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 KiB142 linesraw
1#include "models.h"
2
3void llama_model_olmo::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5    ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,     hparams.f_clamp_kqv, false);
6
7    switch (hparams.n_layer()) {
8        case 22: type = LLM_TYPE_1B; break;
9        case 32: type = LLM_TYPE_7B; break;
10        case 80: type = LLM_TYPE_70B; break;
11        default: type = LLM_TYPE_UNKNOWN;
12    }
13}
14
15void llama_model_olmo::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 = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
22    // if output is NULL, init from the input tok embed
23    if (output == NULL) {
24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25    }
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
31        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
32
33        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
34        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
35        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
36    }
37}
38
39std::unique_ptr<llm_graph_context> llama_model_olmo::build_arch_graph(const llm_graph_params & params) const {
40    return std::make_unique<graph>(*this, params);
41}
42
43llama_model_olmo::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    GGML_ASSERT(n_embd_head == n_rot);
48
49    ggml_tensor * cur;
50    ggml_tensor * inpL;
51
52    inpL = build_inp_embd(model.tok_embd);
53
54    // inp_pos - contains the positions
55    ggml_tensor * inp_pos = build_inp_pos();
56
57    auto * inp_attn = build_attn_inp_kv();
58
59    ggml_tensor * inp_out_ids = build_inp_out_ids();
60
61    for (int il = 0; il < n_layer; ++il) {
62        ggml_tensor * inpSA = inpL;
63
64        // norm
65        cur = build_norm(inpL,
66                NULL, NULL,
67                LLM_NORM, il);
68        cb(cur, "attn_norm", il);
69
70        // self-attention
71        {
72            // compute Q and K and RoPE them
73            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
74                    n_embd_head, n_head, n_head_kv, il);
75
76            Qcur = ggml_rope_ext(
77                    ctx0, Qcur, inp_pos, nullptr,
78                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
79                    ext_factor, attn_factor, beta_fast, beta_slow
80                    );
81
82            Kcur = ggml_rope_ext(
83                    ctx0, Kcur, inp_pos, nullptr,
84                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
85                    ext_factor, attn_factor, beta_fast, beta_slow
86                    );
87
88            cb(Qcur, "Qcur", il);
89            cb(Kcur, "Kcur", il);
90            cb(Vcur, "Vcur", il);
91
92            cur = build_attn(inp_attn,
93                    model.layers[il].wo, nullptr, model.layers[il].wo_s,
94                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
95        }
96        if (il == n_layer - 1 && inp_out_ids) {
97            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
98            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
99        }
100        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
101        cb(ffn_inp, "ffn_inp", il);
102
103        // feed-forward network
104        cur = build_norm(ffn_inp,
105                NULL, NULL,
106                LLM_NORM, il);
107        cb(cur, "ffn_norm", il);
108
109        cur = build_ffn(cur,
110                model.layers[il].ffn_up,   NULL, NULL,
111                model.layers[il].ffn_gate, NULL, NULL,
112                model.layers[il].ffn_down, NULL, NULL,
113                NULL,
114                LLM_FFN_SILU, LLM_FFN_PAR, il);
115        cb(cur, "ffn_out", il);
116
117        cur = ggml_add(ctx0, cur, ffn_inp);
118        cb(cur, "ffn_out", il);
119
120        cur = build_cvec(cur, il);
121        cb(cur, "l_out", il);
122
123        // input for next layer
124        inpL = cur;
125    }
126    cur = inpL;
127
128    cur = build_norm(cur,
129            NULL, NULL,
130            LLM_NORM, -1);
131
132    cb(cur, "result_norm", -1);
133    res->t_embd = cur;
134
135    // lm_head
136    cur = build_lora_mm(model.output, cur, model.output_s);
137
138    cb(cur, "result_output", -1);
139    res->t_logits = cur;
140
141    ggml_build_forward_expand(gf, cur);
142}