Coverage Report

Created: 2026-01-11 07:13

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/llama.cpp/src/models/stablelm.cpp
Line
Count
Source
1
#include "models.h"
2
3
0
llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
0
    const int64_t n_embd_head = hparams.n_embd_head_v;
5
6
0
    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
7
8
0
    ggml_tensor * cur;
9
0
    ggml_tensor * inpL;
10
11
0
    inpL = build_inp_embd(model.tok_embd);
12
13
    // inp_pos - contains the positions
14
0
    ggml_tensor * inp_pos = build_inp_pos();
15
16
0
    auto * inp_attn = build_attn_inp_kv();
17
18
0
    ggml_tensor * inp_out_ids = build_inp_out_ids();
19
20
0
    for (int il = 0; il < n_layer; ++il) {
21
        // norm
22
0
        cur = build_norm(inpL,
23
0
                model.layers[il].attn_norm,
24
0
                model.layers[il].attn_norm_b,
25
0
                LLM_NORM, il);
26
0
        cb(cur, "attn_norm", il);
27
28
0
        ggml_tensor * inpSA = cur;
29
30
        // self-attention
31
0
        {
32
            // compute Q and K and RoPE them
33
0
            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
34
0
            cb(Qcur, "Qcur", il);
35
0
            if (model.layers[il].bq) {
36
0
                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
37
0
                cb(Qcur, "Qcur", il);
38
0
            }
39
40
0
            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
41
0
            cb(Kcur, "Kcur", il);
42
0
            if (model.layers[il].bk) {
43
0
                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
44
0
                cb(Kcur, "Kcur", il);
45
0
            }
46
47
0
            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
48
0
            cb(Vcur, "Vcur", il);
49
0
            if (model.layers[il].bv) {
50
0
                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
51
0
                cb(Vcur, "Vcur", il);
52
0
            }
53
54
0
            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
55
0
            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
56
0
            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
57
58
0
            if (model.layers[il].attn_q_norm) {
59
0
                Qcur = build_norm(Qcur,
60
0
                        model.layers[il].attn_q_norm,
61
0
                        NULL,
62
0
                        LLM_NORM, il);
63
0
                cb(Qcur, "Qcur", il);
64
0
            }
65
0
            if (model.layers[il].attn_k_norm) {
66
0
                Kcur = build_norm(Kcur,
67
0
                        model.layers[il].attn_k_norm,
68
0
                        NULL,
69
0
                        LLM_NORM, il);
70
0
                cb(Kcur, "Kcur", il);
71
0
            }
72
73
0
            Qcur = ggml_rope_ext(
74
0
                    ctx0, Qcur, inp_pos, nullptr,
75
0
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
76
0
                    ext_factor, attn_factor, beta_fast, beta_slow
77
0
                    );
78
79
0
            Kcur = ggml_rope_ext(
80
0
                    ctx0, Kcur, inp_pos, nullptr,
81
0
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82
0
                    ext_factor, attn_factor, beta_fast, beta_slow
83
0
                    );
84
85
0
            cb(Qcur, "Qcur", il);
86
0
            cb(Kcur, "Kcur", il);
87
0
            cb(Vcur, "Vcur", il);
88
89
0
            cur = build_attn(inp_attn,
90
0
                    model.layers[il].wo, NULL,
91
0
                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
92
0
        }
93
0
        if (il == n_layer - 1 && inp_out_ids) {
94
0
            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
95
0
            inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
96
0
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
97
0
        }
98
0
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
99
0
        cb(ffn_inp, "ffn_inp", il);
100
101
        // feed-forward network
102
0
        {
103
0
            if (model.layers[il].ffn_norm) {
104
0
                cur = build_norm(ffn_inp,
105
0
                        model.layers[il].ffn_norm,
106
0
                        model.layers[il].ffn_norm_b,
107
0
                        LLM_NORM, il);
108
0
                cb(cur, "ffn_norm", il);
109
0
            } else {
110
                // parallel residual
111
0
                cur = inpSA;
112
0
            }
113
0
            cur = build_ffn(cur,
114
0
                    model.layers[il].ffn_up,   NULL, NULL,
115
0
                    model.layers[il].ffn_gate, NULL, NULL,
116
0
                    model.layers[il].ffn_down, NULL, NULL,
117
0
                    NULL,
118
0
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
119
0
            cb(cur, "ffn_out", il);
120
0
        }
121
0
        cur = ggml_add(ctx0, cur, ffn_inp);
122
123
0
        cur = build_cvec(cur, il);
124
0
        cb(cur, "l_out", il);
125
126
        // input for next layer
127
0
        inpL = cur;
128
0
    }
129
0
    cur = inpL;
130
131
0
    cur = build_norm(cur,
132
0
            model.output_norm,
133
0
            model.output_norm_b,
134
0
            LLM_NORM, -1);
135
136
0
    cb(cur, "result_norm", -1);
137
0
    res->t_embd = cur;
138
139
    // lm_head
140
0
    cur = build_lora_mm(model.output, cur);
141
142
0
    cb(cur, "result_output", -1);
143
0
    res->t_logits = cur;
144
145
0
    ggml_build_forward_expand(gf, cur);
146
0
}