char-fork/llama.cpp

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

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.8 KiB180 linesraw
1#include "models.h"
2
3void llama_model_bailingmoe::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
6    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
7    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
8    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
9    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
10
11    switch (hparams.n_layer()) {
12        case 28: type = LLM_TYPE_16B; break;
13        case 88: type = LLM_TYPE_290B; break;
14        default: type = LLM_TYPE_UNKNOWN;
15    }
16}
17
18void llama_model_bailingmoe::load_arch_tensors(llama_model_loader &) {
19    LLAMA_LOAD_LOCALS;
20    const int64_t n_expert_shared = hparams.n_expert_shared;
21
22    const int64_t n_ff_exp            = hparams.n_ff_exp();
23
24    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
25
26    // output
27    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
28    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
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_head * n_rot, n_head_kv * n_rot, n_head_kv * n_rot, 0);
36        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);
37        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
38
39        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
40
41        if (n_expert == 0) {
42            throw std::runtime_error("n_expert must be > 0");
43        }
44        if (n_expert_used == 0) {
45            throw std::runtime_error("n_expert_used must be > 0");
46        }
47
48        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
49        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
50        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
51
52        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
53        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
54        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
55    }
56}
57
58std::unique_ptr<llm_graph_context> llama_model_bailingmoe::build_arch_graph(const llm_graph_params & params) const {
59    return std::make_unique<graph>(*this, params);
60}
61
62llama_model_bailingmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
63    ggml_tensor * cur;
64    ggml_tensor * inpL;
65
66    inpL = build_inp_embd(model.tok_embd);
67
68    // inp_pos - contains the positions
69    ggml_tensor * inp_pos = build_inp_pos();
70
71    auto * inp_attn = build_attn_inp_kv();
72
73    ggml_tensor * inp_out_ids = build_inp_out_ids();
74
75    for (int il = 0; il < n_layer; ++il) {
76        ggml_tensor * inpSA = inpL;
77
78        // norm
79        cur = build_norm(inpL,
80                model.layers[il].attn_norm, NULL,
81                LLM_NORM_RMS, il);
82        cb(cur, "attn_norm", il);
83
84        // self-attention
85        {
86            // rope freq factors for llama3; may return nullptr for llama2 and other models
87            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
88
89            // compute Q and K and RoPE them
90            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
91                    n_embd_head_k, n_head, n_head_kv, il);
92
93            Qcur = ggml_rope_ext(
94                    ctx0, Qcur, inp_pos, rope_factors,
95                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
96                    ext_factor, attn_factor, beta_fast, beta_slow
97                    );
98
99            Kcur = ggml_rope_ext(
100                    ctx0, Kcur, inp_pos, rope_factors,
101                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
102                    ext_factor, attn_factor, beta_fast, beta_slow
103                    );
104
105            cb(Qcur, "Qcur", il);
106            cb(Kcur, "Kcur", il);
107            cb(Vcur, "Vcur", il);
108
109            cur = build_attn(inp_attn,
110                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
111                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
112        }
113
114        if (il == n_layer - 1 && inp_out_ids) {
115            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
116            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
117        }
118
119        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
120        cb(ffn_inp, "ffn_inp", il);
121
122        cur = build_norm(ffn_inp,
123                model.layers[il].ffn_norm, NULL,
124                LLM_NORM_RMS, il);
125        cb(cur, "ffn_norm", il);
126
127        ggml_tensor * moe_out =
128            build_moe_ffn(cur,
129                    model.layers[il].ffn_gate_inp,
130                    model.layers[il].ffn_up_exps,
131                    model.layers[il].ffn_gate_exps,
132                    model.layers[il].ffn_down_exps,
133                    nullptr,
134                    n_expert, n_expert_used,
135                    LLM_FFN_SILU, hparams.expert_weights_norm,
136                    hparams.expert_weights_scale,
137                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
138                    il);
139        cb(moe_out, "ffn_moe_out", il);
140
141        // FFN shared expert
142        {
143            ggml_tensor * ffn_shexp = build_ffn(cur,
144                    model.layers[il].ffn_up_shexp,   NULL, NULL,
145                    model.layers[il].ffn_gate_shexp, NULL, NULL,
146                    model.layers[il].ffn_down_shexp, NULL, NULL,
147                    NULL,
148                    LLM_FFN_SILU, LLM_FFN_PAR, il);
149            cb(ffn_shexp, "ffn_shexp", il);
150
151            cur = ggml_add(ctx0, moe_out, ffn_shexp);
152            cb(cur, "ffn_out", il);
153        }
154
155        cur = ggml_add(ctx0, cur, ffn_inp);
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}