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 →
6.2 KiB171 linesraw
1#include "models.h"
2
3void llama_model_openelm::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 16: type = LLM_TYPE_270M; break;
8        case 20: type = LLM_TYPE_450M; break;
9        case 28: type = LLM_TYPE_1B; break;
10        case 36: type = LLM_TYPE_3B; break;
11        default: type = LLM_TYPE_UNKNOWN;
12    }
13}
14
15void llama_model_openelm::load_arch_tensors(llama_model_loader &) {
16    LLAMA_LOAD_LOCALS;
17
18    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
19
20    // output
21    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
22    // init output from the input tok embed
23    output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
24
25    for (int i = 0; i < n_layer; ++i) {
26        const int64_t n_head      =   hparams.n_head(i);
27        const int64_t n_head_qkv  = 2*hparams.n_head_kv(i) + n_head;
28        const int64_t n_ff        =   hparams.n_ff(i);
29
30        auto & layer = layers[i];
31
32        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
33
34        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_head_qkv*n_embd_head_k}, 0);
35        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
36        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
37        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head*n_embd_head_k, n_embd}, 0);
38
39        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
40        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
41        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
42        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
43    }
44}
45
46std::unique_ptr<llm_graph_context> llama_model_openelm::build_arch_graph(const llm_graph_params & params) const {
47    return std::make_unique<graph>(*this, params);
48}
49
50llama_model_openelm::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
51    const int64_t n_embd_head = hparams.n_embd_head_v();
52
53    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
54
55    ggml_tensor * cur;
56    ggml_tensor * inpL;
57    inpL = build_inp_embd(model.tok_embd);
58
59    // inp_pos - contains the positions
60    ggml_tensor * inp_pos = build_inp_pos();
61
62    auto * inp_attn = build_attn_inp_kv();
63
64    ggml_tensor * inp_out_ids = build_inp_out_ids();
65
66    for (int il = 0; il < n_layer; ++il) {
67        const int64_t n_head    = hparams.n_head(il);
68        const int64_t n_head_kv = hparams.n_head_kv(il);
69        const int64_t n_head_qkv = 2*n_head_kv + n_head;
70
71        cur = inpL;
72        ggml_tensor * residual = cur;
73
74        // norm
75        cur = build_norm(inpL,
76                model.layers[il].attn_norm, NULL,
77                LLM_NORM_RMS, il);
78        cb(cur, "attn_norm", il);
79
80        // self-attention
81        {
82            cur = build_lora_mm(model.layers[il].wqkv, cur);
83            cb(cur, "wqkv", il);
84
85            cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);
86
87            ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head,    n_tokens, cur->nb[1], cur->nb[2], 0);
88            cb(Qcur, "Qcur", il);
89
90            ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head);
91            cb(Kcur, "Kcur", il);
92
93            ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv));
94            cb(Vcur, "Vcur", il);
95
96            Qcur = build_norm(Qcur,
97                    model.layers[il].attn_q_norm, NULL,
98                    LLM_NORM_RMS, il);
99            cb(Qcur, "Qcur", il);
100
101            Kcur = build_norm(Kcur,
102                    model.layers[il].attn_k_norm, NULL,
103                    LLM_NORM_RMS, il);
104            cb(Kcur, "Kcur", il);
105
106            Qcur = ggml_rope_ext(
107                    ctx0, Qcur, inp_pos, NULL,
108                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
109                    ext_factor, attn_factor, beta_fast, beta_slow
110                    );
111
112            Kcur = ggml_rope_ext(
113                    ctx0, Kcur, inp_pos, NULL,
114                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
115                    ext_factor, attn_factor, beta_fast, beta_slow
116                    );
117
118            cb(Qcur, "Qcur", il);
119            cb(Kcur, "Kcur", il);
120            cb(Qcur, "Vcur", il);
121
122            cur = build_attn(inp_attn,
123                    model.layers[il].wo, NULL, model.layers[il].wo_s,
124                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
125        }
126        if (il == n_layer - 1 && inp_out_ids) {
127            residual = ggml_get_rows(ctx0, residual, inp_out_ids);
128            cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
129        }
130        ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
131        cb(ffn_inp, "ffn_inp", il);
132
133        // feed-forward network
134        {
135            cur = build_norm(ffn_inp,
136                    model.layers[il].ffn_norm, NULL,
137                    LLM_NORM_RMS, il);
138            cb(cur, "ffn_norm", il);
139
140            cur = build_ffn(cur,
141                    model.layers[il].ffn_up,   NULL, NULL,
142                    model.layers[il].ffn_gate, NULL, NULL,
143                    model.layers[il].ffn_down, NULL, NULL,
144                    NULL,
145                    LLM_FFN_SILU, LLM_FFN_PAR, il);
146            cb(cur, "ffn_out", il);
147        }
148        cur = ggml_add(ctx0, cur, ffn_inp);
149
150        cur = build_cvec(cur, il);
151        cb(cur, "l_out", il);
152
153        inpL = cur;
154    }
155    cur = inpL;
156
157    // norm
158    cur = build_norm(cur,
159            model.output_norm, NULL,
160            LLM_NORM_RMS, -1);
161
162    cb(cur, "result_norm", -1);
163    res->t_embd = cur;
164
165    cur = build_lora_mm(model.output, cur, model.output_s);
166
167    cb(cur, "result_output", -1);
168    res->t_logits = cur;
169
170    ggml_build_forward_expand(gf, cur);
171}