ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : upgrade init_tensor API to return a ggml_status (#11854)

Browse source 
* Upgrade init_tensor API to return a ggml_status

To prepare for an 'abort-free' ggml
(ggml not to abort on OOMs but return a OOM status),
as agreeed with Diego in the ggml repo,
upgrade the init_tensor() and view_init() APIs
to return a ggml_status.

* misc fixes

---------

Co-authored-by: slaren <slarengh@gmail.com>
This commit is contained in:
William Tambellini 2025-02-28 05:41:47 -08:00 • committed by GitHub
parent c43a3e7996
commit 70680c48e5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
16 changed files with 136 additions and 97 deletions
Download patch file Download diff file Expand all files Collapse all files

17
ggml/src/ggml-backend.cpp
View file

@ -126,11 +126,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
return base;
}
void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
enum ggml_status ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
// init_tensor is optional
if (buffer->iface.init_tensor) {
buffer->iface.init_tensor(buffer, tensor);
return buffer->iface.init_tensor(buffer, tensor);
}
return GGML_STATUS_SUCCESS;
}
void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
@ -1641,7 +1642,7 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
// utils
void ggml_backend_view_init(struct ggml_tensor * tensor) {
enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor) {
GGML_ASSERT(tensor->buffer == NULL);
GGML_ASSERT(tensor->view_src != NULL);
GGML_ASSERT(tensor->view_src->buffer != NULL);
@ -1649,10 +1650,10 @@ void ggml_backend_view_init(struct ggml_tensor * tensor) {
tensor->buffer = tensor->view_src->buffer;
tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
return ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
}
void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
GGML_ASSERT(tensor->buffer == NULL);
GGML_ASSERT(tensor->data == NULL);
GGML_ASSERT(tensor->view_src == NULL);
@ -1662,7 +1663,7 @@ void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor
tensor->buffer = buffer;
tensor->data = addr;
ggml_backend_buffer_init_tensor(buffer, tensor);
return ggml_backend_buffer_init_tensor(buffer, tensor);
}
static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
@ -1708,7 +1709,8 @@ static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_
struct ggml_tensor * dst = node_copies[id];
if (dst->view_src != NULL) {
graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
ggml_backend_view_init(dst);
enum ggml_status status = ggml_backend_view_init(dst);
GGML_ASSERT(status == GGML_STATUS_SUCCESS);
}
else {
ggml_backend_tensor_copy(src, dst);
@ -1823,7 +1825,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
assert(g1->n_nodes == g2->n_nodes);
for (int i = 0; i < g1->n_nodes; i++) {
//printf("eval %d/%d\n", i, g1->n_nodes);
struct ggml_tensor * t1 = g1->nodes[i];
struct ggml_tensor * t2 = g2->nodes[i];