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

transpose when loading

Browse source
This commit is contained in:
Lorenzo Toniazzi 2024-07-06 22:59:15 +01:00
parent 798cde72a1
commit 931134b536
1 changed files with 7 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

11
llama.cpp
View file

@ -328,6 +328,12 @@ static struct lora_data * load_lora(struct lora_info * info) {
file.seek((0-file.tell()) & 31, SEEK_CUR); file.seek((0-file.tell()) & 31, SEEK_CUR);
size_t offset = file.tell(); size_t offset = file.tell();
struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne); struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
// Transpose lora matrix A
if (std::string(name_buf.data()).find("loraA") != std::string::npos) {
tensor = ggml_cont(result->ctx,
ggml_transpose(result->ctx, tensor)
);
}
ggml_set_name(tensor, name_buf.data()); ggml_set_name(tensor, name_buf.data());
size_t nbytes = ggml_nbytes(tensor); size_t nbytes = ggml_nbytes(tensor);
size_t nbytes_pad = ggml_nbytes_pad(tensor); size_t nbytes_pad = ggml_nbytes_pad(tensor);
@ -9732,10 +9738,7 @@ struct llm_build_context {
ggml_tensor * t_lora = ggml_mul_mat(ctx0, ggml_tensor * t_lora = ggml_mul_mat(ctx0,
loraB, loraB,
ggml_mul_mat(ctx0, ggml_mul_mat(ctx0, loraA, cur)
ggml_cont(ctx0, ggml_transpose(ctx0, loraA)),
cur
)
); );
if (lctx.lora_scale != 1.0f) { if (lctx.lora_scale != 1.0f) {