CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

use build_ffn

Browse Source
This commit is contained in:
JJJYmmm
2025-10-28 20:45:43 +08:00
parent f84bd67c80
commit 0443a098f3
1 changed files with 10 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
tools/mtmd/clip.cpp
View File

@@ -993,13 +993,11 @@ struct clip_graph {
feat = ggml_reshape_3d(ctx0, feat, n_embd * merge_factor, n_pos / merge_factor, batch_size); feat = ggml_reshape_3d(ctx0, feat, n_embd * merge_factor, n_pos / merge_factor, batch_size);
feat = build_norm(feat, merger.norm_w, merger.norm_b, norm_t, eps, il); feat = build_norm(feat, merger.norm_w, merger.norm_b, norm_t, eps, il);
feat = ggml_mul_mat(ctx0, merger.fc1_w, feat); feat = build_ffn(feat,
feat = ggml_add(ctx0, feat, merger.fc1_b); merger.fc1_w, merger.fc1_b,
nullptr, nullptr,
feat = ggml_gelu(ctx0, feat); merger.fc2_w, merger.fc2_b,
ffn_op_type::FFN_GELU, il);
feat = ggml_mul_mat(ctx0, merger.fc2_w, feat);
feat = ggml_add(ctx0, feat, merger.fc2_b);
if(!deepstack_features) { if(!deepstack_features) {
deepstack_features = feat; deepstack_features = feat;
@@ -1021,15 +1019,11 @@ struct clip_graph {
ggml_tensor * embeddings = inpL; ggml_tensor * embeddings = inpL;
embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size); embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size);
embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); embeddings = build_ffn(embeddings,
embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); model.mm_0_w, model.mm_0_b,
nullptr, nullptr,
// GELU activation model.mm_1_w, model.mm_1_b,
embeddings = ggml_gelu(ctx0, embeddings); ffn_op_type::FFN_GELU, -1);
// Second linear layer
embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension