char-fork/llama.cpp

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

Xuan-Son Nguyenmtmd: add clip_graph::build_mm() (#20751)1e6453457

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.0 KiB98 linesraw
1#include "models.h"
2
3ggml_cgraph * clip_graph_cogvlm::build() {
4    GGML_ASSERT(model.class_embedding != nullptr);
5    GGML_ASSERT(model.position_embeddings != nullptr);
6
7    const int n_pos = n_patches + 1; // +1 for [CLS]
8
9    // build input and concatenate class embedding
10    ggml_tensor * inp = build_inp();
11    inp = ggml_concat(ctx0, inp, model.class_embedding, 1);
12
13    inp = ggml_add(ctx0, inp, model.position_embeddings);
14    cb(inp, "inp_pos", -1);
15
16    ggml_tensor * inpL = inp;
17
18    for (int il = 0; il < n_layer; il++) {
19        auto & layer = model.layers[il];
20        ggml_tensor * cur = inpL;
21
22        cur = build_mm(layer.qkv_w, cur);
23
24        cur = ggml_add(ctx0, cur, layer.qkv_b);
25
26        ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
27            cur->nb[1], 0);
28        ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
29            cur->nb[1], n_embd * sizeof(float));
30        ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, d_head, n_head, n_pos, d_head*sizeof(float),
31            cur->nb[1], 2 * n_embd * sizeof(float));
32
33        cb(Qcur, "Qcur", il);
34        cb(Kcur, "Kcur", il);
35        cb(Vcur, "Vcur", il);
36
37        cur = build_attn(layer.o_w, layer.o_b,
38            Qcur, Kcur, Vcur, nullptr, kq_scale, il);
39        cb(cur, "attn_out", il);
40
41        cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il);
42        cb(cur, "attn_post_norm", il);
43
44        cur = ggml_add(ctx0, cur, inpL);
45        inpL = cur;
46
47        cur = build_ffn(cur,
48            layer.ff_up_w, layer.ff_up_b,
49            layer.ff_gate_w, layer.ff_gate_b,
50            layer.ff_down_w, layer.ff_down_b,
51            hparams.ffn_op, il);
52
53        cb(cur, "ffn_out", il);
54
55        cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il);
56        cb(cur, "ffn_post_norm", il);
57
58        cur = ggml_add(ctx0, cur, inpL);
59        cb(cur, "layer_out", il);
60        inpL = cur;
61
62    }
63
64    // remove CLS token (like build_llama4 does)
65    ggml_tensor * cur = ggml_view_2d(ctx0, inpL,
66        n_embd, n_patches,
67        ggml_row_size(inpL->type, n_embd), 0);
68
69    // Multiply with mm_model_proj
70    cur = build_mm(model.mm_model_proj, cur);
71
72    // Apply layernorm, weight, bias
73    cur = build_norm(cur, model.mm_post_fc_norm_w, model.mm_post_fc_norm_b, NORM_TYPE_NORMAL, 1e-5, -1);
74
75    // Apply GELU
76    cur = ggml_gelu_inplace(ctx0, cur);
77
78    // Branch 1: multiply with mm_h_to_4h_w
79    ggml_tensor * h_to_4h = build_mm(model.mm_h_to_4h_w, cur);
80
81    // Branch 2: multiply with mm_gate_w
82    ggml_tensor * gate = build_mm(model.mm_gate_w, cur);
83
84    // Apply silu
85    gate = ggml_swiglu_split(ctx0, gate, h_to_4h);
86
87    // Apply mm_4h_to_h_w
88    cur = build_mm(model.mm_4h_to_h_w, gate);
89
90    // Concatenate with boi and eoi
91    cur = ggml_concat(ctx0, model.mm_boi, cur, 1);
92    cur = ggml_concat(ctx0, cur, model.mm_eoi, 1);
93
94    // build the graph
95    ggml_build_forward_expand(gf, cur);
96
97    return gf;
98}