char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

megeminimodel: Add PaddleOCR-VL model support (#18825)237958db3

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.7 KiB52 linesraw
1#include "models.h"
2
3ggml_cgraph * clip_graph_paddleocr::build() {
4    const int n_pos            = n_patches;
5    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
6
7    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
8
9    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
10    ggml_set_name(positions, "positions");
11    ggml_set_input(positions);
12
13    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
14        return ggml_rope_multi(
15                    ctx0, cur, positions, nullptr,
16                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,
17                    32768, 10000, 1, 0, 1, 32, 1);
18    };
19
20    ggml_tensor * learned_pos_embd = resize_position_embeddings();
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        // mlp_AR paddleocr projector
33        float proj_norm_eps = 1e-5;
34        cur = build_norm(cur,
35                    model.mm_input_norm_w, model.mm_input_norm_b,
36                    NORM_TYPE_NORMAL, proj_norm_eps, -1);
37
38        const int scale_factor = model.hparams.n_merge;
39        cur = build_patch_merge_permute(cur, scale_factor);
40        cur = build_ffn(cur,
41                    model.mm_1_w, model.mm_1_b,
42                    nullptr, nullptr,
43                    model.mm_2_w, model.mm_2_b,
44                    hparams.ffn_op, -1);
45        cb(cur, "mlp_out", -1);
46    }
47
48    // build the graph
49    ggml_build_forward_expand(gf, cur);
50
51    return gf;
52}