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 →
11.9 KiB285 linesraw
1#include "models.h"
2
3void llama_model_afmoe::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_GATING_FUNC,          hparams.expert_gating_func, false);
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_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
12
13    // Set up interleaved sliding window attention (ISWA)
14    // Pattern: 3 sliding - 1 full (global_attn_every_n_layers = 4)
15    if (hparams.n_swa > 0) {
16        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
17        uint32_t swa_period = 4;
18        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
19        hparams.set_swa_pattern(swa_period);
20
21        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
22        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
23        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
24    } else {
25        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
26    }
27
28    // Default to sigmoid if not set
29    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
30        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
31    }
32
33    switch (hparams.n_layer()) {
34        case 56: type = LLM_TYPE_6B; break;
35        case 32: type = LLM_TYPE_26B; break;
36        default: type = LLM_TYPE_UNKNOWN;
37    }
38}
39
40void llama_model_afmoe::load_arch_tensors(llama_model_loader &) {
41    LLAMA_LOAD_LOCALS;
42    const int64_t n_expert_shared = hparams.n_expert_shared;
43
44    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
45
46    // output
47    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
48    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
49
50    // if output is NULL, init from the input tok embed
51    if (output == NULL) {
52        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
53    }
54
55    const int64_t n_ff_exp = hparams.n_ff_exp();
56
57    for (int i = 0; i < n_layer; ++i) {
58        auto & layer = layers[i];
59
60        // dual attention normalization
61        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
62        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
63
64        // attention projections
65        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
66        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
67
68        // Q/K normalization
69        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
70        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
71
72        // attention gating
73        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
74
75        // dual ffn normalization
76        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);
77        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
78
79        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) {
80            // MoE layers
81            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
82            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
83
84            // grouped expert weights
85            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
86            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
87            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
88
89            // shared expert
90            if (n_expert_shared > 0) {
91                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
92                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0);
93                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
94                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
95            }
96        } else {
97            // Dense layers
98            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
99            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
100            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
101        }
102    }
103}
104
105std::unique_ptr<llm_graph_context> llama_model_afmoe::build_arch_graph(const llm_graph_params & params) const {
106    return std::make_unique<graph>(*this, params);
107}
108
109llama_model_afmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
110    const int64_t n_embd_head = hparams.n_embd_head_v();
111    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
112
113    ggml_tensor * cur;
114    ggml_tensor * inpL;
115
116    inpL = build_inp_embd(model.tok_embd);
117
118    // MuP scaling: embeddings * sqrt(hidden_size)
119    // mup_enabled = true, hidden_size = 1024, scale = 32.0
120    inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
121    cb(inpL, "inp_embd_scaled", -1);
122
123    // inp_pos - contains the positions
124    ggml_tensor * inp_pos = build_inp_pos();
125    auto * inp_attn = build_attn_inp_kv_iswa();
126    ggml_tensor * inp_out_ids = build_inp_out_ids();
127
128    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
129
130    for (int il = 0; il < n_layer; ++il) {
131        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
132        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
133
134        ggml_tensor * inpSA = inpL;
135
136        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
137        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
138                              (il + 1) % hparams.n_no_rope_layer_step != 0;
139
140        // dual attention normalization (pre)
141        cur = build_norm(inpL,
142                model.layers[il].attn_norm, NULL,
143                LLM_NORM_RMS, il);
144        cb(cur, "attn_norm", il);
145
146        // self-attention
147        {
148            ggml_tensor * attn_inp = cur;  // save input for gate computation
149
150            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
151                    n_embd_head, n_head, n_head_kv, il);
152
153            // compute gate from input
154            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
155            cb(gate, "attn_gate_proj", il);
156
157            // Q/K normalization
158            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
159            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
160            cb(Qcur, "Qcur_normed", il);
161            cb(Kcur, "Kcur_normed", il);
162
163            if (use_rope) {
164                Qcur = ggml_rope_ext(
165                        ctx0, Qcur, inp_pos, nullptr,
166                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
167                        ext_factor, attn_factor, beta_fast, beta_slow);
168                cb(Qcur, "Qcur_rope", il);
169
170                Kcur = ggml_rope_ext(
171                        ctx0, Kcur, inp_pos, nullptr,
172                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
173                        ext_factor, attn_factor, beta_fast, beta_slow);
174                cb(Kcur, "Kcur_rope", il);
175            }
176
177            cur = build_attn(inp_attn,
178                    NULL, NULL, NULL,  // wo will be applied after gating
179                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
180            cb(cur, "attn_out", il);
181
182            // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
183            gate = ggml_sigmoid(ctx0, gate);
184            cb(gate, "attn_gate_sig", il);
185            cur = ggml_mul(ctx0, cur, gate);
186            cb(cur, "attn_gated", il);
187
188            // now apply output projection
189            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
190            cb(cur, "attn_o_proj", il);
191        }
192
193        // dual attention normalization (post)
194        cur = build_norm(cur,
195                model.layers[il].attn_post_norm, NULL,
196                LLM_NORM_RMS, il);
197        cb(cur, "attn_post_norm", il);
198
199        if (il == n_layer - 1 && inp_out_ids) {
200            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
201            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
202        }
203
204        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
205        cb(ffn_inp, "ffn_inp", il);
206
207        // dual ffn normalization (pre)
208        cur = build_norm(ffn_inp,
209                model.layers[il].ffn_norm, NULL,
210                LLM_NORM_RMS, il);
211        cb(cur, "ffn_norm", il);
212
213        // MoE or dense FFN
214        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
215            // MoE layer with sigmoid routing, normalization, and scaling
216            ggml_tensor * moe_out = build_moe_ffn(cur,
217                    model.layers[il].ffn_gate_inp,
218                    model.layers[il].ffn_up_exps,
219                    model.layers[il].ffn_gate_exps,
220                    model.layers[il].ffn_down_exps,
221                    model.layers[il].ffn_exp_probs_b,
222                    n_expert, n_expert_used,
223                    LLM_FFN_SILU,
224                    hparams.expert_weights_norm,           // norm_w (route_norm=True)
225                    hparams.expert_weights_scale,          // w_scale (route_scale=2.826)
226                    (llama_expert_gating_func_type) hparams.expert_gating_func,
227                    il);
228            cb(moe_out, "ffn_moe_out", il);
229
230            // shared expert
231            if (hparams.n_expert_shared > 0) {
232                ggml_tensor * ffn_shexp = build_ffn(cur,
233                        model.layers[il].ffn_up_shexp,   NULL, NULL,
234                        model.layers[il].ffn_gate_shexp, NULL, NULL,
235                        model.layers[il].ffn_down_shexp, NULL, NULL,
236                        NULL,
237                        LLM_FFN_SILU, LLM_FFN_PAR, il);
238                cb(ffn_shexp, "ffn_shexp", il);
239
240                cur = ggml_add(ctx0, moe_out, ffn_shexp);
241                cb(cur, "ffn_out", il);
242            } else {
243                cur = moe_out;
244            }
245        } else {
246            // dense layer
247            cur = build_ffn(cur,
248                    model.layers[il].ffn_up,   NULL, NULL,
249                    model.layers[il].ffn_gate, NULL, NULL,
250                    model.layers[il].ffn_down, NULL, NULL,
251                    NULL,
252                    LLM_FFN_SILU, LLM_FFN_PAR, il);
253            cb(cur, "ffn_out", il);
254        }
255
256        // dual ffn normalization (post)
257        cur = build_norm(cur,
258                model.layers[il].ffn_post_norm, NULL,
259                LLM_NORM_RMS, il);
260        cb(cur, "ffn_post_norm", il);
261
262        cur = ggml_add(ctx0, cur, ffn_inp);
263        cur = build_cvec(cur, il);
264        cb(cur, "l_out", il);
265
266        // input for next layer
267        inpL = cur;
268    }
269
270    cur = inpL;
271
272    cur = build_norm(cur,
273            model.output_norm, NULL,
274            LLM_NORM_RMS, -1);
275    cb(cur, "result_norm", -1);
276
277    res->t_embd = cur;
278
279    // lm_head
280    cur = build_lora_mm(model.output, cur, model.output_s);
281    cb(cur, "result_output", -1);
282    res->t_logits = cur;
283
284    ggml_build_forward_expand(gf, cur);
285}