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 →
5.0 KiB136 linesraw
1#include "models.h"
2
3void llama_model_exaone::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    switch (hparams.n_layer()) {
7        case 32: type = LLM_TYPE_8B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_exaone::load_arch_tensors(llama_model_loader &) {
13    LLAMA_LOAD_LOCALS;
14
15    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
16
17    // output
18    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
19    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
20
21    // if output is NULL, init from the input tok embed
22    if (output == NULL) {
23        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
24    }
25
26    for (int i = 0; i < n_layer; ++i) {
27        auto & layer = layers[i];
28
29        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
30
31        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
32        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
33
34        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM,   "weight", i), {n_embd}, 0);
35        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
36        layer.ffn_gate   = create_tensor(tn(LLM_TENSOR_FFN_GATE,   "weight", i), {n_embd,   n_ff}, 0);
37        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN,   "weight", i), {  n_ff, n_embd}, 0);
38        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,     "weight", i), {n_embd,   n_ff}, 0);
39    }
40}
41
42std::unique_ptr<llm_graph_context> llama_model_exaone::build_arch_graph(const llm_graph_params & params) const {
43    return std::make_unique<graph>(*this, params);
44}
45
46llama_model_exaone::graph::graph(const llama_model & model, const llm_graph_params & params) :
47    llm_graph_context(params) {
48    const int64_t n_embd_head = hparams.n_embd_head_v();
49
50    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
51    GGML_ASSERT(n_embd_head == n_rot);
52
53    ggml_tensor * cur;
54    ggml_tensor * inpL;
55
56    inpL = build_inp_embd(model.tok_embd);
57
58    // inp_pos - contains the positions
59    ggml_tensor * inp_pos = build_inp_pos();
60
61    auto * inp_attn = build_attn_inp_kv();
62
63    ggml_tensor * inp_out_ids = build_inp_out_ids();
64
65    for (int il = 0; il < n_layer; ++il) {
66        ggml_tensor * inpSA = inpL;
67
68        // norm
69        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
70        cb(cur, "attn_norm", il);
71
72        // self-attention
73        {
74            // rope freq factors for llama3; may return nullptr for llama2 and other models
75            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
76
77            // compute Q and K and RoPE them
78            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
79                    n_embd_head, n_head, n_head_kv, il);
80
81            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82                                 ext_factor, attn_factor, beta_fast, beta_slow);
83
84            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
85                                 ext_factor, attn_factor, beta_fast, beta_slow);
86
87            cb(Qcur, "Qcur", il);
88            cb(Kcur, "Kcur", il);
89            cb(Vcur, "Vcur", il);
90
91            cur = build_attn(inp_attn,
92                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
93                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
94        }
95        if (il == n_layer - 1 && inp_out_ids) {
96            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
97            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
98        }
99        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
100        cb(ffn_inp, "ffn_inp", il);
101
102        // feed-forward network
103        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
104        cb(cur, "ffn_norm", il);
105
106        cur = build_ffn(cur,
107                model.layers[il].ffn_up, NULL, NULL,
108                model.layers[il].ffn_gate, NULL, NULL,
109                model.layers[il].ffn_down, NULL, NULL,
110                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
111        cb(cur, "ffn_out", il);
112
113        cur = ggml_add(ctx0, cur, ffn_inp);
114        cb(cur, "ffn_out", il);
115
116        cur = build_cvec(cur, il);
117        cb(cur, "l_out", il);
118
119        // input for next layer
120        inpL = cur;
121    }
122    cur = inpL;
123
124    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
125
126    cb(cur, "result_norm", -1);
127    res->t_embd = cur;
128
129    // lm_head
130    cur = build_lora_mm(model.output, cur, model.output_s);
131
132    cb(cur, "result_output", -1);
133    res->t_logits = cur;
134
135    ggml_build_forward_expand(gf, cur);
136}