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 →
10.8 KiB241 linesraw
1#include "models.h"
2
3void llama_model_exaone_moe::load_arch_hparams(llama_model_loader & ml) {
4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
5    hparams.n_swa = 128;
6    uint32_t swa_period = 4;
7    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
8    hparams.set_swa_pattern(swa_period);
9    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
10    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
11
12    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,                hparams.rope_freq_base_train_swa, false);
13    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
14    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
15    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
16    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
17    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
18    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
19    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
20    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
21    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
22
23    switch (hparams.n_layer()) {
24        case 32: type = LLM_TYPE_30B_A3B; break;
25        case 48: type = LLM_TYPE_235B_A22B; break;
26        default: type = LLM_TYPE_UNKNOWN;
27    }
28}
29
30void llama_model_exaone_moe::load_arch_tensors(llama_model_loader &) {
31    LLAMA_LOAD_LOCALS;
32
33    const int64_t n_ff_exp       = hparams.n_ff_exp();
34    const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;
35    const int64_t head_dim       = hparams.n_embd_head_k();
36    const int64_t n_qo_dim       = n_head * head_dim;
37    const int64_t n_kv_dim       = n_head_kv * head_dim;
38
39    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
40
41    // output
42    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
43    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
44
45    if (output == NULL) {
46        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
47    }
48
49    for (int i = 0; i < n_layer_all; ++i) {
50        int flags = 0;
51        if (i >= n_layer) {
52            // skip all tensors in the NextN layers
53            flags |= TENSOR_SKIP;
54        }
55
56        auto & layer = layers[i];
57        create_tensor_qkv(layer, i, n_embd, n_qo_dim, n_kv_dim, n_kv_dim, flags);
58        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, flags);
59
60        layer.rope_freqs   = create_tensor(tn(LLM_TENSOR_ROPE_FREQS,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0) | flags);
61
62        layer.attn_norm    = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, flags);
63        layer.attn_q_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
64        layer.attn_k_norm  = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
65
66        layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd}, flags);
67
68        // dense layers for first n_layer_dense_lead layers or nextn_predict_layers layers at the end
69        if (i < (int) hparams.n_layer_dense_lead || (i >= n_layer)) {
70            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
71            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, flags);
72            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, flags);
73        } else {
74            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, flags);
75            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
76
77            if (n_expert == 0) {
78                throw std::runtime_error("n_expert must be > 0");
79            }
80            if (n_expert_used == 0) {
81                throw std::runtime_error("n_expert_used must be > 0");
82            }
83
84            layer.ffn_gate_exps  = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS,  "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
85            layer.ffn_down_exps  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,  "weight", i), {n_ff_exp, n_embd, n_expert}, flags);
86            layer.ffn_up_exps    = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,    "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
87
88            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
89            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
90            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);
91        }
92
93        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
94        if (i >= n_layer) {
95            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
96            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);
97            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);
98
99            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
100            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
101            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
102        }
103    }
104}
105
106std::unique_ptr<llm_graph_context> llama_model_exaone_moe::build_arch_graph(const llm_graph_params & params) const {
107    return std::make_unique<graph>(*this, params);
108}
109
110llama_model_exaone_moe::graph::graph(const llama_model & model, const llm_graph_params & params) :
111    llm_graph_context(params) {
112    const int64_t n_embd_head = hparams.n_embd_head_k();
113
114    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
115    GGML_ASSERT(n_embd_head == n_rot);
116
117    ggml_tensor * cur;
118    ggml_tensor * inpL;
119
120    inpL = build_inp_embd(model.tok_embd);
121
122    // inp_pos - contains the positions
123    ggml_tensor * inp_pos = build_inp_pos();
124
125    auto * inp_attn_iswa = build_attn_inp_kv_iswa();
126
127    ggml_tensor * inp_out_ids = build_inp_out_ids();
128
129    for (int il = 0; il < n_layer; ++il) {
130        ggml_tensor * inpSA = inpL;
131
132        // use RoPE for SWA layers
133        const bool is_local_layer = hparams.is_swa(il);
134
135        // norm
136        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
137        cb(cur, "attn_norm", il);
138
139        // self-attention
140        {
141            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
142
143            // compute Q and K and RoPE them
144            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
145                    n_embd_head, n_head, n_head_kv, il);
146
147            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
148            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
149            cb(Qcur, "Qcur_normed", il);
150            cb(Kcur, "Kcur_normed", il);
151
152            if (is_local_layer) {
153                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
154                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
155
156                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
157                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
158            }
159            cb(Qcur, "Qcur", il);
160            cb(Kcur, "Kcur", il);
161            cb(Vcur, "Vcur", il);
162
163            cur = build_attn(inp_attn_iswa,
164                model.layers[il].wo, NULL, model.layers[il].wo_s,
165                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
166            cb(cur, "attn_out", il);
167        }
168        if (il == n_layer - 1 && inp_out_ids) {
169            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
170            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
171        }
172        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
173        cb(ffn_inp, "ffn_inp", il);
174
175        // norm
176        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
177        cb(cur, "ffn_norm", il);
178
179        // feed-forward network
180        if (model.layers[il].ffn_gate_inp == nullptr) {
181            // dense branch
182            cur = build_ffn(cur,
183                    model.layers[il].ffn_up, NULL, NULL,
184                    model.layers[il].ffn_gate, NULL, NULL,
185                    model.layers[il].ffn_down, NULL, NULL, NULL,
186                    LLM_FFN_SILU, LLM_FFN_PAR, il);
187            cb(cur, "ffn_out", il);
188        } else {
189            // MoE branch
190            ggml_tensor * moe_out = build_moe_ffn(cur,
191                model.layers[il].ffn_gate_inp,
192                model.layers[il].ffn_up_exps,
193                model.layers[il].ffn_gate_exps,
194                model.layers[il].ffn_down_exps,
195                model.layers[il].ffn_exp_probs_b,
196                n_expert, n_expert_used,
197                LLM_FFN_SILU, hparams.expert_weights_norm,
198                hparams.expert_weights_scale,
199                (llama_expert_gating_func_type) hparams.expert_gating_func,
200                il);
201            cb(moe_out, "ffn_moe_out", il);
202
203            // FFN shared expert
204            {
205                ggml_tensor * ffn_shexp =
206                    build_ffn(cur,
207                        model.layers[il].ffn_up_shexp, NULL, NULL,
208                        model.layers[il].ffn_gate_shexp, NULL, NULL,
209                        model.layers[il].ffn_down_shexp, NULL, NULL,
210                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
211                cb(ffn_shexp, "ffn_shexp", il);
212
213                cur = ggml_add(ctx0, moe_out, ffn_shexp);
214                cb(cur, "ffn_out", il);
215            }
216        }
217
218        cur = ggml_add(ctx0, cur, ffn_inp);
219
220        cur = build_cvec(cur, il);
221        cb(cur, "l_out", il);
222
223        // input for next layer
224        inpL = cur;
225    }
226    cur = inpL;
227
228    // final norm
229    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
230
231    cb(cur, "result_norm", -1);
232    res->t_embd = cur;
233
234    // lm_head
235    cur = build_lora_mm(model.output, cur, model.output_s);
236
237    cb(cur, "result_output", -1);
238    res->t_logits = cur;
239
240    ggml_build_forward_expand(gf, cur);
241}