nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 16 Packages Projects Releases Wiki Activity

mtmd: n_head_kv defaults to n_head (#23782)

Browse Source 
removed AI-generated comment
This commit is contained in:
Saba Fallah
2026-05-28 16:44:36 +02:00
committed by GitHub
parent d6be3158e1
commit 0b56d283bf
2 changed files with 8 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip-graph.h
+1
View File
@@ -29,6 +29,7 @@ struct clip_graph {
const int n_patches; const int n_patches;
const int n_embd; const int n_embd;
const int n_head; const int n_head;
const int n_head_kv;
const int d_head; const int d_head;
const int n_layer; const int n_layer;
const int n_mmproj_embd; const int n_mmproj_embd;
tools/mtmd/clip.cpp
+7 -3
View File
@@ -246,6 +246,7 @@ clip_graph::clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
n_patches(n_patches_x * n_patches_y), n_patches(n_patches_x * n_patches_y),
n_embd(hparams.n_embd), n_embd(hparams.n_embd),
n_head(hparams.n_head), n_head(hparams.n_head),
n_head_kv(hparams.n_head_kv),
d_head(n_embd / n_head), d_head(n_embd / n_head),
n_layer(hparams.n_layer), n_layer(hparams.n_layer),
n_mmproj_embd(clip_n_mmproj_embd(ctx)), n_mmproj_embd(clip_n_mmproj_embd(ctx)),
@@ -401,9 +402,9 @@ ggml_tensor * clip_graph::build_vit(
} }
} }
Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos); Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos);
Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos); Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head_kv, n_pos);
Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos); Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head_kv, n_pos);
if (norm_per_head) { if (norm_per_head) {
if (layer.q_norm) { if (layer.q_norm) {
@@ -1120,6 +1121,9 @@ struct clip_model_loader {
get_u32(string_format(KEY_PROJ_DIM, prefix), hparams.projection_dim); get_u32(string_format(KEY_PROJ_DIM, prefix), hparams.projection_dim);
get_f32(string_format(KEY_LAYER_NORM_EPS, prefix), hparams.eps); get_f32(string_format(KEY_LAYER_NORM_EPS, prefix), hparams.eps);
// n_head_kv is optional (for GQA), default to n_head
hparams.n_head_kv = hparams.n_head;
if (is_vision) { if (is_vision) {
get_u32(KEY_IMAGE_SIZE, hparams.image_size); get_u32(KEY_IMAGE_SIZE, hparams.image_size);
get_u32(KEY_PATCH_SIZE, hparams.patch_size); get_u32(KEY_PATCH_SIZE, hparams.patch_size);