Coverage Report

Created: 2026-03-21 06:49

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/llama.cpp/src/models/baichuan.cpp
Line
Count
Source
1
#include "models.h"
2
3
4
0
llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5
0
    const int64_t n_embd_head = hparams.n_embd_head_v();
6
7
0
    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
8
0
    GGML_ASSERT(n_embd_head == n_rot);
9
10
0
    ggml_tensor * cur;
11
0
    ggml_tensor * inpL;
12
13
0
    inpL = build_inp_embd(model.tok_embd);
14
15
    // inp_pos - contains the positions
16
0
    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
17
18
0
    auto * inp_attn = build_attn_inp_kv();
19
20
0
    ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22
0
    for (int il = 0; il < n_layer; ++il) {
23
0
        ggml_tensor * inpSA = inpL;
24
25
0
        cur = build_norm(inpL,
26
0
                model.layers[il].attn_norm, NULL,
27
0
                LLM_NORM_RMS, il);
28
0
        cb(cur, "attn_norm", il);
29
30
        // self-attention
31
0
        {
32
0
            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
33
0
            cb(Qcur, "Qcur", il);
34
35
0
            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
36
0
            cb(Kcur, "Kcur", il);
37
38
0
            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
39
0
            cb(Vcur, "Vcur", il);
40
41
0
            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
42
0
            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
43
0
            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
44
45
0
            switch (model.type) {
46
0
                case LLM_TYPE_7B:
47
0
                    Qcur = ggml_rope_ext(
48
0
                            ctx0, Qcur, inp_pos, nullptr,
49
0
                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
50
0
                            ext_factor, attn_factor, beta_fast, beta_slow
51
0
                            );
52
0
                    Kcur = ggml_rope_ext(
53
0
                            ctx0, Kcur, inp_pos, nullptr,
54
0
                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
55
0
                            ext_factor, attn_factor, beta_fast, beta_slow
56
0
                            );
57
0
                    break;
58
0
                case LLM_TYPE_13B:
59
0
                case LLM_TYPE_UNKNOWN:
60
0
                    break;
61
0
                default:
62
0
                    GGML_ABORT("fatal error");
63
0
            }
64
65
0
            cb(Qcur, "Qcur", il);
66
0
            cb(Kcur, "Kcur", il);
67
0
            cb(Vcur, "Vcur", il);
68
69
0
            cur = build_attn(inp_attn,
70
0
                    model.layers[il].wo, NULL,
71
0
                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
72
0
        }
73
74
0
        if (il == n_layer - 1 && inp_out_ids) {
75
0
            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
76
0
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
77
0
        }
78
79
0
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
80
0
        cb(ffn_inp, "ffn_inp", il);
81
82
        // feed-forward network
83
0
        {
84
0
            cur = build_norm(ffn_inp,
85
0
                    model.layers[il].ffn_norm, NULL,
86
0
                    LLM_NORM_RMS, il);
87
0
            cb(cur, "ffn_norm", il);
88
89
0
            cur = build_ffn(cur,
90
0
                    model.layers[il].ffn_up,   NULL, NULL,
91
0
                    model.layers[il].ffn_gate, NULL, NULL,
92
0
                    model.layers[il].ffn_down, NULL, NULL,
93
0
                    NULL,
94
0
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
95
0
            cb(cur, "ffn_out", il);
96
0
        }
97
98
0
        cur = ggml_add(ctx0, cur, ffn_inp);
99
100
0
        cur = build_cvec(cur, il);
101
0
        cb(cur, "l_out", il);
102
103
        // input for next layer
104
0
        inpL = cur;
105
0
    }
106
107
0
    cur = inpL;
108
109
0
    cur = build_norm(cur,
110
0
            model.output_norm, NULL,
111
0
            LLM_NORM_RMS, -1);
112
113
0
    cb(cur, "result_norm", -1);
114
0
    res->t_embd = cur;
115
116
    // lm_head
117
0
    cur = build_lora_mm(model.output, cur);
118
119
0
    cb(cur, "result_output", -1);
120
0
    res->t_logits = cur;
121
122
0
    ggml_build_forward_expand(gf, cur);
123
0
}