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 →
6.2 KiB180 linesraw
1#include "models.h"
2
3void llama_model_arctic::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    if (hparams.n_expert == 128) {
7        switch (hparams.n_layer()) {
8            case 35: type = LLM_TYPE_10B_128x3_66B; break;
9            default: type = LLM_TYPE_UNKNOWN;
10        }
11    } else {
12        type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_arctic::load_arch_tensors(llama_model_loader &) {
17    LLAMA_LOAD_LOCALS;
18
19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
20
21    // output
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}, TENSOR_NOT_REQUIRED);
24
25    // if output is NULL, init from the input tok embed
26    if (output == NULL) {
27        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
28    }
29
30    for (int i = 0; i < n_layer; ++i) {
31        auto & layer = layers[i];
32
33        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
34
35        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
36        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
37
38        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
39
40        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd}, 0);
41        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd}, 0);
42        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd}, 0);
43
44        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
45        layer.ffn_norm_exps = create_tensor(tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd}, 0);
46        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
47        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
48        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
49    }
50}
51
52std::unique_ptr<llm_graph_context> llama_model_arctic::build_arch_graph(const llm_graph_params & params) const {
53    return std::make_unique<graph>(*this, params);
54}
55
56llama_model_arctic::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
57    const int64_t n_embd_head = hparams.n_embd_head_v();
58
59    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
60    GGML_ASSERT(n_embd_head == n_rot);
61
62    ggml_tensor * cur;
63    ggml_tensor * inpL;
64
65    inpL = build_inp_embd(model.tok_embd);
66
67    // inp_pos - contains the positions
68    ggml_tensor * inp_pos = build_inp_pos();
69
70    auto * inp_attn = build_attn_inp_kv();
71
72    ggml_tensor * inp_out_ids = build_inp_out_ids();
73
74    for (int il = 0; il < n_layer; ++il) {
75        ggml_tensor * inpSA = inpL;
76
77        // norm
78        cur = build_norm(inpL,
79                model.layers[il].attn_norm, NULL,
80                LLM_NORM_RMS, il);
81        cb(cur, "attn_norm", il);
82
83        // self-attention
84        {
85            // compute Q and K and RoPE them
86            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
87                    n_embd_head, n_head, n_head_kv, il);
88
89            Qcur = ggml_rope_ext(
90                    ctx0, Qcur, 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            Kcur = ggml_rope_ext(
96                    ctx0, Kcur, inp_pos, nullptr,
97                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
98                    ext_factor, attn_factor, beta_fast, beta_slow
99                    );
100
101            cb(Qcur, "Qcur", il);
102            cb(Kcur, "Kcur", il);
103            cb(Vcur, "Vcur", il);
104
105            cur = build_attn(inp_attn,
106                    model.layers[il].wo, NULL, model.layers[il].wo_s,
107                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
108        }
109
110        if (il == n_layer - 1 && inp_out_ids) {
111            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
112            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
113        }
114
115        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
116        cb(ffn_inp, "ffn_inp", il);
117
118        // feed-forward network
119        cur = build_norm(ffn_inp,
120                model.layers[il].ffn_norm, NULL,
121                LLM_NORM_RMS, il);
122        cb(cur, "ffn_norm", il);
123
124        cur = build_ffn(cur,
125                model.layers[il].ffn_up,   NULL, NULL,
126                model.layers[il].ffn_gate, NULL, NULL,
127                model.layers[il].ffn_down, NULL, NULL,
128                NULL,
129                LLM_FFN_SILU, LLM_FFN_PAR, il);
130        cb(cur, "ffn_out", il);
131
132        ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
133        cb(ffn_out, "ffn_out", il);
134
135        // MoE
136        cur = build_norm(inpSA,
137                model.layers[il].ffn_norm_exps, NULL,
138                LLM_NORM_RMS, il);
139        cb(cur, "ffn_norm_exps", il);
140
141        cur = build_moe_ffn(cur,
142                model.layers[il].ffn_gate_inp,
143                model.layers[il].ffn_up_exps,
144                model.layers[il].ffn_gate_exps,
145                model.layers[il].ffn_down_exps,
146                nullptr,
147                n_expert, n_expert_used,
148                LLM_FFN_SILU, true,
149                hparams.expert_weights_scale,
150                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
151                il);
152        cb(cur, "ffn_moe_out", il);
153
154        cur = ggml_add(ctx0, cur, ffn_out);
155        cb(cur, "ffn_out", il);
156
157        cur = build_cvec(cur, il);
158        cb(cur, "l_out", il);
159
160        // input for next layer
161        inpL = cur;
162    }
163
164    cur = inpL;
165
166    cur = build_norm(cur,
167            model.output_norm, NULL,
168            LLM_NORM_RMS, -1);
169
170    cb(cur, "result_norm", -1);
171    res->t_embd = cur;
172
173    // lm_head
174    cur = build_lora_mm(model.output, cur, model.output_s);
175
176    cb(cur, "result_output", -1);
177    res->t_logits = cur;
178
179    ggml_build_forward_expand(gf, cur);
180}