1863,9 +1863,6 @@ static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
static void ggml_backend_kompute_free(ggml_backend_t backend) { static void ggml_backend_kompute_free(ggml_backend_t backend) {
struct ggml_kompute_context * ctx = (struct ggml_kompute_context *)backend->context; struct ggml_kompute_context * ctx = (struct ggml_kompute_context *)backend->context;
ggml_vk_free(ctx); ggml_vk_free(ctx);
// TODO(cebtenzzre): This should only be done if the device was initialized by us, but
// that would require a change to GPT4All.
ggml_vk_free_device();
delete backend; delete backend;
} }
@ -1901,12 +1898,6 @@ static struct ggml_backend_i kompute_backend_i = {
}; };
ggml_backend_t ggml_backend_kompute_init() { ggml_backend_t ggml_backend_kompute_init() {
#if defined(GGML_USE_KOMPUTE)
if (!ggml_vk_has_device()) {
ggml_vk_init_device(0, "gpu");
}
#endif
if (!ggml_vk_has_device()) { if (!ggml_vk_has_device()) {
fprintf(stderr, "%s: error: device was not initialized\n", __func__); fprintf(stderr, "%s: error: device was not initialized\n", __func__);
return nullptr; return nullptr;

11
llama.cpp
View file

@ -9860,6 +9860,13 @@ void llama_backend_init(bool numa) {
#ifdef GGML_USE_MPI #ifdef GGML_USE_MPI
ggml_mpi_backend_init(); ggml_mpi_backend_init();
#endif #endif
#ifdef GGML_USE_KOMPUTE
if (!ggml_vk_has_device()) {
ggml_vk_init_device(0, "gpu");
}
#endif
} }
void llama_backend_free(void) { void llama_backend_free(void) {
@ -9867,6 +9874,10 @@ void llama_backend_free(void) {
ggml_mpi_backend_free(); ggml_mpi_backend_free();
#endif #endif
ggml_quantize_free(); ggml_quantize_free();
#ifdef GGML_USE_KOMPUTE
ggml_vk_free_device();
#endif
} }
int64_t llama_time_us(void) { int64_t llama_time_us(void) {