ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama : support GEGLU for jina-bert-v2 (#14090)

Browse source
This commit is contained in:
Sigbjørn Skjæret 2025-06-10 18:02:08 +02:00 • committed by GitHub
parent 652b70e667
commit 3678b838bb
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 9 additions and 33 deletions
Download patch file Download diff file Expand all files Collapse all files

3
src/llama-graph.cpp
View file

@ -650,6 +650,7 @@ ggml_tensor * llm_graph_context::build_ffn(
{
// Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
int64_t split_point = cur->ne[0] / 2;
// TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
@ -663,7 +664,7 @@ ggml_tensor * llm_graph_context::build_ffn(
{
// Split into two equal parts
int64_t split_point = cur->ne[0] / 2;
// TODO: these conts should not be needed
// TODO: these conts should not be needed, see https://github.com/ggml-org/llama.cpp/pull/14090#discussion_r2137437217
ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));

6
src/llama-model.cpp
View file

@ -2224,8 +2224,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, layer.ffn_gate ? n_ff : n_ff * 2}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
@ -6043,7 +6043,7 @@ struct llm_build_bert : public llm_graph_context {
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_GELU, LLM_FFN_PAR, il);
model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
cur = build_ffn(cur,