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 →
4.5 KiB136 linesraw
1#include "models.h"
2
3void llama_model_plamo::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 40: type = LLM_TYPE_13B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9   }
10}
11
12void llama_model_plamo::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}, 0);
20
21    for (int i = 0; i < n_layer; ++i) {
22        auto & layer = layers[i];
23
24        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
25
26        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
27        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
28
29        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "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        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
32    }
33}
34
35std::unique_ptr<llm_graph_context> llama_model_plamo::build_arch_graph(const llm_graph_params & params) const {
36    return std::make_unique<graph>(*this, params);
37}
38
39llama_model_plamo::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
40    const int64_t n_embd_head = hparams.n_embd_head_v();
41
42    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
43    GGML_ASSERT(n_embd_head == n_rot);
44
45    ggml_tensor * cur;
46    ggml_tensor * inpL;
47
48    inpL = build_inp_embd(model.tok_embd);
49
50    // inp_pos - contains the positions
51    ggml_tensor * inp_pos = build_inp_pos();
52
53    auto * inp_attn = build_attn_inp_kv();
54
55    ggml_tensor * inp_out_ids = build_inp_out_ids();
56
57    for (int il = 0; il < n_layer; ++il) {
58        // norm
59        cur = build_norm(inpL,
60                model.layers[il].attn_norm, NULL,
61                LLM_NORM_RMS, il);
62        cb(cur, "attn_norm", il);
63
64        ggml_tensor * sa_inp = cur;
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_embd_head, 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_embd_head, 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            sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids);
95            inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
96        }
97        ggml_tensor * sa_out = cur;
98
99        cur = sa_inp;
100
101        // feed-forward network
102        {
103            cur = build_ffn(cur,
104                    model.layers[il].ffn_up,   NULL, NULL,
105                    model.layers[il].ffn_gate, NULL, NULL,
106                    model.layers[il].ffn_down, NULL, NULL,
107                    NULL,
108                    LLM_FFN_SILU, LLM_FFN_PAR, il);
109            cb(cur, "ffn_out", il);
110        }
111        cur = ggml_add(ctx0, cur, sa_out);
112        cur = ggml_add(ctx0, cur, inpL);
113
114        cur = build_cvec(cur, il);
115        cb(cur, "l_out", il);
116
117        // input for next layer
118        inpL = cur;
119    }
120    cur = inpL;
121
122    cur = build_norm(cur,
123            model.output_norm, NULL,
124            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}