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 KiB153 linesraw
1#include "models.h"
2
3void llama_model_codeshell::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5
6    switch (hparams.n_layer()) {
7        case 42: type = LLM_TYPE_7B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_codeshell::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}, TENSOR_NOT_REQUIRED);
16
17    // if tok embd is NULL, init from output
18    if (tok_embd == NULL) {
19        tok_embd = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
20    }
21
22    // output
23    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
24    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
25    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
31        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
32
33        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
34
35        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
36        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
37
38        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
39        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
40
41        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
42        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
43
44        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i),   {n_embd, n_ff}, 0);
45        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff}, 0);
46    }
47}
48
49std::unique_ptr<llm_graph_context> llama_model_codeshell::build_arch_graph(const llm_graph_params & params) const {
50    return std::make_unique<graph>(*this, params);
51}
52
53llama_model_codeshell::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
54    const int64_t n_embd_head = hparams.n_embd_head_v();
55
56    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
57    GGML_ASSERT(n_embd_head == n_rot);
58
59    ggml_tensor * cur;
60    ggml_tensor * inpL;
61
62    inpL = build_inp_embd(model.tok_embd);
63
64    // inp_pos - contains the positions
65    ggml_tensor * inp_pos = build_inp_pos();
66
67    auto * inp_attn = build_attn_inp_kv();
68
69    ggml_tensor * inp_out_ids = build_inp_out_ids();
70
71    for (int il = 0; il < n_layer; ++il) {
72        cur = build_norm(inpL,
73                model.layers[il].attn_norm,
74                model.layers[il].attn_norm_b,
75                LLM_NORM, il);
76        cb(cur, "attn_norm", il);
77
78        // self-attention
79        {
80            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
81                    n_embd_head, n_head, n_head_kv, il);
82
83            Qcur = ggml_rope_ext(
84                    ctx0, Qcur, inp_pos, nullptr,
85                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
86                    ext_factor, attn_factor, beta_fast, beta_slow
87                    );
88
89            Kcur = ggml_rope_ext(
90                    ctx0, Kcur, inp_pos, nullptr,
91                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
92                    ext_factor, attn_factor, beta_fast, beta_slow
93                    );
94
95            cb(Qcur, "Qcur", il);
96            cb(Kcur, "Kcur", il);
97            cb(Vcur, "Vcur", il);
98
99            cur = build_attn(inp_attn,
100                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
101                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
102        }
103
104        if (il == n_layer - 1 && inp_out_ids) {
105            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
106            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
107        }
108
109        // add the input
110        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
111        cb(ffn_inp, "ffn_inp", il);
112
113        // FF
114        {
115            cur = build_norm(ffn_inp,
116                    model.layers[il].ffn_norm,
117                    model.layers[il].ffn_norm_b,
118                    LLM_NORM, il);
119            cb(cur, "ffn_norm", il);
120
121            cur = build_ffn(cur,
122                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
123                    NULL,                      NULL,                        NULL,
124                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
125                    NULL,
126                    LLM_FFN_GELU, LLM_FFN_SEQ, 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 = build_norm(inpL,
140            model.output_norm,
141            model.output_norm_b,
142            LLM_NORM, -1);
143
144    cb(cur, "result_norm", -1);
145    res->t_embd = cur;
146
147    cur = build_lora_mm(model.output, cur, model.output_s);
148
149    cb(cur, "result_output", -1);
150    res->t_logits = cur;
151
152    ggml_build_forward_expand(gf, cur);
153}