Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_llama4::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_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);6    ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);7 8    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);9    if (found_swa && hparams.n_swa == 0) {10        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;11        hparams.n_no_rope_layer_step = hparams.n_layer(); // always use rope12    } else {13        hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;14        hparams.n_swa                   = 8192;15        hparams.n_attn_temp_floor_scale = 8192;16        hparams.f_attn_temp_scale       = 0.1f;17        hparams.f_attn_temp_offset      = 1.0f;18 19        load_swa_pattern(ml, 4); // pattern: 3 chunked - 1 full20 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    }25 26    switch (hparams.n_expert) {27        case 0: {28            // MobileLLM (no MoE)29            switch (hparams.n_embd) {30                case 2048: type = LLM_TYPE_140M; break;31                case 4096: type = LLM_TYPE_360M; break;32                case 6144: type = LLM_TYPE_950M; break;33                default:   type = LLM_TYPE_UNKNOWN;34            }35        } break;36        case 16:  type = LLM_TYPE_17B_16E; break;37        case 128: type = LLM_TYPE_17B_128E; break;38        default:  type = LLM_TYPE_UNKNOWN;39    }40 41    hparams.use_kq_norm = type != LLM_TYPE_17B_128E;42}43 44void llama_model_llama4::load_arch_tensors(llama_model_loader &) {45    LLAMA_LOAD_LOCALS;46 47    if (n_expert == 0) {48        throw std::runtime_error(arch_name() + " model cannot have zero experts");49    }50    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);51 52    // output53    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);54    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);55 56    // if output is NULL, init from the input tok embed57    if (output == NULL) {58        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);59    }60 61    for (int i = 0; i < n_layer; ++i) {62        const bool is_moe_layer = hparams.n_moe_layer_step > 0 && (i + 1) % hparams.n_moe_layer_step == 0;63 64        auto & layer = layers[i];65 66        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);67 68        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);69        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);70 71        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);72 73        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));74 75        if (is_moe_layer) {76            const int64_t n_ff_exp = hparams.n_ff_exp();77 78            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);79            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);80            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);81            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);82 83            // Shared expert84            const int64_t n_ff_shexp = n_ff_exp;85            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp}, 0);86            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd    }, 0);87            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);88        } else {89            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);90            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);91            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);92        }93    }94}95 96std::unique_ptr<llm_graph_context> llama_model_llama4::build_arch_graph(const llm_graph_params & params) const {97    if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {98        return std::make_unique<graph<false>>(*this, params);99    } else {100        return std::make_unique<graph<true>>(*this, params);101    }102}103 104template <bool iswa>105llama_model_llama4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {106    const int64_t n_embd_head = hparams.n_embd_head_v();107 108    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());109    GGML_ASSERT(n_embd_head == n_rot);110 111    ggml_tensor * cur;112    ggml_tensor * inpL;113 114    inpL = build_inp_embd(model.tok_embd);115 116    // inp_pos - contains the positions117    ggml_tensor * inp_pos = build_inp_pos();118 119    // temperature tuning120    ggml_tensor * inp_attn_scale = nullptr;121    inp_attn_scale = build_inp_attn_scale();122 123    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;124    inp_attn_type * inp_attn = nullptr;125 126    if constexpr (iswa) {127        inp_attn = build_attn_inp_kv_iswa();128    } else {129        inp_attn = build_attn_inp_kv();130    }131 132    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;133 134    ggml_tensor * inp_out_ids = build_inp_out_ids();135 136    for (int il = 0; il < n_layer; ++il) {137        const float freq_base_l  = model.get_rope_freq_base (cparams, il);138        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);139 140        ggml_tensor * inpSA = inpL;141 142        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous143        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&144                              (il + 1) % hparams.n_no_rope_layer_step != 0;145 146        // norm147        cur = build_norm(inpL,148                model.layers[il].attn_norm, NULL,149                LLM_NORM_RMS, il);150        cb(cur, "attn_norm", il);151 152        // self-attention153        {154            // rope freq factors for llama3; may return nullptr for llama2 and other models155            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);156 157            // compute Q and K and RoPE them158            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,159                    n_embd_head, n_head, n_head_kv, il);160 161            if (use_rope) {162                Qcur = ggml_rope_ext(163                        ctx0, Qcur, inp_pos, rope_factors,164                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,165                        ext_factor, attn_factor, beta_fast, beta_slow166                        );167 168                Kcur = ggml_rope_ext(169                        ctx0, Kcur, inp_pos, rope_factors,170                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,171                        ext_factor, attn_factor, beta_fast, beta_slow172                        );173            } else if (inp_attn_scale) {174                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);175            }176            cb(Qcur, "Qcur", il);177            cb(Kcur, "Kcur", il);178            cb(Vcur, "Vcur", il);179 180            if (use_rope && hparams.use_kq_norm) {181                // Llama4TextL2Norm182                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);183                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);184                cb(Qcur, "Qcur_normed", il);185                cb(Kcur, "Kcur_normed", il);186            }187            cur = build_attn(inp_attn,188                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,189                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);190            cb(cur, "attn_out", il);191        }192        if (il == n_layer - 1 && inp_out_ids) {193            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);194            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);195        }196        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);197        cb(ffn_inp, "ffn_inp", il);198 199        // feed-forward network (non-MoE)200        if (model.layers[il].ffn_gate_inp == nullptr) {201            cur = build_norm(ffn_inp,202                    model.layers[il].ffn_norm, NULL,203                    LLM_NORM_RMS, il);204            cb(cur, "ffn_norm", il);205 206            cur = build_ffn(cur,207                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,208                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,209                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,210                    NULL,211                    LLM_FFN_SILU, LLM_FFN_PAR, il);212            cb(cur, "ffn_out", il);213        } else {214            ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,215                    model.layers[il].ffn_norm, NULL,216                    LLM_NORM_RMS, il);217            cb(cur, "ffn_norm", il);218 219            ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,220                    model.layers[il].ffn_gate_inp,221                    model.layers[il].ffn_up_exps,222                    model.layers[il].ffn_gate_exps,223                    model.layers[il].ffn_down_exps,224                    nullptr,225                    n_expert, n_expert_used,226                    LLM_FFN_SILU, false,227                    hparams.expert_weights_scale,228                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,229                    il);230 231            // Shared experts232            ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,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(shexp_out, "ffn_moe_shexp", il);239 240            cur = ggml_add(ctx0, moe_out, shexp_out);241            cb(cur, "ffn_moe_out_merged", il);242        }243        cur = ggml_add(ctx0, cur, ffn_inp);244        cb(cur, "ffn_out", il);245 246        cur = build_cvec(cur, il);247        cb(cur, "l_out", il);248 249        // input for next layer250        inpL = cur;251    }252    cur = inpL;253 254    cur = build_norm(cur,255            model.output_norm, NULL,256            LLM_NORM_RMS, -1);257 258    cb(cur, "result_norm", -1);259    res->t_embd = cur;260 261    // lm_head262    cur = build_lora_mm(model.output, cur, model.output_s);263 264    cb(cur, "result_output", -1);265    res->t_logits = cur;266 267    ggml_build_forward_expand(gf, cur);268}269 270// Explicit template instantiations271template struct llama_model_llama4::graph<false>;272template struct llama_model_llama4::graph<true>;273