slm_ikllama_sys 0.1.1

ik_llama.cpp rust sys bindings
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120

#include "../llama-build-context.h"
#include "../llama-model.h"
#include "../llama-context.h"

ggml_cgraph * llm_build_context::build_arctic() {
    ggml_cgraph * gf = new_graph_custom();

    // mutable variable, needed during the last layer of the computation to skip unused tokens
    int32_t n_tokens = this->n_tokens;

    const int64_t n_embd_head = hparams.n_embd_head_v(0);
    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k(0));
    GGML_ASSERT(n_embd_head == hparams.n_rot);

    struct ggml_tensor * cur;
    struct ggml_tensor * inpL;

    inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);

    // inp_pos - contains the positions
    struct ggml_tensor * inp_pos = build_inp_pos();

    // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
    struct ggml_tensor * KQ_mask = build_inp_KQ_mask();

    for (int il = 0; il < n_layer; ++il) {
        struct ggml_tensor * inpSA = inpL;

        // norm
        cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
        cb(cur, "attn_norm", il);

        // self-attention
        {
            auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, model.layers[il].wq, nullptr,
                    model.layers[il].wk, nullptr,
                    model.layers[il].wv, nullptr, 0, il);

            Qcur = ggml_rope_ext(
                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );
            cb(Qcur, "Qcur", il);

            Kcur = ggml_rope_ext(
                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );
            cb(Kcur, "Kcur", il);

            cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                    model.layers[il].wo, NULL,
                    Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
        }

        if (il == n_layer - 1) {
            // skip computing output for unused tokens
            struct ggml_tensor * inp_out_ids = build_inp_out_ids();
            n_tokens = n_outputs;
            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }

        struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);

        // feed-forward network
        cur = llm_build_ffn(ctx0, lctx, model.layers[il].ffn_norm, ffn_inp,
                model.layers[il].ffn_up,   NULL, NULL,
                model.layers[il].ffn_gate, NULL, NULL,
                model.layers[il].ffn_down, NULL, NULL,
                NULL,
                LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
        cb(cur, "ffn_out", il);

        struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
        cb(ffn_out, "ffn_out", il);

        // MoE
        cur = llm_build_norm(ctx0, inpSA, hparams, model.layers[il].ffn_norm_exps, NULL, LLM_NORM_RMS, cb, il);
        cb(cur, "ffn_norm_exps", il);

        cur = llm_build_moe_ffn(ctx0, lctx, cur,
                model.layers[il].ffn_gate_inp,
                model.layers[il].ffn_up_exps,
                model.layers[il].ffn_gate_exps,
                model.layers[il].ffn_down_exps,
                nullptr,
                n_expert, n_expert_used,
                LLM_FFN_SILU, true,
                false, 0.0,
                LLM_EXPERT_GATING_FUNC_SOFTMAX,
                cb, il, gf);
        cb(cur, "ffn_moe_out", il);

        cur = ggml_add(ctx0, cur, ffn_out);
        cb(cur, "ffn_out", il);

        cur = lctx.cvec.apply_to(ctx0, cur, il);
        cb(cur, "l_out", il);

        // input for next layer
        inpL = cur;
    }

    cur = inpL;

    cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1);
    cb(cur, "result_norm", -1);

    // lm_head
    cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
    cb(cur, "result_output", -1);

    ggml_build_forward_expand(gf, cur);

    return gf;
}