sleepy/llama.cpp
1
0
Fork 0
Code Issues 16 Pull Requests Actions 30 Packages Projects Releases Wiki Activity

model: support GLM-OCR (#19677)

Browse Source 
* model: support GLM-OCR

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Xuan-Son Nguyen
2026-02-18 17:51:40 +01:00
committed by GitHub
parent e99f1083a0
commit eeef3cfced
8 changed files with 122 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip.cpp
+11 -3
View File
@@ -342,9 +342,17 @@ ggml_tensor * clip_graph::build_vit(
/* nb2 */ cur->nb[1],
/* offset */ ggml_row_size(cur->type, 2 * n_embd));
// TODO: q/k norm requires row size == n_embd, while here it's d_head
// we can add support in the future if needed
GGML_ASSERT(layer.q_norm == nullptr && layer.k_norm == nullptr);
if (layer.q_norm) {
GGML_ASSERT(layer.q_norm->ne[0] == Qcur->ne[0]);
Qcur = build_norm(Qcur, layer.q_norm, NULL, norm_t, eps, il);
cb(Qcur, "Qcur_norm", il);
}
if (layer.k_norm) {
GGML_ASSERT(layer.k_norm->ne[0] == Kcur->ne[0]);
Kcur = build_norm(Kcur, layer.k_norm, NULL, norm_t, eps, il);
cb(Kcur, "Kcur_norm", il);
}
} else {
// separate q, k, v
tools/mtmd/models/glm4v.cpp
+16 -14
View File
@@ -2,7 +2,6 @@
ggml_cgraph * clip_graph_glm4v::build() {
GGML_ASSERT(model.patch_bias != nullptr);
GGML_ASSERT(model.position_embeddings != nullptr);
GGML_ASSERT(model.class_embedding == nullptr);
const int batch_size = 1;
@@ -45,19 +44,22 @@ ggml_cgraph * clip_graph_glm4v::build() {
// pos-conv norm
inp = build_norm(inp, model.norm_embd_w, model.norm_embd_b, norm_t, eps, -1);
// calculate absolute position embedding and apply
ggml_tensor * learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);
learned_pos_embd = ggml_cont_4d(
ctx0, learned_pos_embd,
n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
learned_pos_embd = ggml_reshape_4d(
ctx0, learned_pos_embd,
n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
learned_pos_embd = ggml_cont_3d(
ctx0, learned_pos_embd,
n_embd, n_patches_x * n_patches_y, batch_size);
cb(learned_pos_embd, "learned_pos_embd", -1);
ggml_tensor * learned_pos_embd = nullptr;
// Note: GLM-OCR does not have learned position embeddings
if (model.position_embeddings != nullptr) {
learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BICUBIC);
learned_pos_embd = ggml_cont_4d(
ctx0, learned_pos_embd,
n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
learned_pos_embd = ggml_reshape_4d(
ctx0, learned_pos_embd,
n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
learned_pos_embd = ggml_permute(ctx0, learned_pos_embd, 0, 2, 1, 3);
learned_pos_embd = ggml_cont_3d(
ctx0, learned_pos_embd,
n_embd, n_patches_x * n_patches_y, batch_size);
cb(learned_pos_embd, "learned_pos_embd", -1);
}
auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
return ggml_rope_multi(