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 →
5.3 KiB151 linesraw
1#include "models.h"
2
3void llama_model_maincoder::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_1B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_maincoder::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    // if output is NULL, init from the input tok embed
21    if (output == NULL) {
22        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
23    }
24
25    for (int i = 0; i < n_layer; ++i) {
26        auto & layer = layers[i];
27
28        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
29
30        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
31        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
32
33        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
34        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
35
36        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
37        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
38        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
39        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
40    }
41}
42
43std::unique_ptr<llm_graph_context> llama_model_maincoder::build_arch_graph(const llm_graph_params & params) const {
44    return std::make_unique<graph>(*this, params);
45}
46
47llama_model_maincoder::graph::graph(const llama_model & model, const llm_graph_params & params) : 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,
70                model.layers[il].attn_norm, NULL,
71                LLM_NORM_RMS, il);
72        cb(cur, "attn_norm", il);
73
74        // self-attention
75        {
76            // compute Q and K and RoPE them
77            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
78                    n_embd_head, n_head, n_head_kv, il);
79
80            Qcur = ggml_rope_ext(
81                    ctx0, Qcur, inp_pos, nullptr,
82                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
83                    ext_factor, attn_factor, beta_fast, beta_slow
84                    );
85
86            Kcur = ggml_rope_ext(
87                    ctx0, Kcur, inp_pos, nullptr,
88                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
89                    ext_factor, attn_factor, beta_fast, beta_slow
90                    );
91
92            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
93            cb(Qcur, "Qcur_normed", il);
94
95            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
96            cb(Kcur, "Kcur_normed", il);
97
98            cb(Qcur, "Qcur", il);
99            cb(Kcur, "Kcur", il);
100            cb(Vcur, "Vcur", il);
101
102            cur = build_attn(inp_attn,
103                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
104                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
105        }
106        if (il == n_layer - 1 && inp_out_ids) {
107            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
108            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
109        }
110        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
111        cb(ffn_inp, "ffn_inp", il);
112
113        // feed-forward network
114        cur = build_norm(ffn_inp,
115                model.layers[il].ffn_norm, NULL,
116                LLM_NORM_RMS, il);
117        cb(cur, "ffn_norm", il);
118
119        cur = build_ffn(cur,
120                model.layers[il].ffn_up,   NULL, NULL,
121                model.layers[il].ffn_gate, NULL, NULL,
122                model.layers[il].ffn_down, NULL, NULL,
123                NULL,
124                LLM_FFN_SILU, LLM_FFN_PAR, il);
125        cb(cur, "ffn_out", il);
126
127        cur = ggml_add(ctx0, cur, ffn_inp);
128
129        cur = build_cvec(cur, il);
130        cb(cur, "l_out", il);
131
132        // input for next layer
133        inpL = cur;
134    }
135    cur = inpL;
136
137    cur = build_norm(cur,
138            model.output_norm, NULL,
139            LLM_NORM_RMS, -1);
140
141    cb(cur, "result_norm", -1);
142    res->t_embd = cur;
143
144    // lm_head
145    cur = build_lora_mm(model.output, cur, model.output_s);
146
147    cb(cur, "result_output", -1);
148    res->t_logits = cur;
149
150    ggml_build_forward_expand(gf, cur);
151}