echodict/llama.cpp
version https://git-lfs.github.com/spec/v1 oid sha256:cfc44b7ba25614df70e6b65e3341cae0310163bd32fd31a6b928a542df433faf size 30786
0479
1#include "models.h"2 3ggml_cgraph * clip_graph_kimivl::build() {4 // 2D input positions5 ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);6 ggml_set_name(pos_h, "pos_h");7 ggml_set_input(pos_h);8 9 ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);10 ggml_set_name(pos_w, "pos_w");11 ggml_set_input(pos_w);12 13 ggml_tensor * learned_pos_embd = resize_position_embeddings();14 15 // build ViT with 2D position embeddings16 auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {17 // first half is X axis and second half is Y axis18 return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);19 };20 21 ggml_tensor * inp = build_inp();22 ggml_tensor * cur = build_vit(23 inp, n_patches,24 NORM_TYPE_NORMAL,25 hparams.ffn_op,26 learned_pos_embd,27 add_pos);28 29 cb(cur, "vit_out", -1);30 31 {32 // patch_merger33 const int scale_factor = model.hparams.n_merge;34 cur = build_patch_merge_permute(cur, scale_factor);35 36 // projection norm37 int proj_inp_dim = cur->ne[0];38 cur = ggml_view_2d(ctx0, cur,39 n_embd, cur->ne[1] * scale_factor * scale_factor,40 ggml_row_size(cur->type, n_embd), 0);41 cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm42 cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);43 cur = ggml_add(ctx0, cur, model.mm_input_norm_b);44 cur = ggml_view_2d(ctx0, cur,45 proj_inp_dim, cur->ne[1] / scale_factor / scale_factor,46 ggml_row_size(cur->type, proj_inp_dim), 0);47 cb(cur, "proj_inp_normed", -1);48 49 // projection mlp50 cur = build_ffn(cur,51 model.mm_1_w, model.mm_1_b,52 nullptr, nullptr,53 model.mm_2_w, model.mm_2_b,54 FFN_GELU,55 -1);56 cb(cur, "proj_out", -1);57 }58 59 // build the graph60 ggml_build_forward_expand(gf, cur);61 62 return gf;63}64 