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 KiB145 linesraw
1#include "models.h"
2
3void llama_model_starcoder::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 36: type = LLM_TYPE_3B; break;
9        case 42: type = LLM_TYPE_7B; break;
10        case 40: type = LLM_TYPE_15B; break;
11        default: type = LLM_TYPE_UNKNOWN;
12    }
13}
14
15void llama_model_starcoder::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    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train}, 0);
20
21    // output
22    {
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}, TENSOR_NOT_REQUIRED);
26        if (!output) {
27            // needs to be on GPU
28            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
29        }
30
31    }
32
33    for (int i = 0; i < n_layer; ++i) {
34        auto & layer = layers[i];
35
36        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
37        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
38
39        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
40        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
41
42        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
43        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
44
45        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
46        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
47
48        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
49        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
50
51        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i),   {n_embd, n_ff}, 0);
52        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff}, 0);
53    }
54}
55
56std::unique_ptr<llm_graph_context> llama_model_starcoder::build_arch_graph(const llm_graph_params & params) const {
57    return std::make_unique<graph>(*this, params);
58}
59
60llama_model_starcoder::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
61    const int64_t n_embd_head = hparams.n_embd_head_v();
62
63    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
64
65    ggml_tensor * cur;
66    ggml_tensor * inpL;
67
68    inpL = build_inp_embd(model.tok_embd);
69
70    // inp_pos - contains the positions
71    ggml_tensor * inp_pos = build_inp_pos();
72
73    auto * inp_attn = build_attn_inp_kv();
74
75    ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
76    cb(pos, "pos_embd", -1);
77
78    inpL = ggml_add(ctx0, inpL, pos);
79    cb(inpL, "inpL", -1);
80
81    ggml_tensor * inp_out_ids = build_inp_out_ids();
82
83    for (int il = 0; il < n_layer; ++il) {
84        cur = build_norm(inpL,
85                model.layers[il].attn_norm,
86                model.layers[il].attn_norm_b,
87                LLM_NORM, il);
88        cb(cur, "attn_norm", il);
89
90        // self-attention
91        {
92            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
93                    n_embd_head, n_head, n_head_kv, il);
94
95            cur = build_attn(inp_attn,
96                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
97                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
98        }
99        if (il == n_layer - 1 && inp_out_ids) {
100            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
101            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
102        }
103        // add the input
104        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
105        cb(ffn_inp, "ffn_inp", il);
106
107        // FF
108        {
109            cur = build_norm(ffn_inp,
110                    model.layers[il].ffn_norm,
111                    model.layers[il].ffn_norm_b,
112                    LLM_NORM, il);
113            cb(cur, "ffn_norm", il);
114
115            cur = build_ffn(cur,
116                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
117                    NULL,                      NULL,                        NULL,
118                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
119                    NULL,
120                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
121            cb(cur, "ffn_out", il);
122        }
123        cur = ggml_add(ctx0, cur, ffn_inp);
124
125        cur = build_cvec(cur, il);
126        cb(cur, "l_out", il);
127
128        // input for next layer
129        inpL = cur;
130    }
131    cur = build_norm(inpL,
132            model.output_norm,
133            model.output_norm_b,
134            LLM_NORM, -1);
135
136    cb(cur, "result_norm", -1);
137    res->t_embd = cur;
138
139    cur = build_lora_mm(model.output, cur, model.output_s);
140
141    cb(cur, "result_output", -1);
142    res->t_logits = cur;
143
144    ggml_build_forward_expand(gf, cur);
145}