Brunobkr/llama.cpp_AlgMor24_github

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_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());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(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, 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    switch (hparams.n_layer()) {23        case 30: type = LLM_TYPE_26B_A4B; break;24        case 35: type = LLM_TYPE_E2B; break;25        case 42: type = LLM_TYPE_E4B; break;26        case 60: type = LLM_TYPE_31B; break;27        default: type = LLM_TYPE_UNKNOWN;28    }29}30 31void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {32    LLAMA_LOAD_LOCALS;33 34    const uint32_t n_embd_per_layer = hparams.n_embd_per_layer;35    const int64_t  n_ff_exp         = hparams.n_ff_exp;36 37    if (n_embd_head_k != n_embd_head_v) {38        throw std::runtime_error("Gemma 4 requires n_embd_head_k == n_embd_head_v");39    }40    if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {41        throw std::runtime_error("Gemma 4 requires n_embd_head_k_swa == n_embd_head_v_swa");42    }43 44    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);45    // if output is NULL, init from the input tok embed46    if (output == NULL) {47        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);48    }49 50    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);51 52    if (n_embd_per_layer > 0) {53        per_layer_tok_embd   = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"),    {n_embd_per_layer * n_layer, n_vocab}, 0);54        per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0);55        per_layer_proj_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight", 0), {n_embd_per_layer}, 0);56    }57 58    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);59 60    int rope_freqs_flag = 0;61 62    for (int i = 0; i < n_layer; ++i) {63        auto & layer = layers[i];64        const int64_t n_head      = hparams.n_head(i);65        const int64_t n_embd_head = hparams.n_embd_head_k(i);66        const int64_t n_embd_k    = hparams.n_embd_k_gqa(i);67        const int64_t n_embd_v    = hparams.n_embd_v_gqa(i);68        const int     kv_flags    = hparams.has_kv(i) ? 0 : TENSOR_NOT_REQUIRED;69 70        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);71 72        // note: use_alternative_attention (v_proj is optional, if it's not present, use k_proj)73        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head * n_head}, 0);74        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k}, kv_flags);75        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);76        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);77 78        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head}, 0);79        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head}, kv_flags);80        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);81 82        layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, TENSOR_NOT_REQUIRED);83 84        if (!hparams.is_swa(i)) {85            // full_attention layers use rope_freqs for proportional rope86            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_embd_head/2}, rope_freqs_flag);87            rope_freqs_flag = TENSOR_DUPLICATED;88        }89 90        // handle use_double_wide_mlp91        int64_t n_ff_cur = hparams.n_ff(i);92 93        // for expert layers, we use normal FFN as shared expert (same as python code)94        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);95        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff_cur}, 0);96        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff_cur}, 0);97        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);98        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);99 100        // MoE router101        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);102        bool has_expert = layer.ffn_gate_inp != nullptr;103 104        // norm105        if (has_expert) {106            layer.ffn_gate_inp_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "scale", i), {n_embd}, 0);107 108            layer.ffn_pre_norm_2  = create_tensor(tn(LLM_TENSOR_FFN_PRE_NORM_2,  "weight", i), {n_embd}, 0);109            layer.ffn_post_norm_1 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_1, "weight", i), {n_embd}, 0);110            layer.ffn_post_norm_2 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_2, "weight", i), {n_embd}, 0);111 112            // MoE FFN113            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);114 115            if (layer.ffn_gate_up_exps == nullptr) {116                layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);117                layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);118            }119 120            layer.ffn_down_exps     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,     "weight", i), {n_ff_exp, n_embd, n_expert}, 0);121 122            // per-expert scale will be loaded as down_exps_s at the end of the current switch case123        }124 125        // per-layer embeddings126        if (n_embd_per_layer > 0) {127            layer.per_layer_inp_gate   = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE,  "weight", i), {n_embd, n_embd_per_layer}, 0);128            layer.per_layer_proj       = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ,      "weight", i), {n_embd_per_layer, n_embd}, 0);129            layer.per_layer_post_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);130        }131    }132}133 134std::unique_ptr<llm_graph_context> llama_model_gemma4::build_arch_graph(const llm_graph_params & params) const {135    return std::make_unique<graph>(*this, params);136}137 138// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim139static ggml_tensor * ggml_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, int idx) {140    GGML_ASSERT(idx < (int) x->ne[2]);141    return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]),142                        idx * x->ne[0] * x->ne[1] * ggml_element_size(x));143}144 145llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_params & params) :146        llm_graph_context(params),147        model(model),148        n_embd_per_layer(model.hparams.n_embd_per_layer) {149    ggml_tensor * cur;150    ggml_tensor * inpL;151 152    inpL = build_inp_embd(model.tok_embd);153 154    // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)155    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);156    cb(inpL, "inp_scaled", -1);157 158    // inp_pos - contains the positions159    ggml_tensor * inp_pos = build_inp_pos();160 161    // TODO: is causal == true correct? might need some changes162    auto * inp_attn = build_attn_inp_kv_iswa();163 164    ggml_tensor * inp_out_ids = build_inp_out_ids();165 166    ggml_tensor * inp_per_layer = nullptr;167    if (model.per_layer_tok_embd) {168        inp_per_layer = build_inp_per_layer();169        ggml_build_forward_expand(gf, inp_per_layer);170 171        // inp_per_layer shape: [n_embd_per_layer, n_tokens, n_layer]172        inp_per_layer = project_per_layer_inputs(inpL, inp_per_layer);173    }174 175    for (int il = 0; il < n_layer; ++il) {176        const int64_t n_embd_head = hparams.n_embd_head_k(il);177        GGML_ASSERT(n_embd_head == hparams.n_embd_head_v(il));178 179        const int64_t n_head    = hparams.n_head(il);180        const int64_t n_head_kv = hparams.n_head_kv(il);181 182        const float freq_base_l  = model.get_rope_freq_base(cparams, il);183        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);184        const int   n_rot_l      = hparams.n_rot(il);185 186        res->t_layer_inp[il] = inpL;187 188        // norm189        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);190        cb(cur, "attn_norm", il);191 192        ggml_tensor * freq_factors = nullptr;193        if (!hparams.is_swa(il)) {194            // full_attention layers use rope_freqs for proportional rope195            freq_factors = model.layers[il].rope_freqs;196        }197 198        // Q projection (shared for both non-KV and KV layers)199        // this is to mirror Gemma4Attention in pytorch code200        ggml_tensor * Qcur;201        {202            Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);203            cb(Qcur, "Qcur", il);204 205            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);206 207            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);208            cb(Qcur, "Qcur_normed", il);209 210            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,211                                 ext_factor, attn_factor, beta_fast, beta_slow);212            cb(Qcur, "Qcur_pos", il);213        }214 215        // self-attention216        if (hparams.has_kv(il)) {217            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);218            cb(Kcur, "Kcur", il);219 220            ggml_tensor * Vcur = model.layers[il].wv221                                    ? build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s)222                                    : Kcur; // if v_proj is not present, use Kcur as Vcur223            cb(Vcur, "Vcur", il);224 225            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);226            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);227 228            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);229            Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);230 231            cb(Kcur, "Kcur_normed", il);232            cb(Vcur, "Vcur_normed", il);233 234            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, freq_factors, n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,235                                 ext_factor, attn_factor, beta_fast, beta_slow);236 237            cb(Kcur, "Kcur_pos", il);238 239            cur = build_attn(inp_attn, model.layers[il].wo,240                    nullptr, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,241                    hparams.f_attention_scale, il);242        } else {243            // reuse KV cache of earlier layers244            cur = build_attn(inp_attn,245                    model.layers[il].wo, nullptr, model.layers[il].wo_s,246                    Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il);247        }248 249        // TODO @ngxson : strip unused token right after the last KV layer to speed up prompt processing250        // keep all rows when extracting unmasked nextn embeddings (MTP target needs the hidden state for every token)251        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {252            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);253            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);254        }255        cur = build_norm(cur,256                model.layers[il].attn_post_norm, nullptr,257                LLM_NORM_RMS, il);258        cb(cur, "attn_post_norm", il);259 260        ggml_tensor * attn_out = ggml_add(ctx0, cur, inpL);261        cb(attn_out, "attn_out", il);262 263        // feed-forward network264        const bool is_moe_layer = model.layers[il].ffn_gate_inp != nullptr;265        if (is_moe_layer) {266            // MLP (shared exp)267            ggml_tensor * cur_mlp = build_norm(attn_out,268                    model.layers[il].ffn_norm, nullptr,269                    LLM_NORM_RMS, il);270            cb(cur_mlp, "ffn_norm_1", il);271 272            cur_mlp = build_ffn(cur_mlp,273                    model.layers[il].ffn_up,   nullptr, model.layers[il].ffn_up_s,274                    model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s,275                    model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,276                    nullptr,277                    LLM_FFN_GELU, LLM_FFN_PAR, il);278            cur_mlp = build_norm(cur_mlp,279                    model.layers[il].ffn_post_norm_1, nullptr,280                    LLM_NORM_RMS, il);281            cb(cur_mlp, "ffn_mlp", il);282 283            // Expert FFN284            ggml_tensor * cur_moe = build_norm(attn_out,285                    model.layers[il].ffn_pre_norm_2, nullptr,286                    LLM_NORM_RMS, il);287            cb(cur_moe, "ffn_norm_2", il);288 289            // custom MoE logits calculation (router operates on attn_out, not cur)290            ggml_tensor * tmp = ggml_rms_norm(ctx0, attn_out, hparams.f_norm_rms_eps);291            tmp = ggml_scale(ctx0, tmp, 1.0f / sqrtf((float) n_embd));292            tmp = ggml_mul(ctx0, tmp, model.layers[il].ffn_gate_inp_s);293            ggml_tensor * logits = build_lora_mm(model.layers[il].ffn_gate_inp, tmp); // [n_expert, n_tokens]294            cb(logits, "ffn_moe_logits", il);295 296            cur_moe = build_moe_ffn(cur_moe,297                    nullptr, // gate_inp298                    model.layers[il].ffn_up_exps,299                    model.layers[il].ffn_gate_exps,300                    model.layers[il].ffn_down_exps,301                    nullptr, // exp_probs_b (not used for gemma4)302                    n_expert, n_expert_used,303                    LLM_FFN_GELU, true,304                    1.0f,305                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,306                    il, logits,307                    model.layers[il].ffn_gate_up_exps,308                    model.layers[il].ffn_up_exps_s,309                    model.layers[il].ffn_gate_exps_s,310                    model.layers[il].ffn_down_exps_s);311            cur_moe = build_norm(cur_moe,312                    model.layers[il].ffn_post_norm_2, nullptr,313                    LLM_NORM_RMS, il);314            cb(cur_moe, "ffn_moe", il);315 316            cur = ggml_add(ctx0, cur_mlp, cur_moe);317            cb(cur, "ffn_moe_combined", il);318        } else {319            cur = build_norm(attn_out,320                    model.layers[il].ffn_norm, nullptr,321                    LLM_NORM_RMS, il);322            cb(cur, "ffn_norm", il);323 324            cur = build_ffn(cur,325                    model.layers[il].ffn_up,   nullptr, model.layers[il].ffn_up_s,326                    model.layers[il].ffn_gate, nullptr, model.layers[il].ffn_gate_s,327                    model.layers[il].ffn_down, nullptr, model.layers[il].ffn_down_s,328                    nullptr,329                    LLM_FFN_GELU, LLM_FFN_PAR, il);330            cb(cur, "ffn_out", il);331        }332        cur = build_norm(cur,333                model.layers[il].ffn_post_norm, nullptr,334                LLM_NORM_RMS, -1);335        cb(cur, "ffn_post_norm", il);336 337        // residual connection338        cur = ggml_add(ctx0, cur, attn_out);339 340        // per-layer embedding341        if (inp_per_layer) {342            ggml_tensor * pe_in = cur;343            cb(cur, "pe_in", il);344 345            cur = build_lora_mm(model.layers[il].per_layer_inp_gate, cur); // [n_embd_per_layer, n_tokens]346            cur = ggml_gelu(ctx0, cur);347 348            ggml_tensor * inp_this_layer = ggml_view_2d_slice(ctx0, inp_per_layer, il); // [n_embd_per_layer, n_tokens]349 350            // TODO @ngxson : improve this351            if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {352                inp_this_layer = ggml_get_rows(ctx0, inp_this_layer, inp_out_ids);353            }354 355            cur = ggml_mul(ctx0, cur, inp_this_layer);356            cur = build_lora_mm(model.layers[il].per_layer_proj, cur); // [n_embd, n_tokens]357            cur = build_norm(cur, model.layers[il].per_layer_post_norm, nullptr, LLM_NORM_RMS, il);358            cb(cur, "per_layer_embd_out", il);359 360            // residual connection361            cur = ggml_add(ctx0, pe_in, cur);362        }363 364        // layer_scalar365        if (model.layers[il].out_scale) {366            cur = ggml_mul(ctx0, cur, model.layers[il].out_scale);367            cb(cur, "out_scaled", il);368        }369 370        cur = build_cvec(cur, il);371        cb(cur, "l_out", il);372 373        // input for next layer374        inpL = cur;375    }376    cur = inpL;377 378    cur = build_norm(cur,379            model.output_norm, nullptr,380            LLM_NORM_RMS, -1);381 382    // Expose the post-output-norm hidden state (the LM-head input feature) so that383    // MTP draft contexts can read it via llama_get_embeddings_nextn_ith() as the384    // recurrent h input. This matches the reference (transformers/vLLM/SGLang),385    // which feeds the drafter the target's post-final-norm hidden state.386    cb(cur, "h_nextn", -1);387    res->t_h_nextn = cur;388 389    if (!cparams.embeddings_nextn_masked && inp_out_ids) {390        cur = ggml_get_rows(ctx0, cur, inp_out_ids);391    }392 393    cb(cur, "result_norm", -1);394    res->t_embd = cur;395 396    // lm_head397    cur = build_lora_mm(model.output, cur, model.output_s);398 399    if (hparams.f_final_logit_softcapping) {400        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);401        cur = ggml_tanh(ctx0, cur);402        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);403    }404 405    cb(cur, "result_output", -1);406    res->t_logits = cur;407 408    ggml_build_forward_expand(gf, cur);409}410 411// equivalent to get_per_layer_inputs() in python code412// output shape: [n_embd_per_layer, n_layer, n_tokens]413ggml_tensor * llama_model_gemma4::graph::build_inp_per_layer() {414    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);415 416    ggml_tensor * inp_per_layer;417    float tok_embd_scale = sqrtf((float) n_embd_per_layer);418    if (ubatch.token) {419        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);420        ggml_set_input(inp->tokens);421        res->t_inp_tokens = inp->tokens;422 423        inp_per_layer = ggml_get_rows  (ctx0, model.per_layer_tok_embd, inp->tokens);424        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, n_tokens);425        inp_per_layer = ggml_scale     (ctx0, inp_per_layer, tok_embd_scale);426        cb(inp_per_layer, "inp_per_layer_selected", -1);427 428        res->add_input(std::move(inp));429    } else {430        // Multimodal embedding path: use padding token (ID=0) embedding431        // TODO: verify if this is the correct behavior in transformers implementation432        const int64_t embd_size = model.per_layer_tok_embd->ne[0];  // n_embd_per_layer * n_layer433 434        // Extract and dequantize padding token embedding (row 0)435        ggml_tensor * padding = ggml_view_1d(ctx0, model.per_layer_tok_embd, embd_size, 0);436        inp_per_layer = ggml_cast (ctx0, padding, GGML_TYPE_F32);437        inp_per_layer = ggml_scale(ctx0, inp_per_layer, tok_embd_scale);438 439        // Reshape to [n_embd_per_layer, n_layer, 1]440        inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_per_layer, n_layer, 1);441        cb(inp_per_layer, "inp_per_layer_multimodal", -1);442    }443    return inp_per_layer;444}445 446// equivalent to project_per_layer_inputs() in python code447// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim448// inp_batch     shape: [n_embd, n_tokens]449// inp_per_layer shape: [n_embd_per_layer, n_layer, n_tokens] (from build_inp_per_layer)450// output shape: [n_embd_per_layer, n_tokens, n_layer]451ggml_tensor * llama_model_gemma4::graph::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {452    const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);453    const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);454 455    // note: this matrix multiplication will be performed in the input layer (i.e. on the CPU)456    ggml_tensor * per_layer_proj;457    per_layer_proj = ggml_mul_mat   (ctx0, model.per_layer_model_proj, inp_batch);458    per_layer_proj = ggml_scale     (ctx0, per_layer_proj, per_layer_projection_scale);459    per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_per_layer, n_layer, n_tokens);460 461    per_layer_proj = build_norm(per_layer_proj, model.per_layer_proj_norm, nullptr, LLM_NORM_RMS, -1);462    cb(per_layer_proj, "per_layer_proj", -1);463 464    inp_per_layer = ggml_add  (ctx0, per_layer_proj, inp_per_layer);465    inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);466    cb(inp_per_layer, "inp_per_layer", -1);467 468    // permute to shape: [n_embd_per_layer, n_tokens, n_layer]469    inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));470    return inp_per_layer;471}472