char-fork/llama.cpp

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

Sigbjørn Skjæretmodels : fix incorrect uses of get_key_or_arr (#28868)97e4ca735

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
22.9 KiB498 linesraw
1#include "models.h"
2
3void llama_model_gemma4::load_arch_hparams(llama_model_loader & ml) {
4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
5    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
6
7    uint32_t n_kv_shared_layers = 0;
8    ml.get_key(LLM_KV_ATTENTION_SHARED_KV_LAYERS, n_kv_shared_layers, false);
9
10    hparams.n_layer_kv_from_start = hparams.n_layer_all - (int32_t)n_kv_shared_layers;
11    hparams.f_attention_scale     = 1.0f; // Gemma4 uses self.scaling = 1.0 (no pre-attn scaling)
12
13    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
14    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
15    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
16    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
17    ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER,  hparams.n_embd_per_layer);
18    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_SWA,    hparams.n_embd_head_k_swa);
19    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,  hparams.n_embd_head_v_swa);
20    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
21
22    // when non_causal is set, the model will use bidirectional attention on SWA layers only, while dense layers will remain causal
23    // ref: use_bidirectional_attention == "vision" in HF config
24    // note: E2B/E4B are always causal, bypassing this logic
25    hparams.non_causal_type = LLAMA_NON_CAUSAL_TYPE_SWA_ONLY;
26
27    switch (hparams.n_layer()) {
28        case 30: type = LLM_TYPE_26B_A4B; break;
29        case 35: type = LLM_TYPE_E2B; break;
30        case 42: type = LLM_TYPE_E4B; break;
31        case 60: type = LLM_TYPE_31B; break;
32        default: type = LLM_TYPE_UNKNOWN;
33    }
34}
35
36void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {
37    LLAMA_LOAD_LOCALS;
38
39    const uint32_t n_embd_per_layer = hparams.n_embd_per_layer;
40    const int64_t  n_ff_exp         = hparams.n_ff_exp();
41
42    if (n_embd_head_k != n_embd_head_v) {
43        throw std::runtime_error("Gemma 4 requires n_embd_head_k == n_embd_head_v");
44    }
45    if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {
46        throw std::runtime_error("Gemma 4 requires n_embd_head_k_swa == n_embd_head_v_swa");
47    }
48
49    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
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    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
56
57    if (n_embd_per_layer > 0) {
58        per_layer_tok_embd   = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"),    {n_embd_per_layer * n_layer, n_vocab}, TENSOR_READ_LAZY);
59        per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0);
60        per_layer_proj_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight", 0), {n_embd_per_layer}, 0);
61    }
62
63    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
64
65    int rope_freqs_flag = 0;
66
67    for (int i = 0; i < n_layer; ++i) {
68        auto & layer = layers[i];
69        const int64_t n_head      = hparams.n_head(i);
70        const int64_t n_embd_head = hparams.n_embd_head_k(i);
71        const int64_t n_embd_k    = hparams.n_embd_k_gqa(i);
72        const int64_t n_embd_v    = hparams.n_embd_v_gqa(i);
73        const int     kv_flags    = hparams.has_kv(i) ? 0 : TENSOR_NOT_REQUIRED;
74
75        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
76
77        // note: use_alternative_attention (v_proj is optional, if it's not present, use k_proj)
78        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i),
79            {n_embd, n_embd_head * n_head + n_embd_k + n_embd_v}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
80        if (!layer.wqkv) {
81            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head * n_head}, 0);
82            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k}, kv_flags);
83            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);
84        }
85        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);
86
87        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head}, 0);
88        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head}, kv_flags);
89        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
90
91        layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, TENSOR_NOT_REQUIRED);
92
93        if (!hparams.is_swa(i)) {
94            // full_attention layers use rope_freqs for proportional rope
95            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_embd_head/2}, rope_freqs_flag);
96            rope_freqs_flag = TENSOR_DUPLICATED;
97        }
98
99        // handle use_double_wide_mlp
100        int64_t n_ff_cur = hparams.n_ff(i);
101
102        // for expert layers, we use normal FFN as shared expert (same as python code)
103        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
104        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff_cur}, 0);
105        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff_cur}, 0);
106        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);
107        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
108
109        // MoE router
110        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
111        bool has_expert = layer.ffn_gate_inp != nullptr;
112
113        // norm
114        if (has_expert) {
115            layer.ffn_gate_inp_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "scale", i), {n_embd}, 0);
116
117            layer.ffn_pre_norm_2  = create_tensor(tn(LLM_TENSOR_FFN_PRE_NORM_2,  "weight", i), {n_embd}, 0);
118            layer.ffn_post_norm_1 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_1, "weight", i), {n_embd}, 0);
119            layer.ffn_post_norm_2 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_2, "weight", i), {n_embd}, 0);
120
121            // MoE FFN
122            layer.ffn_gate_up_exps  = create_tensor(tn(LLM_TENSOR_FFN_GATE_UP_EXPS,  "weight", i), {n_embd, n_ff_exp * 2, n_expert}, TENSOR_NOT_REQUIRED);
123
124            if (layer.ffn_gate_up_exps == nullptr) {
125                layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
126                layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
127            }
128
129            layer.ffn_down_exps     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,     "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
130
131            // per-expert scale will be loaded as down_exps_s at the end of the current switch case
132        }
133
134        // per-layer embeddings
135        if (n_embd_per_layer > 0) {
136            layer.per_layer_inp_gate   = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE,  "weight", i), {n_embd, n_embd_per_layer}, 0);
137            layer.per_layer_proj       = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ,      "weight", i), {n_embd_per_layer, n_embd}, 0);
138            layer.per_layer_post_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);
139        }
140    }
141}
142
143std::unique_ptr<llm_graph_context> llama_model_gemma4::build_arch_graph(const llm_graph_params & params) const {
144    return std::make_unique<graph>(*this, params);
145}
146
147// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
148static ggml_tensor * gemma4_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, int idx) {
149    GGML_ASSERT(idx < (int) x->ne[2]);
150    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),
151                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
152}
153
154llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_params & params) :
155        llm_graph_context(params),
156        model(model),
157        n_embd_per_layer(model.hparams.n_embd_per_layer) {
158    ggml_tensor * cur;
159    ggml_tensor * inpL;
160
161    inpL = build_inp_embd(model.tok_embd);
162
163    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
164    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
165    cb(inpL, "inp_scaled", -1);
166
167    // inp_pos - contains the positions
168    ggml_tensor * inp_pos = build_inp_pos();
169
170    // TODO: is causal == true correct? might need some changes
171    auto * inp_attn = build_attn_inp_kv_iswa();
172
173    ggml_tensor * inp_out_ids = build_inp_out_ids();
174
175    ggml_tensor * inp_per_layer = nullptr;
176    if (model.per_layer_tok_embd) {
177        inp_per_layer = build_inp_per_layer();
178        ggml_build_forward_expand(gf, inp_per_layer);
179
180        // inp_per_layer shape: [n_embd_per_layer, n_tokens, n_layer]
181        inp_per_layer = project_per_layer_inputs(inpL, inp_per_layer);
182    }
183
184    for (int il = 0; il < n_layer; ++il) {
185        const int64_t n_embd_head = hparams.n_embd_head_k(il);
186        GGML_ASSERT(n_embd_head == hparams.n_embd_head_v(il));
187
188        const int64_t n_head    = hparams.n_head(il);
189        const int64_t n_head_kv = hparams.n_head_kv(il);
190
191        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
192        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
193        const int   n_rot_l      = hparams.n_rot(il);
194
195        res->t_layer_inp[il] = inpL;
196
197        // norm
198        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
199        cb(cur, "attn_norm", il);
200
201        ggml_tensor * freq_factors = nullptr;
202        if (!hparams.is_swa(il)) {
203            // full_attention layers use rope_freqs for proportional rope
204            freq_factors = model.layers[il].rope_freqs;
205        }
206
207        // Q projection (shared for both non-KV and KV layers)
208        // this is to mirror Gemma4Attention in pytorch code
209        ggml_tensor * qkv_fused = nullptr;
210        ggml_tensor * Qcur;
211        if (model.layers[il].wqkv) {
212            qkv_fused = build_lora_mm(model.layers[il].wqkv, cur, model.layers[il].wqkv_s);
213            cb(qkv_fused, "wqkv", il);
214            const int64_t q_dim = n_embd_head * n_head;
215            Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, q_dim, n_tokens, qkv_fused->nb[1], 0));
216        } else {
217            Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
218        }
219        {
220            cb(Qcur, "Qcur", il);
221
222            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
223
224            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
225            cb(Qcur, "Qcur_normed", il);
226
227            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
228                                 ext_factor, attn_factor, beta_fast, beta_slow);
229            cb(Qcur, "Qcur_pos", il);
230        }
231
232        // self-attention
233        if (hparams.has_kv(il)) {
234            ggml_tensor * Kcur;
235            ggml_tensor * Vcur;
236            if (qkv_fused) {
237                const int64_t q_dim = n_embd_head * n_head;
238                const int64_t k_dim = n_embd_head * n_head_kv;
239                const int64_t v_dim = n_embd_head * n_head_kv;
240                const size_t  esize = ggml_element_size(qkv_fused);
241                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, k_dim, n_tokens, qkv_fused->nb[1], q_dim * esize));
242                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, v_dim, n_tokens, qkv_fused->nb[1], (q_dim + k_dim) * esize));
243            } else {
244                Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
245                Vcur = model.layers[il].wv
246                       ? build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s)
247                       : Kcur; // if v_proj is not present, use Kcur as Vcur
248            }
249            cb(Kcur, "Kcur", il);
250            cb(Vcur, "Vcur", il);
251
252            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
253            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
254
255            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
256            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
257
258            cb(Kcur, "Kcur_normed", il);
259            cb(Vcur, "Vcur_normed", il);
260
261            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
262                                 ext_factor, attn_factor, beta_fast, beta_slow);
263
264            cb(Kcur, "Kcur_pos", il);
265
266            cur = build_attn(inp_attn, model.layers[il].wo,
267                    nullptr, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
268                    hparams.f_attention_scale, il);
269        } else {
270            // reuse KV cache of earlier layers
271            cur = build_attn(inp_attn,
272                    model.layers[il].wo, nullptr, model.layers[il].wo_s,
273                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);
274        }
275
276        // TODO @ngxson : strip unused token right after the last KV layer to speed up prompt processing
277        // keep all rows when extracting unmasked nextn embeddings (MTP target needs the hidden state for every token)
278        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
279            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
280            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
281        }
282        cur = build_norm(cur,
283                model.layers[il].attn_post_norm, nullptr,
284                LLM_NORM_RMS, il);
285        cb(cur, "attn_post_norm", il);
286
287        ggml_tensor * attn_out = ggml_add(ctx0, cur, inpL);
288        cb(attn_out, "attn_out", il);
289
290        // feed-forward network
291        const bool is_moe_layer = model.layers[il].ffn_gate_inp != nullptr;
292        if (is_moe_layer) {
293            // MLP (shared exp)
294            ggml_tensor * cur_mlp = build_norm(attn_out,
295                    model.layers[il].ffn_norm, nullptr,
296                    LLM_NORM_RMS, il);
297            cb(cur_mlp, "ffn_norm_1", il);
298
299            cur_mlp = build_ffn(cur_mlp,
300                    model.layers[il].ffn_up,   nullptr, model.layers[il].ffn_up_s,
301                    model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s,
302                    model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,
303                    nullptr,
304                    LLM_FFN_GELU, LLM_FFN_PAR, il);
305            cur_mlp = build_norm(cur_mlp,
306                    model.layers[il].ffn_post_norm_1, nullptr,
307                    LLM_NORM_RMS, il);
308            cb(cur_mlp, "ffn_mlp", il);
309
310            // Expert FFN
311            ggml_tensor * cur_moe = build_norm(attn_out,
312                    model.layers[il].ffn_pre_norm_2, nullptr,
313                    LLM_NORM_RMS, il);
314            cb(cur_moe, "ffn_norm_2", il);
315
316            // custom MoE logits calculation (router operates on attn_out, not cur)
317            ggml_tensor * tmp = ggml_rms_norm(ctx0, attn_out, hparams.f_norm_rms_eps);
318            tmp = ggml_scale(ctx0, tmp, 1.0f / sqrtf((float) n_embd));
319            tmp = ggml_mul(ctx0, tmp, model.layers[il].ffn_gate_inp_s);
320            ggml_tensor * logits = build_lora_mm(model.layers[il].ffn_gate_inp, tmp); // [n_expert, n_tokens]
321            cb(logits, "ffn_moe_logits", il);
322
323            cur_moe = build_moe_ffn(cur_moe,
324                    nullptr, // gate_inp
325                    model.layers[il].ffn_up_exps,
326                    model.layers[il].ffn_gate_exps,
327                    model.layers[il].ffn_down_exps,
328                    nullptr, // exp_probs_b (not used for gemma4)
329                    n_expert, n_expert_used,
330                    LLM_FFN_GELU, true,
331                    1.0f,
332                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
333                    il, logits,
334                    model.layers[il].ffn_gate_up_exps,
335                    model.layers[il].ffn_up_exps_s,
336                    model.layers[il].ffn_gate_exps_s,
337                    model.layers[il].ffn_down_exps_s);
338            cur_moe = build_norm(cur_moe,
339                    model.layers[il].ffn_post_norm_2, nullptr,
340                    LLM_NORM_RMS, il);
341            cb(cur_moe, "ffn_moe", il);
342
343            cur = ggml_add(ctx0, cur_mlp, cur_moe);
344            cb(cur, "ffn_moe_combined", il);
345        } else {
346            cur = build_norm(attn_out,
347                    model.layers[il].ffn_norm, nullptr,
348                    LLM_NORM_RMS, il);
349            cb(cur, "ffn_norm", il);
350
351            cur = build_ffn(cur,
352                    model.layers[il].ffn_up,   nullptr, model.layers[il].ffn_up_s,
353                    model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s,
354                    model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,
355                    nullptr,
356                    LLM_FFN_GELU, LLM_FFN_PAR, il);
357            cb(cur, "ffn_out", il);
358        }
359        cur = build_norm(cur,
360                model.layers[il].ffn_post_norm, nullptr,
361                LLM_NORM_RMS, -1);
362        cb(cur, "ffn_post_norm", il);
363
364        // residual connection
365        cur = ggml_add(ctx0, cur, attn_out);
366
367        // per-layer embedding
368        if (inp_per_layer) {
369            ggml_tensor * pe_in = cur;
370            cb(cur, "pe_in", il);
371
372            cur = build_lora_mm(model.layers[il].per_layer_inp_gate, cur); // [n_embd_per_layer, n_tokens]
373            cur = ggml_gelu(ctx0, cur);
374
375            ggml_tensor * inp_this_layer = gemma4_view_2d_slice(ctx0, inp_per_layer, il); // [n_embd_per_layer, n_tokens]
376
377            // TODO @ngxson : improve this
378            if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
379                inp_this_layer = ggml_get_rows(ctx0, inp_this_layer, inp_out_ids);
380            }
381
382            cur = ggml_mul(ctx0, cur, inp_this_layer);
383            cur = build_lora_mm(model.layers[il].per_layer_proj, cur); // [n_embd, n_tokens]
384            cur = build_norm(cur, model.layers[il].per_layer_post_norm, nullptr, LLM_NORM_RMS, il);
385            cb(cur, "per_layer_embd_out", il);
386
387            // residual connection
388            cur = ggml_add(ctx0, pe_in, cur);
389        }
390
391        // layer_scalar
392        if (model.layers[il].out_scale) {
393            cur = ggml_mul(ctx0, cur, model.layers[il].out_scale);
394            cb(cur, "out_scaled", il);
395        }
396
397        cur = build_cvec(cur, il);
398        cb(cur, "l_out", il);
399
400        // input for next layer
401        inpL = cur;
402    }
403    cur = inpL;
404
405    cur = build_norm(cur,
406            model.output_norm, nullptr,
407            LLM_NORM_RMS, -1);
408
409    // Expose the post-output-norm hidden state (the LM-head input feature) so that
410    // MTP draft contexts can read it via llama_get_embeddings_nextn_ith() as the
411    // recurrent h input. This matches the reference (transformers/vLLM/SGLang),
412    // which feeds the drafter the target's post-final-norm hidden state.
413    cb(cur, "h_nextn", -1);
414    res->t_h_nextn = cur;
415
416    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
417        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
418    }
419
420    cb(cur, "result_norm", -1);
421    res->t_embd = cur;
422
423    // lm_head
424    cur = build_lora_mm(model.output, cur, model.output_s);
425
426    if (hparams.f_final_logit_softcapping) {
427        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
428        cur = ggml_tanh(ctx0, cur);
429        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
430    }
431
432    cb(cur, "result_output", -1);
433    res->t_logits = cur;
434
435    ggml_build_forward_expand(gf, cur);
436}
437
438// equivalent to get_per_layer_inputs() in python code
439// output shape: [n_embd_per_layer, n_layer, n_tokens]
440ggml_tensor * llama_model_gemma4::graph::build_inp_per_layer() {
441    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
442
443    ggml_tensor * inp_per_layer;
444    float tok_embd_scale = sqrtf((float) n_embd_per_layer);
445    if (ubatch.token) {
446        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
447        ggml_set_input(inp->tokens);
448        res->t_inp_tokens = inp->tokens;
449
450        inp_per_layer = ggml_get_rows  (ctx0, model.per_layer_tok_embd, inp->tokens);
451        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, n_tokens);
452        inp_per_layer = ggml_scale     (ctx0, inp_per_layer, tok_embd_scale);
453        cb(inp_per_layer, "inp_per_layer_selected", -1);
454
455        res->add_input(std::move(inp));
456    } else {
457        // Multimodal embedding path: use padding token (ID=0) embedding
458        // TODO: verify if this is the correct behavior in transformers implementation
459        const int64_t embd_size = model.per_layer_tok_embd->ne[0];  // n_embd_per_layer * n_layer
460
461        // Extract and dequantize padding token embedding (row 0)
462        ggml_tensor * padding = ggml_view_1d(ctx0, model.per_layer_tok_embd, embd_size, 0);
463        inp_per_layer = ggml_cast (ctx0, padding, GGML_TYPE_F32);
464        inp_per_layer = ggml_scale(ctx0, inp_per_layer, tok_embd_scale);
465
466        // Reshape to [n_embd_per_layer, n_layer, 1]
467        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, 1);
468        cb(inp_per_layer, "inp_per_layer_multimodal", -1);
469    }
470    return inp_per_layer;
471}
472
473// equivalent to project_per_layer_inputs() in python code
474// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
475// inp_batch     shape: [n_embd, n_tokens]
476// inp_per_layer shape: [n_embd_per_layer, n_layer, n_tokens] (from build_inp_per_layer)
477// output shape: [n_embd_per_layer, n_tokens, n_layer]
478ggml_tensor * llama_model_gemma4::graph::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {
479    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
480    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
481
482    // note: this matrix multiplication will be performed in the input layer (i.e. on the CPU)
483    ggml_tensor * per_layer_proj;
484    per_layer_proj = ggml_mul_mat   (ctx0, model.per_layer_model_proj, inp_batch);
485    per_layer_proj = ggml_scale     (ctx0, per_layer_proj, per_layer_projection_scale);
486    per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_per_layer, n_layer, n_tokens);
487
488    per_layer_proj = build_norm(per_layer_proj, model.per_layer_proj_norm, nullptr, LLM_NORM_RMS, -1);
489    cb(per_layer_proj, "per_layer_proj", -1);
490
491    inp_per_layer = ggml_add  (ctx0, per_layer_proj, inp_per_layer);
492    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
493    cb(inp_per_layer, "inp_per_layer", -1);
494
495    // permute to shape: [n_embd_per_layer, n_tokens, n_layer]
496    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
497    return inp_per_layer;
498}