char-fork/llama.cpp

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

Xuan-Son Nguyenclip: move model cgraphs into their own files (#17965)e39a2ce66

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.0 KiB63 linesraw
1#include "models.h"
2
3ggml_cgraph * clip_graph_kimivl::build() {
4    // 2D input positions
5    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 embeddings
16    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
17        // first half is X axis and second half is Y axis
18        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_merger
33        const int scale_factor = model.hparams.n_merge;
34        cur = build_patch_merge_permute(cur, scale_factor);
35
36        // projection norm
37        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.LayerNorm
42        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 mlp
50        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 graph
60    ggml_build_forward_expand(gf, cur);
61
62    return gf;
63}