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 →
5.2 KiB144 linesraw
1#include "models.h"
2
3void llama_model_command_r::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_LOGIT_SCALE,             hparams.f_logit_scale, false);
5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
6
7    switch (hparams.n_layer()) {
8        case 40: type = LLM_TYPE_35B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_command_r::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    // init output from the input tok embed
21    output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
22
23    for (int i = 0; i < n_layer; ++i) {
24        auto & layer = layers[i];
25
26        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
27
28        if (n_layer >= 64){
29            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head}, 0);
30            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, 0);
31        }
32
33        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
34        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
35
36        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
37        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
38        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
39    }
40}
41
42std::unique_ptr<llm_graph_context> llama_model_command_r::build_arch_graph(const llm_graph_params & params) const {
43    return std::make_unique<graph>(*this, params);
44}
45
46llama_model_command_r::graph::graph(const llama_model & model, const llm_graph_params & params) :
47    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
52    const float f_logit_scale = hparams.f_logit_scale;
53
54    ggml_tensor * cur;
55    ggml_tensor * inpL;
56
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        // norm
68        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
69        cb(cur, "attn_norm", il);
70
71        ggml_tensor * ffn_inp = cur;
72
73        // self-attention
74        {
75            // compute Q and K and RoPE them
76            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
77                    n_embd_head, n_head, n_head_kv, il);
78
79            if (model.layers[il].attn_q_norm) {
80                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il);
81                cb(Qcur, "Qcur", il);
82            }
83            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
84                                 ext_factor, attn_factor, beta_fast, beta_slow);
85
86            if (model.layers[il].attn_k_norm) {
87                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il);
88                cb(Kcur, "Kcur", il);
89            }
90            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
91                                 ext_factor, attn_factor, beta_fast, beta_slow);
92
93            cb(Qcur, "Qcur", il);
94            cb(Kcur, "Kcur", il);
95            cb(Vcur, "Vcur", il);
96
97            cur = build_attn(inp_attn,
98                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
99                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
100        }
101        if (il == n_layer - 1 && inp_out_ids) {
102            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);
103            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);
104            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
105        }
106        ggml_tensor * attn_out = cur;
107
108        // feed-forward network
109        {
110            cur = build_ffn(ffn_inp,
111                    model.layers[il].ffn_up, NULL, NULL,
112                    model.layers[il].ffn_gate, NULL, NULL,
113                    model.layers[il].ffn_down, NULL, NULL,
114                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
115            cb(cur, "ffn_out", il);
116        }
117        // add together residual + FFN + self-attention
118        cur = ggml_add(ctx0, cur, inpL);
119        cur = ggml_add(ctx0, cur, attn_out);
120
121        cur = build_cvec(cur, il);
122        cb(cur, "l_out", il);
123
124        // input for next layer
125        inpL = cur;
126    }
127    cur = inpL;
128
129    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
130
131    cb(cur, "result_norm", -1);
132    res->t_embd = cur;
133
134    // lm_head
135    cur = build_lora_mm(model.output, cur, model.output_s);
136
137    if (f_logit_scale) {
138        cur = ggml_scale(ctx0, cur, f_logit_scale);
139    }
140    cb(cur, "result_output", -1);
141    res->t_logits = cur;
142
143    ggml_build_forward_expand(gf, cur);
144}