Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {4    if (hparams.n_layer() == 64) {    // 32B5        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;6        hparams.n_swa = 4096;7        uint32_t swa_period = 4;8        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);9        hparams.set_swa_pattern(swa_period);10 11        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;12        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;13        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);14    }15 16    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);17    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);18    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.n_layer_nextn, false);19 20    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");21 22    switch (hparams.n_layer()) {23        case 30: type = LLM_TYPE_1_2B; break;24        case 64: type = LLM_TYPE_32B; break;25        default: type = LLM_TYPE_UNKNOWN;26    }27}28 29void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {30    LLAMA_LOAD_LOCALS;31 32    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);33 34    // output35    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);36    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);37 38    // if output is NULL, init from the input tok embed39    if (output == NULL) {40        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);41    }42 43    for (int i = 0; i < n_layer_all; ++i) {44        const bool is_nextn = i >= n_layer;45        int flags = 0;46        if (is_nextn) {47            // NextN/MTP layers are preserved in GGUF but are not executed yet.48            flags |= TENSOR_SKIP;49        }50 51        auto & layer = layers[i];52 53        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);54        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, flags);55 56        if (!is_nextn) {57            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));58        }59 60        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);61        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);62        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);63 64        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);65        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);66        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);67        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);68 69        if (is_nextn) {70            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);71            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);72            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);73            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);74        }75    }76}77 78std::unique_ptr<llm_graph_context> llama_model_exaone4::build_arch_graph(const llm_graph_params & params) const {79    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {80        return std::make_unique<graph<true>>(*this, params);81    } else {82        return std::make_unique<graph<false>>(*this, params);83    }84}85 86template <bool iswa>87llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :88    llm_graph_context(params) {89    const int64_t n_embd_head = hparams.n_embd_head_k();90 91    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());92    GGML_ASSERT(n_embd_head == n_rot);93 94    ggml_tensor * cur;95    ggml_tensor * inpL;96 97    inpL = build_inp_embd(model.tok_embd);98 99    // inp_pos - contains the positions100    ggml_tensor * inp_pos = build_inp_pos();101 102    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;103    inp_attn_type * inp_attn = nullptr;104 105    if constexpr (iswa) {106        inp_attn = build_attn_inp_kv_iswa();107    } else {108        inp_attn = build_attn_inp_kv();109    }110    ggml_tensor * inp_out_ids = build_inp_out_ids();111 112    for (int il = 0; il < n_layer; ++il) {113        ggml_tensor * inpSA = inpL;114 115        // use RoPE for SWA layers or non-SWA models116        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;117 118        cur = inpL;119 120        // self-attention121        {122            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);123 124            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,125                    n_embd_head, n_head, n_head_kv, il);126 127            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);128            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);129            cb(Qcur, "Qcur_normed", il);130            cb(Kcur, "Kcur_normed", il);131 132            if (use_rope) {133                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,134                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);135 136                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,137                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);138            }139            cb(Qcur, "Qcur", il);140            cb(Kcur, "Kcur", il);141            cb(Vcur, "Vcur", il);142 143            cur = build_attn(inp_attn,144                    model.layers[il].wo, NULL, model.layers[il].wo_s,145                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);146            cb(cur, "attn_out", il);147        }148        if (il == n_layer - 1 && inp_out_ids) {149            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);150            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);151        }152        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);153        cb(cur, "attn_post_norm", il);154 155        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);156        cb(ffn_inp, "ffn_inp", il);157 158        // feed-forward network159        cur = build_ffn(ffn_inp,160                model.layers[il].ffn_up, NULL, NULL,161                model.layers[il].ffn_gate, NULL, NULL,162                model.layers[il].ffn_down, NULL, NULL, NULL,163                LLM_FFN_SILU, LLM_FFN_PAR, il);164        cb(cur, "ffn_out", il);165 166        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);167        cb(cur, "ffn_post_norm", -1);168 169        cur = ggml_add(ctx0, cur, ffn_inp);170 171        cur = build_cvec(cur, il);172        cb(cur, "l_out", il);173 174        // input for next layer175        inpL = cur;176    }177    cur = inpL;178 179    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);180 181    cb(cur, "result_norm", -1);182    res->t_embd = cur;183 184    // lm_head185    cur = build_lora_mm(model.output, cur, model.output_s);186 187    cb(cur, "result_output", -1);188    res->t_logits = cur;189 190    ggml_build_forward_expand(gf, cur);191}192 193// Explicit template instantiations194template struct llama_model_exaone4::graph<false>;195template struct llama_model_exaone4::graph<true>;196