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

kompute : init device automatically and remove an unnecessary free

Browse source
This commit is contained in:
Jared Van Bortel 2024-01-26 14:42:11 -05:00
parent 8ca33dec7d
commit 6af02b19d1
3 changed files with 8 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

4
examples/main/main.cpp
View file

@ -186,10 +186,6 @@ int main(int argc, char ** argv) {
g_model = &model;
g_ctx = &ctx;
#if defined(GGML_USE_KOMPUTE)
ggml_vk_init_device(0, "gpu");
#endif
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
std::tie(model, ctx) = llama_init_from_gpt_params(params);

9
ggml-kompute.cpp
View file

@ -1837,6 +1837,8 @@ static const char * ggml_backend_kompute_name(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;
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;
}
@ -1873,6 +1875,12 @@ static struct ggml_backend_i kompute_backend_i = {
};
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()) {
fprintf(stderr, "%s: error: device was not initialized\n", __func__);
return nullptr;
@ -1897,6 +1905,5 @@ extern "C" ggml_backend_t ggml_backend_reg_kompute_init(const char * params, voi
ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {
GGML_UNUSED(params);
GGML_UNUSED(user_data);
ggml_vk_init_device(0, "gpu");
return ggml_backend_kompute_init();
}

3
llama.cpp
View file

@ -10162,9 +10162,6 @@ struct llama_context * llama_new_context_with_model(
void llama_free(struct llama_context * ctx) {
delete ctx;
#ifdef GGML_USE_KOMPUTE
ggml_vk_free_device();
#endif
}
const llama_model * llama_get_model(const struct llama_context * ctx) {