CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-01 09:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

optimize deepstack feature saving

Browse Source
This commit is contained in:
JJJYmmm
2025-10-29 14:55:48 +08:00
parent 3271877207
commit f321b9fdf1
1 changed files with 22 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
tools/mtmd/clip.cpp
View File

@@ -846,9 +846,12 @@ struct clip_graph {
GGML_ASSERT(model.patch_bias != nullptr);
GGML_ASSERT(model.position_embeddings != nullptr);
GGML_ASSERT(model.class_embedding == nullptr);
GGML_ASSERT(hparams.spatial_merge_size == 2);
const int batch_size = 1;
const int merge_factor = 4;
const int n_pos = n_patches;
const int n_pos_merged = n_pos / merge_factor;
const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
norm_type norm_t = NORM_TYPE_NORMAL;
@@ -911,9 +914,23 @@ struct clip_graph {
inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1);
}
// deepstack features (stack along the feature dimension), [n_embd * len(deepstack_layers), n_patches_x * n_patches_y, batch_size]
ggml_tensor * deepstack_features = nullptr;
const int merge_factor = hparams.spatial_merge_size > 0 ? hparams.spatial_merge_size * hparams.spatial_merge_size : 4; // default 2x2=4 for qwen3vl
int deepstack_layer_idx = 1; // begin with 1 to jump main feature
const int llm_n_embd = model.mm_1_w->ne[1]; // llm token dim
const int n_deepstack_layers = std::count(hparams.is_deepstack_layers.begin(), hparams.is_deepstack_layers.end(), true);
const size_t element_size = ggml_type_size(inpL->type);
const size_t slice_offsets = llm_n_embd * n_pos_merged * batch_size * element_size;
ggml_tensor * final_embedding = ggml_new_tensor_3d(ctx0, inpL->type,
llm_n_embd * (n_deepstack_layers + 1), n_pos_merged, batch_size);
auto make_deepstack_slice = [&](int idx) {
return ggml_view_3d(ctx0, final_embedding,
llm_n_embd, n_pos_merged, batch_size,
llm_n_embd * element_size,
slice_offsets,
idx * slice_offsets);
};
// loop over layers
for (int il = 0; il < n_layer; il++) {
@@ -990,13 +1007,7 @@ struct clip_graph {
nullptr, nullptr,
layer.deepstack_fc2_w, layer.deepstack_fc2_b,
ffn_op_type::FFN_GELU, il);
if(!deepstack_features) {
deepstack_features = feat;
} else {
// concat along the feature dimension
deepstack_features = ggml_concat(ctx0, deepstack_features, feat, 0);
}
ggml_cpy(ctx0, feat, make_deepstack_slice(deepstack_layer_idx++));
}
inpL = cur;
@@ -1017,7 +1028,7 @@ struct clip_graph {
model.mm_1_w, model.mm_1_b,
ffn_op_type::FFN_GELU, -1);
embeddings = ggml_concat(ctx0, embeddings, deepstack_features, 0); // concat along the feature dimension
ggml_cpy(ctx0, embeddings, make_deepstack_slice(0));
// build the graph
ggml_build_forward_expand(gf, embeddings);