Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_qwen3::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6    switch (hparams.n_layer()) {7        case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;8        case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break;9        case 40: type = LLM_TYPE_14B; break;10        case 64: type = LLM_TYPE_32B; break;11        default: type = LLM_TYPE_UNKNOWN;12    }13}14 15void llama_model_qwen3::load_arch_tensors(llama_model_loader &) {16    LLAMA_LOAD_LOCALS;17 18    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);19 20    // output21    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);22    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);23    // if output is NULL, init from the input tok embed24    if (output == NULL) {25        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);26    }27 28    // output rerank head29    cls_out = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);30 31    for (int i = 0; i < n_layer; ++i) {32        auto & layer = layers[i];33 34        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);35 36        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);37        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);38 39        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);40        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);41 42        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);43        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);44        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);45        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);46    }47}48 49std::unique_ptr<llm_graph_context> llama_model_qwen3::build_arch_graph(const llm_graph_params & params) const {50    return std::make_unique<graph>(*this, params);51}52 53llama_model_qwen3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {54    const int64_t n_embd_head = hparams.n_embd_head_v();55 56    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());57    GGML_ASSERT(n_embd_head == n_rot);58 59    ggml_tensor * cur;60    ggml_tensor * inpL;61 62    inpL = build_inp_embd(model.tok_embd);63 64    // inp_pos - contains the positions65    ggml_tensor * inp_pos = build_inp_pos();66 67    auto * inp_attn = build_attn_inp_kv();68 69    ggml_tensor * inp_out_ids = build_inp_out_ids();70 71    for (int il = 0; il < n_layer; ++il) {72        res->t_layer_inp[il] = inpL;73 74        ggml_tensor * inpSA = inpL;75 76        // norm77        cur = build_norm(inpL,78                model.layers[il].attn_norm, NULL,79                LLM_NORM_RMS, il);80        cb(cur, "attn_norm", il);81 82        // self-attention83        {84            // compute Q and K and RoPE them85            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,86                    n_embd_head, n_head, n_head_kv, il);87 88            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);89            cb(Qcur, "Qcur_normed", il);90 91            Qcur = ggml_rope_ext(92                    ctx0, Qcur, inp_pos, nullptr,93                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,94                    ext_factor, attn_factor, beta_fast, beta_slow95                    );96 97            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);98            cb(Kcur, "Kcur_normed", il);99 100            Kcur = ggml_rope_ext(101                    ctx0, Kcur, inp_pos, nullptr,102                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,103                    ext_factor, attn_factor, beta_fast, beta_slow104                    );105 106            cb(Qcur, "Qcur", il);107            cb(Kcur, "Kcur", il);108            cb(Vcur, "Vcur", il);109 110            cur = build_attn(inp_attn,111                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,112                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);113        }114        if (il == n_layer - 1 && inp_out_ids) {115            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);116            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);117        }118        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);119        cb(ffn_inp, "ffn_inp", il);120 121        // feed-forward network122        cur = build_norm(ffn_inp,123                model.layers[il].ffn_norm, NULL,124                LLM_NORM_RMS, il);125        cb(cur, "ffn_norm", il);126 127        cur = build_ffn(cur,128                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,129                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,130                model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,131                NULL,132                LLM_FFN_SILU, LLM_FFN_PAR, il);133        cb(cur, "ffn_out", il);134 135        cur = ggml_add(ctx0, cur, ffn_inp);136 137        cur = build_cvec(cur, il);138        cb(cur, "l_out", il);139 140        // input for next layer141        inpL = cur;142    }143    cur = inpL;144 145    cur = build_norm(cur,146            model.output_norm, NULL,147            LLM_NORM_RMS, -1);148 149    cb(cur, "result_norm", -1);150    res->t_embd = cur;151 152    // lm_head153    cur = build_lora_mm(model.output, cur, model.output_s);154 155    cb(cur, "result_output", -1);156    res->t_logits = cur;157 158    ggml_build_forward_expand(gf, cur);159}160