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 KiB154 linesraw
1#include "models.h"
2
3void llama_model_dbrx::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5    ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,     hparams.f_clamp_kqv);
6
7    switch (hparams.n_layer()) {
8        case 40: type = LLM_TYPE_16x12B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_dbrx::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    if (n_expert == 0) {
17        throw std::runtime_error("DBRX model cannot have zero experts");
18    }
19
20    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
21
22    // output
23    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
24    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
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        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
32        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
33
34        layer.attn_out_norm = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
35
36        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
37        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
38        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
39        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
40    }
41}
42
43std::unique_ptr<llm_graph_context> llama_model_dbrx::build_arch_graph(const llm_graph_params & params) const {
44    return std::make_unique<graph>(*this, params);
45}
46
47llama_model_dbrx::graph::graph(const llama_model & model, const llm_graph_params & params) : 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    GGML_ASSERT(n_embd_head == n_rot);
52
53    ggml_tensor * cur;
54    ggml_tensor * inpL;
55
56    inpL = build_inp_embd(model.tok_embd);
57
58    // inp_pos - contains the positions
59    ggml_tensor * inp_pos = build_inp_pos();
60
61    auto * inp_attn = build_attn_inp_kv();
62
63    ggml_tensor * inp_out_ids = build_inp_out_ids();
64
65    for (int il = 0; il < n_layer; ++il) {
66        ggml_tensor * inpSA = inpL;
67
68        // norm
69        cur = build_norm(inpL,
70                model.layers[il].attn_norm, NULL,
71                LLM_NORM, il);
72        cb(cur, "attn_norm", il);
73
74        // self-attention
75        {
76            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
77                    n_embd_head, n_head, n_head_kv, il);
78
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
85            Kcur = ggml_rope_ext(
86                    ctx0, Kcur, inp_pos, nullptr,
87                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
88                    ext_factor, attn_factor, beta_fast, beta_slow
89                    );
90
91            cb(Qcur, "Qcur", il);
92            cb(Kcur, "Kcur", il);
93            cb(Vcur, "Vcur", il);
94
95            cur = build_attn(inp_attn,
96                    model.layers[il].wo, NULL, model.layers[il].wo_s,
97                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
98        }
99
100        if (il == n_layer - 1 && inp_out_ids) {
101            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
102            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
103        }
104
105        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
106        cb(ffn_inp, "ffn_inp", il);
107
108        // feed-forward network
109        // MoE branch
110        cur = build_norm(ffn_inp,
111                model.layers[il].attn_out_norm, NULL,
112                LLM_NORM, il);
113        cb(cur, "attn_out_norm", il);
114
115        cur = build_moe_ffn(cur,
116                model.layers[il].ffn_gate_inp,
117                model.layers[il].ffn_up_exps,
118                model.layers[il].ffn_gate_exps,
119                model.layers[il].ffn_down_exps,
120                nullptr,
121                n_expert, n_expert_used,
122                LLM_FFN_SILU, true,
123                hparams.expert_weights_scale,
124                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
125                il);
126        cb(cur, "ffn_moe_out", il);
127
128        cur = ggml_add(ctx0, cur, ffn_inp);
129        cb(cur, "ffn_out", il);
130
131        cur = build_cvec(cur, il);
132        cb(cur, "l_out", il);
133
134        // input for next layer
135        inpL = cur;
136    }
137
138    cur = inpL;
139
140    cur = build_norm(cur,
141            model.output_norm, NULL,
142            LLM_NORM, -1);
143
144    cb(cur, "result_norm", -1);
145    res->t_embd = cur;
146
147    // lm_head
148    cur = build_lora_mm(model.output, cur, model.output_s);
149
150    cb(cur, "result_output", -1);
151    res->t_logits = cur;
152
153    ggml_build_forward_expand(gf, cur);
154}