upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-03 16:40:45 -05:00
Code Issues Packages Projects Releases Wiki Activity

llama : load MTP tensors only if they are really used (#26296)

Browse Source 
* llama : load MTP tensors only if they are really used

* llama : skip loading MTP (if not used) in remaining models that support MTP

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming
2026-07-31 14:57:02 +02:00
committed by GitHub
parent 6f3c0a790b
commit 82dbc4f017
13 changed files with 65 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
src/models/step35.cpp
View File

@@ -48,7 +48,11 @@ void llama_model_step35::load_arch_tensors(llama_model_loader & ml) {
const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
const int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
int mtp_flags = trunk_only ? TENSOR_NOT_REQUIRED : 0;
if (!ml.load_mtp) {
mtp_flags |= TENSOR_SKIP;
}
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);