vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 13:34:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix kappa and memory bounds, account for look-up table and input/output layer delay

Browse source
This commit is contained in:
Lizonghang 2025-01-25 22:31:40 +04:00
parent 9e4ba4f06a
commit f3dd5776eb
2 changed files with 34 additions and 34 deletions
Download patch file Download diff file Expand all files Collapse all files

7
src/llama.cpp
View file

@ -5255,14 +5255,11 @@ struct llama_model_loader {
const auto & mapping = mappings.at(idx);
*addr = mapping->addr;
auto merge_tensor_range = [&](ggml_context * context, bool keep_only_inp_out) {
auto merge_tensor_range = [&](ggml_context * context, bool keep_output) {
for (ggml_tensor * tensor = ggml_get_first_tensor(context); tensor; tensor = ggml_get_next_tensor(context, tensor)) {
try {
const char * tname = ggml_get_name(tensor);
if (keep_only_inp_out && !(
// strcmp(tname, "token_embd.weight") == 0 || // lookup table is used so we do not need to keep it in metal memory
strcmp(tname, "output_norm.weight") == 0 ||
strcmp(tname, "output.weight") == 0)) {
if (keep_output && !(strcmp(tname, "output_norm.weight") == 0 || strcmp(tname, "output.weight") == 0)) {
continue;
}