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 KiB134 linesraw
1#include "models.h"
2
3void llama_model_neo_bert::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    if (hparams.n_layer() == 28) {
7        type = LLM_TYPE_250M;
8    }
9}
10
11void llama_model_neo_bert::load_arch_tensors(llama_model_loader &) {
12    LLAMA_LOAD_LOCALS;
13
14    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
15
16    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
17    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
18
19    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
20    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
21
22    output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}, 0);
23
24    for (int i = 0; i < n_layer; ++i) {
25        auto & layer = layers[i];
26
27        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
28
29        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
30        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
31
32        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
33
34        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff*2}, 0);
35        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
36    }
37}
38
39std::unique_ptr<llm_graph_context> llama_model_neo_bert::build_arch_graph(const llm_graph_params & params) const {
40    return std::make_unique<graph>(*this, params);
41}
42
43llama_model_neo_bert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
44    const int64_t n_embd_head = hparams.n_embd_head_v();
45
46    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
47
48    ggml_tensor * cur;
49    ggml_tensor * inpL;
50    ggml_tensor * inp_pos = build_inp_pos();
51
52    // construct input embeddings (token, type, position)
53    inpL = build_inp_embd(model.tok_embd);
54    cb(inpL, "inp_embd", -1);
55
56    auto * inp_attn = build_attn_inp_no_cache();
57
58    ggml_tensor * inp_out_ids = build_inp_out_ids();
59
60    for (int il = 0; il < n_layer; ++il) {
61        ggml_tensor * cur = inpL;
62
63        // pre-norm
64        cur = build_norm(inpL,
65                model.layers[il].attn_norm, NULL,
66                LLM_NORM_RMS, il);
67
68        {
69            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
70                    n_embd_head, n_head, n_head_kv, il);
71
72            // RoPE
73            Qcur = ggml_rope_ext(
74                    ctx0, Qcur, inp_pos, nullptr,
75                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
76                    ext_factor, attn_factor, beta_fast, beta_slow
77                    );
78
79            Kcur = ggml_rope_ext(
80                    ctx0, Kcur, inp_pos, nullptr,
81                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82                    ext_factor, attn_factor, beta_fast, beta_slow
83                    );
84
85            cb(Qcur, "Qcur", il);
86            cb(Kcur, "Kcur", il);
87            cb(Vcur, "Vcur", il);
88
89            cur = build_attn(inp_attn,
90                    model.layers[il].wo, nullptr, model.layers[il].wo_s,
91                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
92            cb(cur, "kqv_out", il);
93        }
94        if (il == n_layer - 1 && inp_out_ids) {
95            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
96            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
97        }
98        // re-add the layer input
99        cur = ggml_add(ctx0, cur, inpL);
100
101        ggml_tensor * ffn_inp = cur;
102        cb(ffn_inp, "ffn_inp", il);
103
104        // pre-norm
105        cur = build_norm(ffn_inp,
106                model.layers[il].ffn_norm, NULL,
107                LLM_NORM_RMS, il);
108        cb(cur, "ffn_norm", il);
109
110        // feed-forward network
111        cur = build_ffn(cur,
112                model.layers[il].ffn_up,
113                NULL, NULL, NULL, NULL, NULL,
114                model.layers[il].ffn_down,
115                NULL, NULL, NULL,
116                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
117
118        // attentions bypass the intermediate layer
119        cur = ggml_add(ctx0, cur, ffn_inp);
120
121        // input for next layer
122        inpL = cur;
123    }
124    cur = inpL;
125
126    cur = build_norm(cur,
127            model.output_norm_enc, NULL,
128            LLM_NORM_RMS, -1);
129
130    cb(cur, "result_embd", -1);
131    res->t_embd = cur;
132
133    ggml_build_forward_expand(gf, cur);
134}