Coverage Report

Created: 2026-03-07 06:35

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/llama.cpp/src/models/paddleocr.cpp
Line
Count
Source
1
#include "models.h"
2
3
llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_graph_params & params) :
4
0
    llm_graph_context(params) {
5
6
    // NOTE: same with qwen2vl.cpp, but bias tensors are optional
7
8
0
    const int64_t n_embd_head = hparams.n_embd_head_v;
9
10
0
    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
11
0
    GGML_ASSERT(n_embd_head == hparams.n_rot);
12
13
0
    ggml_tensor * cur;
14
0
    ggml_tensor * inpL;
15
16
0
    inpL = build_inp_embd(model.tok_embd);
17
18
0
    int sections[4];
19
0
    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
20
21
    // inp_pos - contains the positions
22
0
    ggml_tensor * inp_pos = build_inp_pos();
23
24
0
    auto * inp_attn = build_attn_inp_kv();
25
26
0
    ggml_tensor * inp_out_ids = build_inp_out_ids();
27
28
0
    for (int il = 0; il < n_layer; ++il) {
29
0
        ggml_tensor * inpSA = inpL;
30
31
        // norm
32
0
        {
33
0
            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
34
0
            cb(cur, "attn_norm", il);
35
0
        }
36
        // self-attention
37
0
        {
38
0
            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
39
0
            cb(Qcur, "Qcur", il);
40
0
            if (model.layers[il].bq) {
41
0
                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
42
0
                cb(Qcur, "Qcur", il);
43
0
            }
44
0
            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
45
0
            cb(Kcur, "Kcur", il);
46
0
            if (model.layers[il].bk) {
47
0
                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
48
0
                cb(Kcur, "Kcur", il);
49
0
            }
50
0
            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
51
0
            cb(Vcur, "Vcur", il);
52
0
            if (model.layers[il].bv) {
53
0
                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
54
0
                cb(Vcur, "Vcur", il);
55
0
            }
56
0
            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
57
0
            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
58
0
            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
59
60
0
            Qcur = ggml_rope_multi(
61
0
                    ctx0, Qcur, inp_pos, nullptr,
62
0
                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
63
0
                    ext_factor, attn_factor, beta_fast, beta_slow
64
0
                    );
65
66
0
            Kcur = ggml_rope_multi(
67
0
                    ctx0, Kcur, inp_pos, nullptr,
68
0
                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
69
0
                    ext_factor, attn_factor, beta_fast, beta_slow
70
0
                    );
71
72
0
            cb(Qcur, "Qcur", il);
73
0
            cb(Kcur, "Kcur", il);
74
0
            cb(Vcur, "Vcur", il);
75
76
0
            cur = build_attn(inp_attn,
77
0
                    model.layers[il].wo, model.layers[il].bo,
78
0
                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
79
0
        }
80
0
        if (il == n_layer - 1) {
81
            // skip computing output for unused tokens
82
0
            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
83
0
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
84
0
        }
85
0
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
86
0
        cb(ffn_inp, "ffn_inp", il);
87
88
        // feed-forward network
89
0
        {
90
0
            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
91
0
            cb(cur, "ffn_norm", il);
92
93
0
            cur = build_ffn(cur,
94
0
                    model.layers[il].ffn_up, NULL, NULL,
95
0
                    model.layers[il].ffn_gate, NULL, NULL,
96
0
                    model.layers[il].ffn_down, NULL, NULL,
97
0
                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
98
0
            cb(cur, "ffn_out", il);
99
0
        }
100
0
        cur = ggml_add(ctx0, cur, ffn_inp);
101
102
0
        cur = build_cvec(cur, il);
103
0
        cb(cur, "l_out", il);
104
105
        // input for next layer
106
0
        inpL = cur;
107
0
    }
108
0
    cur = inpL;
109
110
0
    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
111
112
0
    cb(cur, "result_norm", -1);
113
0
    res->t_embd = cur;
114
115
    // lm_head
116
0
    cur = build_lora_mm(model.output, cur);
117
118
0
    cb(cur, "result_output", -1);
119
0
    res->t_logits = cur;
120
121
0
    ggml_build_forward_expand(gf, cur);
122
0
}