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 →
9.5 KiB211 linesraw
1#include "models.h"
2
3void llama_model_bailingmoe2::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_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
8    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
9    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
10    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
11    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
12
13    switch (hparams.n_layer()) {
14        case 20: type = LLM_TYPE_16B_A1B; break;
15        case 32: type = LLM_TYPE_100B_A6B; break;
16        default: type = LLM_TYPE_UNKNOWN;
17    }
18}
19
20void llama_model_bailingmoe2::load_arch_tensors(llama_model_loader &) {
21    LLAMA_LOAD_LOCALS;
22    const int64_t n_expert_shared = hparams.n_expert_shared;
23
24    const int64_t n_ff_exp        = hparams.n_ff_exp();
25
26    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
27
28    // output
29    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
30    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
31
32    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for bailingmoe2");
33    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for bailingmoe2");
34
35    for (int i = 0; i < n_layer_all; ++i) {
36        int flags = 0;
37        if (i >= n_layer) {
38            // skip all tensors in the NextN layers
39            flags |= TENSOR_SKIP;
40        }
41
42        auto & layer = layers[i];
43
44        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
45
46        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, flags);
47        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
48
49        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
50        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
51
52        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
53
54        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
55            const int64_t n_ff_shexp = (hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp) * n_expert_shared;
56
57            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
58            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
59
60            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
61            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
62            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
63
64            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
65            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
66            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);
67        } else { // Dense layers
68            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
69            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
70            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
71        }
72
73        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
74        if (i >= n_layer) {
75            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
76            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
77            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
78            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
79            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
80            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
81            layer.layer_out_norm         = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, flags);
82        }
83    }
84}
85
86std::unique_ptr<llm_graph_context> llama_model_bailingmoe2::build_arch_graph(const llm_graph_params & params) const {
87    return std::make_unique<graph>(*this, params);
88}
89
90llama_model_bailingmoe2::graph::graph(const llama_model & model, const llm_graph_params & params) :
91    llm_graph_context(params) {
92    const int64_t n_embd_head = hparams.n_embd_head_v();
93
94    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
95
96    ggml_tensor * cur;
97    ggml_tensor * inpL;
98
99    inpL = build_inp_embd(model.tok_embd);
100
101    // inp_pos - contains the positions
102    ggml_tensor * inp_pos = build_inp_pos();
103
104    auto * inp_attn = build_attn_inp_kv();
105
106    ggml_tensor * inp_out_ids = build_inp_out_ids();
107
108    for (int il = 0; il < n_layer; ++il) {
109        ggml_tensor * inpSA = inpL;
110
111        // norm
112        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
113        cb(cur, "attn_norm", il);
114
115        // self_attention
116        {
117            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
118                    n_embd_head, n_head, n_head_kv, il);
119
120            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
121            cb(Qcur, "Qcur_normed", il);
122
123            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
124                                 ext_factor, attn_factor, beta_fast, beta_slow);
125
126            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
127            cb(Kcur, "Kcur_normed", il);
128
129            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
130                                 ext_factor, attn_factor, beta_fast, beta_slow);
131
132            cb(Qcur, "Qcur", il);
133            cb(Kcur, "Kcur", il);
134            cb(Vcur, "Vcur", il);
135
136            cur = build_attn(inp_attn,
137                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
138                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
139        }
140
141        if (il == n_layer - 1 && inp_out_ids) {
142            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
143            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
144        }
145
146        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);
147        cb(sa_out, "sa_out", il);
148
149        // MoE branch
150        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
151        cb(cur, "ffn_norm", il);
152
153        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
154            cur = build_ffn(cur,
155                    model.layers[il].ffn_up, NULL, NULL,
156                    model.layers[il].ffn_gate, NULL, NULL,
157                    model.layers[il].ffn_down, NULL, NULL,
158                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
159            cb(cur, "ffn_out", il);
160        } else {
161            ggml_tensor * moe_out = build_moe_ffn(cur,
162                model.layers[il].ffn_gate_inp,
163                model.layers[il].ffn_up_exps,
164                model.layers[il].ffn_gate_exps,
165                model.layers[il].ffn_down_exps,
166                model.layers[il].ffn_exp_probs_b,
167                n_expert, n_expert_used,
168                LLM_FFN_SILU, hparams.expert_weights_norm,
169                hparams.expert_weights_scale,
170                (llama_expert_gating_func_type) hparams.expert_gating_func,
171                il);
172            cb(moe_out, "ffn_moe_out", il);
173
174            {
175                ggml_tensor * ffn_shexp =
176                    build_ffn(cur,
177                        model.layers[il].ffn_up_shexp, NULL, NULL,
178                        model.layers[il].ffn_gate_shexp, NULL, NULL,
179                        model.layers[il].ffn_down_shexp, NULL, NULL,
180                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
181                cb(ffn_shexp, "ffn_shexp", il);
182
183                cur = ggml_add(ctx0, moe_out, ffn_shexp);
184                cb(cur, "ffn_out", il);
185            }
186        }
187
188        cur = ggml_add(ctx0, cur, sa_out);
189
190        cur = build_cvec(cur, il);
191        cb(cur, "l_out", il);
192
193        // input for next layer
194        inpL = cur;
195    }
196
197    cur = inpL;
198
199    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
200
201    cb(cur, "result_norm", -1);
202    res->t_embd = cur;
203
204    // lm_head
205    cur = build_lora_mm(model.output, cur, model.output_s);
206
207    cb(cur, "result_output", -1);
208    res->t_logits = cur;
209
210    ggml_build_forward_expand(gf, cur);
211}