char-fork/llama.cpp

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

Georgi Gerganovhparams : refactor `hparams.n_layer` (#24060)7acb4e8cd

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
5.1 KiB149 linesraw
1#include "models.h"
2
3void llama_model_talkie::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
6
7    switch (hparams.n_layer()) {
8        case 40: type = LLM_TYPE_13B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_talkie::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
17    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), {n_embd, n_vocab}, 0);
18
19    for (int i = 0; i < n_layer; ++i) {
20        auto & layer = layers[i];
21
22        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
23        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
24
25        // no k gain
26        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {1, n_head}, 0);
27
28        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
29        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
30        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
31
32        layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1}, 0);
33    }
34}
35
36std::unique_ptr<llm_graph_context> llama_model_talkie::build_arch_graph(const llm_graph_params & params) const {
37    return std::make_unique<graph>(*this, params);
38}
39
40llama_model_talkie::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
41    const int64_t n_embd_head = hparams.n_embd_head_k();
42
43    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
44    GGML_ASSERT(n_embd_head == n_rot);
45
46    ggml_tensor * cur;
47    ggml_tensor * inpL;
48
49    inpL = build_inp_embd(model.tok_embd);
50    inpL = build_norm(inpL, nullptr, nullptr, LLM_NORM_RMS, -1);
51    cb(inpL, "inp_norm", -1);
52
53    ggml_tensor * embd_skip = inpL;
54
55    // inp_pos - contains the positions
56    ggml_tensor * inp_pos = build_inp_pos();
57
58    auto * inp_attn = build_attn_inp_kv();
59
60    ggml_tensor * inp_out_ids = build_inp_out_ids();
61
62    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
63
64    for (int il = 0; il < n_layer; ++il) {
65        ggml_tensor * inpSA = inpL;
66        ggml_tensor * inp_skip = embd_skip;
67
68        cur = build_norm(inpL, nullptr, nullptr, LLM_NORM_RMS, il);
69        cb(cur, "attn_norm", il);
70
71        // self-attention
72        {
73            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
74                    n_embd_head, n_head, n_head_kv, il);
75
76            Qcur = ggml_rope_ext(
77                    ctx0, Qcur, inp_pos, nullptr,
78                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
79                    ext_factor, attn_factor, beta_fast, beta_slow);
80
81            Kcur = ggml_rope_ext(
82                    ctx0, Kcur, inp_pos, nullptr,
83                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
84                    ext_factor, attn_factor, beta_fast, beta_slow);
85
86            // reference applies qknorm after rope
87            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
88            cb(Qcur, "Qcur_norm", il);
89
90            Kcur = build_norm(Kcur, nullptr, nullptr, LLM_NORM_RMS, il);
91            cb(Kcur, "Kcur_norm", il);
92
93            cb(Vcur, "Vcur", il);
94
95            cur = build_attn(inp_attn,
96                    model.layers[il].wo, nullptr, model.layers[il].wo_s,
97                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
98            cb(cur, "attn_out", il);
99        }
100
101        if (il == n_layer - 1 && inp_out_ids) {
102            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
103            inpSA    = ggml_get_rows(ctx0, inpSA,    inp_out_ids);
104            inp_skip = ggml_get_rows(ctx0, inp_skip, inp_out_ids);
105        }
106
107        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
108        cb(ffn_inp, "ffn_inp", il);
109
110        cur = build_norm(ffn_inp, nullptr, nullptr, LLM_NORM_RMS, il);
111        cb(cur, "ffn_norm", il);
112
113        cur = build_ffn(cur,
114                model.layers[il].ffn_up,   nullptr, nullptr,
115                model.layers[il].ffn_gate, nullptr, nullptr,
116                model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,
117                nullptr,
118                LLM_FFN_SILU, LLM_FFN_PAR, il);
119        cb(cur, "ffn_out", il);
120
121        cur = ggml_add(ctx0, cur, ffn_inp);
122
123        ggml_tensor * skip = ggml_mul(ctx0, inp_skip, model.layers[il].out_scale);
124        cb(skip, "embd_skip", il);
125
126        cur = ggml_add(ctx0, cur, skip);
127
128        cur = build_cvec(cur, il);
129        cb(cur, "l_out", il);
130
131        // input for next layer
132        inpL = cur;
133    }
134
135    cur = inpL;
136
137    cur = build_norm(cur, nullptr, nullptr, LLM_NORM_RMS, -1);
138    cb(cur, "result_norm", -1);
139
140    res->t_embd = cur;
141
142    cur = build_lora_mm(model.output, cur);
143    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
144    cb(cur, "result_output", -1);
145
146    res->t_logits = cur;
147
148    ggml_build_forward_expand(gf, cur);
149}