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

leave the schedule to ggml_backend_sched entirely

Browse source
This commit is contained in:
Meng, Hengyu 2024-03-23 08:29:54 +00:00
parent 4b9f3b432b
commit 0c2aa1a249
2 changed files with 17 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml.c
View file

@ -16059,12 +16059,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU); GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU);
#endif // GGML_USE_VULKAN #endif // GGML_USE_VULKAN
#ifdef GGML_USE_SYCL
bool skip_cpu = ggml_sycl_compute_forward(params, tensor);
if (skip_cpu) {
return;
}
#endif // GGML_USE_SYCL
switch (tensor->op) { switch (tensor->op) {
case GGML_OP_DUP: case GGML_OP_DUP:
{ {

36
llama.cpp
View file

@ -13158,30 +13158,28 @@ struct llama_context * llama_new_context_with_model(
} }
} }
#elif defined(GGML_USE_SYCL) #elif defined(GGML_USE_SYCL)
if (model->n_gpu_layers > 0) { // with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used
// with split_mode LLAMA_SPLIT_MODE_NONE or LLAMA_SPLIT_MODE_ROW, only the main GPU backend is used if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) {
if (model->split_mode == LLAMA_SPLIT_MODE_NONE || model->split_mode == LLAMA_SPLIT_MODE_ROW) { ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu);
ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); if (backend == nullptr) {
int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu);
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
} else {
// LLAMA_SPLIT_LAYER requires a backend for each GPU
for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
ggml_backend_t backend = ggml_backend_sycl_init(i);
if (backend == nullptr) { if (backend == nullptr) {
int main_gpu_id = ggml_backend_sycl_get_device_id(model->main_gpu); int id_list[GGML_SYCL_MAX_DEVICES];
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, main_gpu_id, model->main_gpu); ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, id_list[i], i);
llama_free(ctx); llama_free(ctx);
return nullptr; return nullptr;
} }
ctx->backends.push_back(backend); ctx->backends.push_back(backend);
} else {
// LLAMA_SPLIT_LAYER requires a backend for each GPU
for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) {
ggml_backend_t backend = ggml_backend_sycl_init(i);
if (backend == nullptr) {
int id_list[GGML_SYCL_MAX_DEVICES];
ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES);
LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d) backend\n", __func__, id_list[i], i);
llama_free(ctx);
return nullptr;
}
ctx->backends.push_back(backend);
}
} }
} }
#elif defined(GGML_USE_KOMPUTE) #elif defined(GGML_USE_KOMPUTE)