CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

hparams : add n_embd_inp() to support extended embed (#16928)

Browse Source 
* add n_embd_full to support extended embed

* don't change output

* rename to n_embd_inp

* restore n_embd where applicable
This commit is contained in:
Sigbjørn Skjæret
2025-11-07 19:27:58 +01:00
committed by GitHub
parent 16bcc1259d
commit 9008027aa3
9 changed files with 29 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-graph.cpp
View File

@@ -1142,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
// input embeddings with optional lora
ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_inp();
auto inp = std::make_unique<llm_graph_input_embd>();
@@ -1279,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
// return cur;
//}
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);