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

Disable Vulkan async backend functions for now

Browse source
This commit is contained in:
0cc4m 2024-02-03 14:43:21 +01:00
parent bdec4a2263
commit 91e7766848
2 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-backend.c
View file

@ -1437,7 +1437,7 @@ static void sched_compute_splits(ggml_backend_sched_t sched) {
// this is important to avoid copying constants such as KQ_mask and inp_pos multiple times
ggml_backend_tensor_copy_async(split_backend, input, input_cpy);
}
ggml_backend_synchronize(split_backend);
//ggml_backend_synchronize(split_backend); // necessary to measure copy time
int64_t copy_end_us = ggml_time_us();
copy_us[split_backend_id] += copy_end_us - copy_start_us;

8
ggml-vulkan.cpp
View file

@ -4775,10 +4775,10 @@ 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 = */ ggml_backend_vk_set_tensor_async,
/* .get_tensor_async = */ ggml_backend_vk_get_tensor_async,
/* .cpy_tensor_async = */ ggml_backend_vk_cpy_tensor_async,
/* .synchronize = */ ggml_backend_vk_synchronize,
/* .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,
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,