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 →
5.4 KiB151 linesraw
1#include "models.h"
2
3void llama_model_bloom::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 24: type = LLM_TYPE_1B; break;
8        case 30:
9            switch (hparams.n_embd) {
10                case 2560: type = LLM_TYPE_3B; break;
11                case 4096: type = LLM_TYPE_7B; break;
12                default: type = LLM_TYPE_UNKNOWN;
13            } break;
14        default: type = LLM_TYPE_UNKNOWN;
15    }
16
17    // TODO: become GGUF KV parameter
18    hparams.f_max_alibi_bias = 8.0f;
19}
20
21void llama_model_bloom::load_arch_tensors(llama_model_loader &) {
22    LLAMA_LOAD_LOCALS;
23
24    tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
25    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
26    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
27
28    // output
29    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
30    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
31    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
32
33    // if output is NULL, init from the input tok embed
34    if (output == NULL) {
35        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
36    }
37
38    for (int i = 0; i < n_layer; ++i) {
39        auto & layer = layers[i];
40
41        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
42        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), {n_embd}, 0);
43
44        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
45        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
46
47        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
48        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd}, 0);
49
50        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
51        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias",   i), {n_embd}, 0);
52
53        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
54        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
55
56        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
57        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias",   i), {n_ff}, 0);
58    }
59}
60
61std::unique_ptr<llm_graph_context> llama_model_bloom::build_arch_graph(const llm_graph_params & params) const {
62    return std::make_unique<graph>(*this, params);
63}
64
65llama_model_bloom::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
66    const int64_t n_embd_head = hparams.n_embd_head_v();
67
68    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
69
70    ggml_tensor * cur;
71    ggml_tensor * inpL;
72
73    inpL = build_inp_embd(model.tok_embd);
74
75    auto * inp_attn = build_attn_inp_kv();
76
77    inpL = build_norm(inpL,
78            model.tok_norm,
79            model.tok_norm_b,
80            LLM_NORM, 0);
81    cb(inpL, "inp_norm", 0);
82
83    ggml_tensor * inp_out_ids = build_inp_out_ids();
84
85    for (int il = 0; il < n_layer; ++il) {
86        cur = build_norm(inpL,
87                model.layers[il].attn_norm,
88                model.layers[il].attn_norm_b,
89                LLM_NORM, il);
90        cb(cur, "attn_norm", il);
91
92        // self-attention
93        {
94            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
95                    n_embd_head, n_head, n_head_kv, 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
102        if (il == n_layer - 1 && inp_out_ids) {
103            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
104            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
105        }
106
107        // Add the input
108        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
109        cb(ffn_inp, "ffn_inp", il);
110
111        // FF
112        {
113            cur = build_norm(ffn_inp,
114                    model.layers[il].ffn_norm,
115                    model.layers[il].ffn_norm_b,
116                    LLM_NORM, il);
117            cb(cur, "ffn_norm", il);
118
119            cur = build_ffn(cur,
120                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
121                    NULL,                      NULL,                        NULL,
122                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
123                    NULL,
124                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
125            cb(cur, "ffn_out", il);
126        }
127
128        cur = ggml_add(ctx0, cur, ffn_inp);
129
130        cur = build_cvec(cur, il);
131        cb(cur, "l_out", il);
132
133        // input for next layer
134        inpL = cur;
135    }
136
137    cur = build_norm(inpL,
138            model.output_norm,
139            model.output_norm_b,
140            LLM_NORM, -1);
141
142    cb(cur, "result_norm", -1);
143    res->t_embd = cur;
144
145    cur = build_lora_mm(model.output, cur, model.output_s);
146
147    cb(cur, "result_output", -1);
148    res->t_logits = cur;
149
150    ggml_build_forward_expand(gf, cur);
151}