Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_openai_moe::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);7 8    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;9    load_swa_pattern(ml, 2);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    switch (hparams.n_layer()) {16        case 24: type = LLM_TYPE_20B; break;17        case 36: type = LLM_TYPE_120B; break;18        default: type = LLM_TYPE_UNKNOWN;19    }20}21 22void llama_model_openai_moe::load_arch_tensors(llama_model_loader &) {23    LLAMA_LOAD_LOCALS;24 25    const int64_t n_ff_exp = hparams.n_ff_exp();26 27    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);28 29    // output30    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);31    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);32 33    for (int i = 0; i < n_layer; ++i) {34        auto & layer = layers[i];35 36        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);37        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);38 39        create_tensor_qkv(layer, i, n_embd, n_head * n_rot, n_head_kv * n_rot, n_head_kv * n_rot, 0);40        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);41 42        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);43 44        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {  n_embd, n_expert}, 0);45        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);46        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);47        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);48 49        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);50 51        layer.ffn_gate_inp_b  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "bias", i), {n_expert}, 0);52        layer.ffn_gate_exps_b = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0);53        layer.ffn_down_exps_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "bias", i), {  n_embd, n_expert}, 0);54        layer.ffn_up_exps_b   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "bias", i), {n_ff_exp, n_expert}, 0);55    }56}57 58std::unique_ptr<llm_graph_context> llama_model_openai_moe::build_arch_graph(const llm_graph_params & params) const {59    return std::make_unique<graph>(*this, params);60}61 62llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {63    ggml_tensor * cur;64    ggml_tensor * inpL;65 66    inpL = build_inp_embd(model.tok_embd);67 68    // inp_pos - contains the positions69    ggml_tensor * inp_pos = build_inp_pos();70 71    auto * inp_attn = build_attn_inp_kv_iswa();72 73    ggml_tensor * inp_out_ids = build_inp_out_ids();74 75    for (int il = 0; il < n_layer; ++il) {76        res->t_layer_inp[il] = inpL;77 78        const float freq_base_l  = model.get_rope_freq_base (cparams, il);79        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);80 81        ggml_tensor * inpSA = inpL;82 83        // norm84        cur = build_norm(inpL,85                model.layers[il].attn_norm, nullptr,86                LLM_NORM_RMS, il);87        cb(cur, "attn_norm", il);88 89        // self-attention90        {91            // compute Q and K and RoPE them92            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,93                    n_rot, n_head, n_head_kv, il);94 95            Qcur = ggml_rope_ext(96                    ctx0, Qcur, inp_pos, nullptr,97                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,98                    ext_factor, attn_factor, beta_fast, beta_slow99                    );100 101            Kcur = ggml_rope_ext(102                    ctx0, Kcur, inp_pos, nullptr,103                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,104                    ext_factor, attn_factor, beta_fast, beta_slow105                    );106 107            cb(Qcur, "Qcur", il);108            cb(Kcur, "Kcur", il);109            cb(Vcur, "Vcur", il);110 111            cur = build_attn(inp_attn,112                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,113                    Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);114 115            cb(cur, "attn_out", il);116        }117        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {118            // skip computing output for unused tokens119            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);120            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);121        }122        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);123        cb(ffn_inp, "ffn_inp", il);124 125        cur = ffn_inp;126        cur = build_norm(cur,127                model.layers[il].attn_post_norm, nullptr,128                LLM_NORM_RMS, il);129        cb(cur, "attn_post_norm", il);130 131        // MoE branch132        cur = build_moe_ffn(cur,133                model.layers[il].ffn_gate_inp,  model.layers[il].ffn_gate_inp_b,134                model.layers[il].ffn_up_exps,   model.layers[il].ffn_up_exps_b,135                model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,136                model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,137                nullptr,138                n_expert, n_expert_used,139                LLM_FFN_SWIGLU_OAI_MOE, false,140                hparams.expert_weights_scale,141                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,142                il);143        cb(cur, "ffn_moe_out", il);144 145        cur = ggml_add(ctx0, cur, ffn_inp);146 147        cur = build_cvec(cur, il);148        cb(cur, "l_out", il);149 150        // input for next layer151        inpL = cur;152    }153    cur = inpL;154 155    res->t_h_nextn = cur;156 157    if (!cparams.embeddings_nextn_masked && inp_out_ids) {158        cur = ggml_get_rows(ctx0, cur, inp_out_ids);159    }160 161    cur = build_norm(cur,162            model.output_norm, NULL,163            LLM_NORM_RMS, -1);164 165    cb(cur, "result_norm", -1);166    res->t_embd = cur;167 168    // lm_head169    cur = build_lora_mm(model.output, cur, model.output_s);170 171    cb(cur, "result_output", -1);172    res->t_logits = cur;173 174    ggml_build_forward_expand(gf, cur);175}176