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 →
5.1 KiB155 linesraw
1#include "models.h"
2
3void llama_model_baichuan::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    switch (hparams.n_layer()) {
6        case 32: type = LLM_TYPE_7B; break;
7        case 40: type = LLM_TYPE_13B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10
11    if (type == LLM_TYPE_13B) {
12        // TODO: become GGUF KV parameter
13        hparams.f_max_alibi_bias = 8.0f;
14    }
15}
16
17void llama_model_baichuan::load_arch_tensors(llama_model_loader &) {
18    LLAMA_LOAD_LOCALS;
19
20    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
21    {
22        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23        output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
24    }
25
26    for (int i = 0; i < n_layer; ++i) {
27        auto & layer = layers[i];
28
29        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
30
31        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
32        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
33
34        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {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_baichuan::build_arch_graph(const llm_graph_params & params) const {
43    return std::make_unique<graph>(*this, params);
44}
45
46llama_model_baichuan::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
47    const int64_t n_embd_head = hparams.n_embd_head_v();
48
49    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
50    GGML_ASSERT(n_embd_head == n_rot);
51
52    ggml_tensor * cur;
53    ggml_tensor * inpL;
54
55    inpL = build_inp_embd(model.tok_embd);
56
57    // inp_pos - contains the positions
58    ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
59
60    auto * inp_attn = build_attn_inp_kv();
61
62    ggml_tensor * inp_out_ids = build_inp_out_ids();
63
64    for (int il = 0; il < n_layer; ++il) {
65        ggml_tensor * inpSA = inpL;
66
67        cur = build_norm(inpL,
68                model.layers[il].attn_norm, NULL,
69                LLM_NORM_RMS, il);
70        cb(cur, "attn_norm", il);
71
72        // self-attention
73        {
74            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
75                    n_embd_head, n_head, n_head_kv, il);
76
77            switch (model.type) {
78                case LLM_TYPE_7B:
79                    Qcur = ggml_rope_ext(
80                            ctx0, Qcur, inp_pos, nullptr,
81                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82                            ext_factor, attn_factor, beta_fast, beta_slow
83                            );
84                    Kcur = ggml_rope_ext(
85                            ctx0, Kcur, inp_pos, nullptr,
86                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
87                            ext_factor, attn_factor, beta_fast, beta_slow
88                            );
89                    break;
90                case LLM_TYPE_13B:
91                case LLM_TYPE_UNKNOWN:
92                    break;
93                default:
94                    GGML_ABORT("fatal error");
95            }
96
97            cb(Qcur, "Qcur", il);
98            cb(Kcur, "Kcur", il);
99            cb(Vcur, "Vcur", il);
100
101            cur = build_attn(inp_attn,
102                    model.layers[il].wo, NULL, model.layers[il].wo_s,
103                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
104        }
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
111        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
112        cb(ffn_inp, "ffn_inp", il);
113
114        // feed-forward network
115        {
116            cur = build_norm(ffn_inp,
117                    model.layers[il].ffn_norm, NULL,
118                    LLM_NORM_RMS, il);
119            cb(cur, "ffn_norm", il);
120
121            cur = build_ffn(cur,
122                    model.layers[il].ffn_up,   NULL, NULL,
123                    model.layers[il].ffn_gate, NULL, NULL,
124                    model.layers[il].ffn_down, NULL, NULL,
125                    NULL,
126                    LLM_FFN_SILU, LLM_FFN_PAR, il);
127            cb(cur, "ffn_out", il);
128        }
129
130        cur = ggml_add(ctx0, cur, ffn_inp);
131
132        cur = build_cvec(cur, il);
133        cb(cur, "l_out", il);
134
135        // input for next layer
136        inpL = cur;
137    }
138
139    cur = inpL;
140
141    cur = build_norm(cur,
142            model.output_norm, NULL,
143            LLM_NORM_RMS, -1);
144
145    cb(cur, "result_norm", -1);
146    res->t_embd = cur;
147
148    // lm_head
149    cur = build_lora_mm(model.output, cur, model.output_s);
150
151    cb(cur, "result_output", -1);
152    res->t_logits = cur;
153
154    ggml_build_forward_expand(gf, cur);
155}