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

lora : include embd and output layers in size calculation

Browse source
This commit is contained in:
slaren 2023-12-16 17:04:39 +01:00
parent b7c7845277
commit 93c9a5d12a
1 changed files with 0 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

5
llama.cpp
View file

@ -8676,11 +8676,6 @@ static int llama_apply_lora_from_file_internal(
std::unordered_map<std::string, struct ggml_tensor*> model_tensors; std::unordered_map<std::string, struct ggml_tensor*> model_tensors;
for (const auto & kv : model.tensors_by_name) { for (const auto & kv : model.tensors_by_name) {
model_tensors.insert(kv); model_tensors.insert(kv);
// skip input and output layers as they are not often finetuned and can be very large
if (kv.first.find("token_embd") != std::string::npos ||
kv.first.find("output") != std::string::npos) {
continue;
}
size_t f32_size = ggml_nelements(kv.second) * sizeof(float); size_t f32_size = ggml_nelements(kv.second) * sizeof(float);
max_tensor_size = std::max(max_tensor_size, f32_size); max_tensor_size = std::max(max_tensor_size, f32_size);
} }