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

llama : only use default buffer types for the KV cache (#10358)

Browse source
This commit is contained in:
Diego Devesa 2024-11-17 12:25:45 +01:00 • committed by GitHub
parent 20a780c7b6
commit be5caccef9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 9 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml/src/ggml-backend.cpp
View file

@ -689,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
}
static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
ggml_backend_buffer_t buffer = tensor->buffer;
ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
if (buffer == NULL) {
return -1;
}
@ -722,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
// returns the backend that should be used for the node based on the current locations
static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
// TODO: use supports_op to check if the backend supports the op
// assign pre-allocated nodes to their backend
int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
if (cur_backend_id != -1) {
@ -742,7 +740,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
// since the tensor is pre-allocated, it cannot be moved to another backend
GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
}
// graph input
@ -886,6 +884,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int i = 0; i < graph->n_nodes; i++) {
struct ggml_tensor * node = graph->nodes[i];
int * node_backend_id = &tensor_backend_id(node);
if (ggml_is_view_op(node->op)) {
continue;
}
// do not overwrite user assignments
if (*node_backend_id == -1) {
*node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);