char-fork/llama.cpp

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

Xuan-Son Nguyenmtmd, model: allow skip build_vit() (#24077)a731805ce

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.8 KiB71 linesraw
1#include "models.h"
2#include <cmath>
3
4ggml_cgraph * clip_graph_gemma4uv::build() {
5    ggml_tensor * inp_raw = build_inp_raw();
6
7    // Gemma4UnifiedVisionEmbedder uses default pytorch LayerNorm, not RMSNorm
8    float eps = 1e-5f; // default eps for pytorch LayerNorm
9
10    ggml_tensor * inp = nullptr;
11    {
12        // note: we cannot use ggml_conv_2d here because we need to apply norm after im2col
13        auto c = inp_raw->ne[2];
14        ggml_tensor * kernel = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, patch_size, patch_size, c);
15        inp = ggml_im2col(ctx0, kernel, inp_raw, patch_size, patch_size, 0, 0, 1, 1, true, inp_raw->type);
16        // inp shape: [patch_size * patch_size * c, n_patches_w, n_patches_h]
17
18        inp = ggml_reshape_2d(ctx0, inp, inp->ne[0], inp->ne[1] * inp->ne[2] * inp->ne[3]);
19        inp = build_norm(inp, model.patch_norm_1_w, model.patch_norm_1_b, NORM_TYPE_NORMAL, eps, -1);
20        // inp shape: [patch_size * patch_size * c, n_patches]
21
22        inp = ggml_mul_mat(ctx0, model.patch_embeddings_0, inp);
23        inp = ggml_add(ctx0, inp, model.patch_bias);
24        // inp shape: [n_embd, n_patches]
25
26        inp = build_norm(inp, model.patch_norm_2_w, model.patch_norm_2_b, NORM_TYPE_NORMAL, eps, -1);
27    }
28
29    ggml_tensor * pos_x = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
30    ggml_set_name(pos_x, "pos_x");
31    ggml_set_input(pos_x);
32
33    ggml_tensor * pos_y = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
34    ggml_set_name(pos_y, "pos_y");
35    ggml_set_input(pos_y);
36
37    {
38        const int64_t pos_size = model.position_embeddings->ne[1];
39        const size_t  nb1      = ggml_row_size(model.position_embeddings->type, n_embd);
40
41        // positional embeddings are stored as lookup tables (one for x, one for y)
42        ggml_tensor * tbl_x = ggml_view_2d(ctx0, model.position_embeddings,
43                                             n_embd, pos_size, nb1, 0);
44        ggml_tensor * tbl_y = ggml_view_2d(ctx0, model.position_embeddings,
45                                             n_embd, pos_size, nb1, pos_size * nb1);
46
47        // ggml_get_rows: [n_embd, n_patches]
48        ggml_tensor * emb_x = ggml_get_rows(ctx0, tbl_x, pos_x);
49        ggml_tensor * emb_y = ggml_get_rows(ctx0, tbl_y, pos_y);
50
51        inp = ggml_add(ctx0, inp, emb_x);
52        inp = ggml_add(ctx0, inp, emb_y);
53        cb(inp, "pos_embd", -1);
54
55        // pos_norm
56        inp = build_norm(inp, model.patch_norm_3_w, model.patch_norm_3_b, NORM_TYPE_NORMAL, eps, -1);
57    }
58
59    auto cur = inp;
60
61    // Gemma4UnifiedMultimodalEmbedder
62    {
63        // embedding_pre_projection_norm
64        cur = ggml_rms_norm(ctx0, cur, hparams.eps);
65        cur = build_mm(model.mm_input_proj_w, cur);
66        cb(cur, "projected", -1);
67    }
68
69    ggml_build_forward_expand(gf, cur);
70    return gf;
71}