Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_mellum::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_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);7 8    if (hparams.n_swa > 0) {9        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;10 11        load_swa_pattern(ml, 4);12 13        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;14        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;15 16        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);17    } else {18        hparams.swa_type = LLAMA_SWA_TYPE_NONE;19    }20 21    switch (hparams.n_layer()) {22        case 28: type = LLM_TYPE_12B_A2_5B; break;23        default: type = LLM_TYPE_UNKNOWN;24    }25}26 27void llama_model_mellum::load_arch_tensors(llama_model_loader &) {28    LLAMA_LOAD_LOCALS;29 30    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);31 32    // output33    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);34    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);35 36    for (int i = 0; i < n_layer; ++i) {37        auto & layer = layers[i];38 39        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);40 41        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);42        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);43 44        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);45        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);46 47        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);48 49        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);50 51        if (n_expert == 0) {52            throw std::runtime_error("n_expert must be > 0 for Mellum");53        }54        if (n_expert_used == 0) {55            throw std::runtime_error("n_expert_used must be > 0 for Mellum");56        }57 58        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;59 60        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);61        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);62        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);63    }64}65 66std::unique_ptr<llm_graph_context> llama_model_mellum::build_arch_graph(const llm_graph_params & params) const {67    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {68        return std::make_unique<graph<true>>(*this, params);69    }70    return std::make_unique<graph<false>>(*this, params);71}72 73template <bool iswa>74llama_model_mellum::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {75    const int64_t n_embd_head = hparams.n_embd_head_v();76 77    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());78    GGML_ASSERT(n_embd_head == n_rot);79 80    ggml_tensor * cur;81    ggml_tensor * inpL;82 83    inpL = build_inp_embd(model.tok_embd);84 85    // inp_pos - contains the positions86    ggml_tensor * inp_pos = build_inp_pos();87 88    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;89    inp_attn_type * inp_attn = nullptr;90 91    if constexpr (iswa) {92        inp_attn = build_attn_inp_kv_iswa();93    } else {94        inp_attn = build_attn_inp_kv();95    }96 97    ggml_tensor * inp_out_ids = build_inp_out_ids();98 99    for (int il = 0; il < n_layer; ++il) {100        ggml_tensor * inpSA = inpL;101 102        // norm103        cur = build_norm(inpL,104                model.layers[il].attn_norm, nullptr,105                LLM_NORM_RMS, il);106        cb(cur, "attn_norm", il);107 108        // self_attention109        {110            // compute Q and K and RoPE them111            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,112                    n_embd_head, n_head, n_head_kv, il);113 114            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);115            cb(Qcur, "Qcur_normed", il);116 117            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);118            cb(Kcur, "Kcur_normed", il);119 120            const bool is_swa = hparams.is_swa(il);121 122            if (is_swa) {123                // For sliding window layers, use regular rope with no yarn rope scaling.124                // This is achieved here by setting freq_scale and attn_factor to 1.125                // We also set ext_factor to 0 to avoid a few unnecessary computations.126                Qcur = ggml_rope_ext(127                    ctx0, Qcur, inp_pos, nullptr,128                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,129                    0.0, 1.0, beta_fast, beta_slow130                    );131 132                Kcur = ggml_rope_ext(133                    ctx0, Kcur, inp_pos, nullptr,134                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,135                    0.0, 1.0, beta_fast, beta_slow136                    );137            } else {138                Qcur = ggml_rope_ext(139                    ctx0, Qcur, inp_pos, nullptr,140                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,141                    ext_factor, attn_factor, beta_fast, beta_slow142                    );143 144                Kcur = ggml_rope_ext(145                    ctx0, Kcur, inp_pos, nullptr,146                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,147                    ext_factor, attn_factor, beta_fast, beta_slow148                    );149            }150 151            cb(Qcur, "Qcur", il);152            cb(Kcur, "Kcur", il);153            cb(Vcur, "Vcur", il);154 155            cur = build_attn(inp_attn,156                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,157                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);158        }159        if (il == n_layer - 1 && inp_out_ids) {160            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);161            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);162        }163        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);164        cb(ffn_inp, "ffn_inp", il);165 166        // MoE167        cur = build_norm(ffn_inp,168                model.layers[il].ffn_norm, nullptr,169                LLM_NORM_RMS, il);170        cb(cur, "ffn_norm", il);171 172        ggml_tensor * moe_out =173            build_moe_ffn(cur,174                    model.layers[il].ffn_gate_inp,175                    model.layers[il].ffn_up_exps,176                    model.layers[il].ffn_gate_exps,177                    model.layers[il].ffn_down_exps,178                    nullptr,179                    n_expert, n_expert_used,180                    LLM_FFN_SILU, true,181                    hparams.expert_weights_scale,182                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,183                    il,184                    nullptr, nullptr,185                    model.layers[il].ffn_up_exps_s,186                    model.layers[il].ffn_gate_exps_s,187                    model.layers[il].ffn_down_exps_s);188        cb(moe_out, "ffn_moe_out", il);189        cur = moe_out;190 191        cur = ggml_add(ctx0, cur, ffn_inp);192        cb(cur, "ffn_out", il);193 194        cur = build_cvec(cur, il);195        cb(cur, "l_out", il);196 197        // input for next layer198        inpL = cur;199    }200    cur = inpL;201 202    cur = build_norm(cur,203            model.output_norm, nullptr,204            LLM_NORM_RMS, -1);205 206    cb(cur, "result_norm", -1);207    res->t_embd = cur;208 209    // lm_head210    cur = build_lora_mm(model.output, cur, model.output_s);211 212    cb(cur, "result_output", -1);213    res->t_logits = cur;214 215    ggml_build_forward_expand(gf, cur);216}217 218template struct llama_model_mellum::graph<false>;219template struct llama_model_mellum::graph<true>;220