echodict/llama.cpp

1#include "models.h"2 3llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {4    const int64_t n_embd_head = hparams.n_embd_head_v();5 6    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());7    GGML_ASSERT(n_embd_head == n_rot);8 9    ggml_tensor * cur;10    ggml_tensor * inpL;11 12    inpL = build_inp_embd(model.tok_embd);13 14    ggml_tensor * inp_pos  = build_inp_pos();15    auto *        inp_attn = build_attn_inp_kv();16 17    const float kq_scale =18        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    for (int il = 0; il < n_layer; ++il) {23        ggml_tensor * inpSA = inpL;24 25        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);26        cb(cur, "attn_norm", il);27 28        // self-attention29        {30            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);31 32            // compute Q and K and RoPE them33            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,34                    n_embd_head, n_head, n_head_kv, il);35 36            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);37            cb(Qcur, "Qcur_normed", il);38 39            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);40            cb(Kcur, "Kcur_normed", il);41 42            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,43                                 ext_factor, attn_factor, beta_fast, beta_slow);44 45            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,46                                 ext_factor, attn_factor, beta_fast, beta_slow);47 48            cb(Qcur, "Qcur_pos", il);49            cb(Kcur, "Kcur_pos", il);50            cb(Vcur, "Vcur_pos", il);51 52            cur = build_attn(inp_attn,53                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,54                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);55            cb(cur, "attn_out", il);56        }57 58        if (il == n_layer - 1 && inp_out_ids) {59            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);60            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);61        }62 63        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);64        cb(ffn_inp, "ffn_inp", il);65 66        // feed-forward network with xIELU activation67        {68            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);69            cb(cur, "ffn_norm", il);70 71            // Up projection72            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);73            cb(up, "ffn_up", il);74 75            float alpha_n_val = hparams.xielu_alpha_n[il];76            float alpha_p_val = hparams.xielu_alpha_p[il];77            float beta_val    = hparams.xielu_beta[il];78            float eps_val     = hparams.xielu_eps[il];79 80            // Apply xIELU activation81            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);82            cb(activated, "ffn_xielu", il);83 84            // Down projection85            cur = build_lora_mm(model.layers[il].ffn_down, activated);86            cb(cur, "ffn_down", il);87        }88 89        cur = ggml_add(ctx0, cur, ffn_inp);90        cb(cur, "ffn_out", il);91 92        cur = build_cvec(cur, il);93        cb(cur, "l_out", il);94 95        // input for next layer96        inpL = cur;97    }98 99    cur = inpL;100 101    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);102 103    cb(cur, "result_norm", -1);104    res->t_embd = cur;105 106    // lm_head107    cur = build_lora_mm(model.output, cur);108 109    cb(cur, "result_output", -1);110    res->t_logits = cur;111 112    ggml_build_forward_expand(gf, cur);113}114