char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

ynankanimodel : NvFP4 quantized LM head support (#23046)42928bc14

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
5.0 KiB133 linesraw
1#include "models.h"
2
3std::unique_ptr<llm_graph_context> llama_model_ernie4_5_moe::build_arch_graph(const llm_graph_params & params) const {
4    return std::make_unique<graph>(*this, params);
5}
6
7llama_model_ernie4_5_moe::graph::graph(const llama_model & model, const llm_graph_params & params) :
8    llm_graph_context(params) {
9    const int64_t n_embd_head = hparams.n_embd_head_v();
10
11    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
12    GGML_ASSERT(n_embd_head == n_rot);
13
14    ggml_tensor * cur;
15    ggml_tensor * inpL;
16
17    inpL = build_inp_embd(model.tok_embd);
18
19    // inp_pos - contains the positions
20    ggml_tensor * inp_pos = build_inp_pos();
21
22    auto * inp_attn = build_attn_inp_kv();
23
24    ggml_tensor * inp_out_ids = build_inp_out_ids();
25
26    GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0");
27    for (int il = 0; il < n_layer; ++il) {
28        ggml_tensor * inpSA = inpL;
29        // norm
30        {
31            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
32            cb(cur, "attn_norm", il);
33        }
34        // self-attention
35        {
36            // compute Q and K and RoPE them
37            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
38                    n_embd_head, n_head, n_head_kv, il);
39
40            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
41                                 ext_factor, attn_factor, beta_fast, beta_slow);
42
43            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
44                                 ext_factor, attn_factor, beta_fast, beta_slow);
45
46            cb(Qcur, "Qcur", il);
47            cb(Kcur, "Kcur", il);
48            cb(Vcur, "Vcur", il);
49
50            cur = build_attn(inp_attn,
51                    model.layers[il].wo, NULL, model.layers[il].wo_s,
52                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
53            cb(cur, "attn_out", il);
54        }
55        if (il == n_layer - 1 && inp_out_ids) {
56            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
57            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
58        }
59        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
60        cb(ffn_inp, "ffn_inp", il);
61
62        // feed-forward network
63        bool is_moe_layer =
64            static_cast<uint32_t>(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0;
65
66        if (!is_moe_layer) {
67            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
68            cb(cur, "ffn_norm", il);
69
70            cur = build_ffn(cur,
71                    model.layers[il].ffn_up, NULL, NULL,
72                    model.layers[il].ffn_gate, NULL, NULL,
73                    model.layers[il].ffn_down, NULL, NULL,
74                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
75            cb(cur, "ffn_out", il);
76        } else {
77            // MoE branch
78            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
79            cb(cur, "ffn_norm", il);
80
81            ggml_tensor * moe_out = build_moe_ffn(cur,
82                                        model.layers[il].ffn_gate_inp,
83                                        model.layers[il].ffn_up_exps,
84                                        model.layers[il].ffn_gate_exps,
85                                        model.layers[il].ffn_down_exps,
86                                        model.layers[il].ffn_exp_probs_b,
87                                        n_expert, n_expert_used,
88                                        LLM_FFN_SILU, true,
89                                        hparams.expert_weights_scale,
90                                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
91                                        il);
92            cb(moe_out, "ffn_moe_out", il);
93
94            // Shared expert (if present)
95            if (hparams.n_ff_shexp > 0) {
96                ggml_tensor * ffn_shexp =
97                    build_ffn(cur,
98                        model.layers[il].ffn_up_shexp, NULL, NULL,
99                        model.layers[il].ffn_gate_shexp, NULL, NULL,
100                        model.layers[il].ffn_down_shexp, NULL, NULL,
101                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
102                cb(ffn_shexp, "ffn_shexp", il);
103
104                cur = ggml_add(ctx0, moe_out, ffn_shexp);
105            } else {
106                cur = moe_out;
107            }
108            cb(cur, "ffn_out", il);
109        }
110        cur = ggml_add(ctx0, cur, ffn_inp);
111        cb(cur, "ffn_out", il);
112
113        cur = build_cvec(cur, il);
114        cb(cur, "l_out", il);
115
116        // input for next layer
117        inpL = cur;
118    }
119    cur = inpL;
120
121    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
122
123    cb(cur, "result_norm", -1);
124    res->t_embd = cur;
125
126    // lm_head
127    cur = build_lora_mm(model.output, cur, model.output_s);
128
129    cb(cur, "result_output", -1);
130    res->t_logits = cur;
131
132    ggml_build_forward_expand(gf, cur);
133}