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 →
4.6 KiB136 linesraw
1#include "models.h"
2
3void llama_model_xverse::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_7B; break;
8        case 40: type = LLM_TYPE_13B; break;
9        case 80: type = LLM_TYPE_65B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_xverse::load_arch_tensors(llama_model_loader &) {
15    LLAMA_LOAD_LOCALS;
16
17    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
18
19    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
20    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
21
22    for (int i = 0; i < n_layer; ++i) {
23        auto & layer = layers[i];
24
25        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
26
27        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
28        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
29
30        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
31        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
32        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
33        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
34    }
35}
36
37std::unique_ptr<llm_graph_context> llama_model_xverse::build_arch_graph(const llm_graph_params & params) const {
38    return std::make_unique<graph>(*this, params);
39}
40
41llama_model_xverse::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
42    const int64_t n_embd_head = hparams.n_embd_head_v();
43
44    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
45    GGML_ASSERT(n_embd_head == n_rot);
46
47    ggml_tensor * cur;
48    ggml_tensor * inpL;
49
50    inpL = build_inp_embd(model.tok_embd);
51
52    // inp_pos - contains the positions
53    ggml_tensor * inp_pos = build_inp_pos();
54
55    auto * inp_attn = build_attn_inp_kv();
56
57    ggml_tensor * inp_out_ids = build_inp_out_ids();
58
59    for (int il = 0; il < n_layer; ++il) {
60        ggml_tensor * inpSA = inpL;
61
62        cur = build_norm(inpL,
63                model.layers[il].attn_norm, NULL,
64                LLM_NORM_RMS, il);
65        cb(cur, "attn_norm", il);
66
67        // self-attention
68        {
69            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
70                    n_embd_head, n_head, n_head_kv, il);
71
72            Qcur = ggml_rope_ext(
73                    ctx0, Qcur, inp_pos, nullptr,
74                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
75                    ext_factor, attn_factor, beta_fast, beta_slow
76                    );
77
78            Kcur = ggml_rope_ext(
79                    ctx0, Kcur, inp_pos, nullptr,
80                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
81                    ext_factor, attn_factor, beta_fast, beta_slow
82                    );
83
84            cb(Qcur, "Qcur", il);
85            cb(Kcur, "Kcur", il);
86            cb(Vcur, "Vcur", il);
87
88            cur = build_attn(inp_attn,
89                    model.layers[il].wo, NULL, model.layers[il].wo_s,
90                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
91        }
92        if (il == n_layer - 1 && inp_out_ids) {
93            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
94            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
95        }
96        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
97        cb(ffn_inp, "ffn_inp", il);
98
99        // feed-forward network
100        {
101            cur = build_norm(ffn_inp,
102                    model.layers[il].ffn_norm, NULL,
103                    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,
111                    LLM_FFN_SILU, LLM_FFN_PAR, il);
112            cb(cur, "ffn_out", il);
113        }
114        cur = ggml_add(ctx0, cur, ffn_inp);
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}