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 →
4.7 KiB139 linesraw
1#include "models.h"
2
3void llama_model_gemma::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 18: type = LLM_TYPE_2B; break;
8        case 28: type = LLM_TYPE_7B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10   }
11}
12
13void llama_model_gemma::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
18    // output
19    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
20    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
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_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
28        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, 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_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
33        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
34    }
35}
36
37std::unique_ptr<llm_graph_context> llama_model_gemma::build_arch_graph(const llm_graph_params & params) const {
38    return std::make_unique<graph>(*this, params);
39}
40
41llama_model_gemma::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_tensor * cur;
45    ggml_tensor * inpL;
46
47    inpL = build_inp_embd(model.tok_embd);
48
49    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
50    cb(inpL, "inp_scaled", -1);
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        // norm
61        cur = build_norm(inpL,
62                model.layers[il].attn_norm, NULL,
63                LLM_NORM_RMS, il);
64        cb(cur, "attn_norm", il);
65
66        // self-attention
67        {
68            // compute Q and K and RoPE them
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            Kcur = ggml_rope_ext(
78                    ctx0, Kcur, inp_pos, nullptr,
79                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
80                    ext_factor, attn_factor, beta_fast, beta_slow);
81
82            cb(Qcur, "Qcur", il);
83            cb(Kcur, "Kcur", il);
84            cb(Vcur, "Vcur", il);
85
86            Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
87            cb(Qcur, "Qcur_scaled", il);
88
89            cur = build_attn(inp_attn,
90                    model.layers[il].wo, NULL, model.layers[il].wo_s,
91                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
92        }
93        if (il == n_layer - 1 && inp_out_ids) {
94            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
95            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
96        }
97        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
98        cb(sa_out, "sa_out", il);
99
100        cur = build_norm(sa_out,
101                model.layers[il].ffn_norm, NULL,
102                LLM_NORM_RMS, il);
103        cb(cur, "ffn_norm", il);
104
105        // feed-forward network
106        {
107            cur = build_ffn(cur,
108                    model.layers[il].ffn_up,   NULL, NULL,
109                    model.layers[il].ffn_gate, NULL, NULL,
110                    model.layers[il].ffn_down, NULL, NULL,
111                    NULL,
112                    LLM_FFN_GELU, LLM_FFN_PAR, il);
113            cb(cur, "ffn_out", il);
114        }
115        cur = ggml_add(ctx0, cur, sa_out);
116
117        cur = build_cvec(cur, il);
118        cb(cur, "l_out", il);
119
120        // input for next layer
121        inpL = cur;
122    }
123    cur = inpL;
124
125    cur = build_norm(cur,
126            model.output_norm, NULL,
127            LLM_NORM_RMS, -1);
128
129    cb(cur, "result_norm", -1);
130    res->t_embd = cur;
131
132    // lm_head
133    cur = build_lora_mm(model.output, cur, model.output_s);
134
135    cb(cur, "result_output", -1);
136    res->t_logits = cur;
137
138    ggml_build_forward_expand(gf, cur);
139}