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

metal : relax conditions on fast matrix multiplication kernel (#3168)

Browse source 
* metal : relax conditions on fast matrix multiplication kernel

* metal : revert the concurrnecy change because it was wrong

* llama : remove experimental stuff
This commit is contained in:
Georgi Gerganov 2023-09-15 11:09:24 +03:00 • committed by GitHub
parent 76164fe2e6
commit a51b687657
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 100 additions and 51 deletions
Download patch file Download diff file Expand all files Collapse all files

20
ggml.c
View file

@ -4303,10 +4303,21 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) {
}
size_t ggml_nbytes(const struct ggml_tensor * tensor) {
size_t nbytes = tensor->ne[0]*tensor->nb[0]/ggml_blck_size(tensor->type);
for (int i = 1; i < GGML_MAX_DIMS; ++i) {
nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
size_t nbytes;
size_t blck_size = ggml_blck_size(tensor->type);
if (blck_size == 1) {
nbytes = ggml_type_size(tensor->type);
for (int i = 0; i < GGML_MAX_DIMS; ++i) {
nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
}
}
else {
nbytes = tensor->ne[0]*tensor->nb[0]/blck_size;
for (int i = 1; i < GGML_MAX_DIMS; ++i) {
nbytes += (tensor->ne[i] - 1)*tensor->nb[i];
}
}
return nbytes;
}
@ -18340,7 +18351,8 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n",
i,
node->ne[0], node->ne[1],
ggml_op_name(node->op));
ggml_op_name(node->op),
ggml_get_name(node));
}
for (int i = 0; i < GGML_OP_COUNT; i++) {