echodict/llama.cpp

1#include "models.h"2 3llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5    const int64_t n_embd_head = hparams.n_embd_head_v();6 7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());8    GGML_ASSERT(n_embd_head == n_rot);9 10    ggml_tensor * cur;11    ggml_tensor * inpL;12 13    inpL = build_inp_embd(model.tok_embd);14 15    // inp_pos - contains the positions16    ggml_tensor * inp_pos = build_inp_pos();17 18    auto * inp_attn = build_attn_inp_kv();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        // norm26        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);27        cb(cur, "attn_norm", il);28 29        // self_attention30        {31            // compute Q and K and RoPE them32            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,33                    n_embd_head, n_head, n_head_kv, il);34 35            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);36            cb(Qcur, "Qcur_normed", il);37 38            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,39                                 ext_factor, attn_factor, beta_fast, beta_slow);40 41            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);42            cb(Kcur, "Kcur_normed", il);43 44            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,45                                 ext_factor, attn_factor, beta_fast, beta_slow);46 47            cb(Qcur, "Qcur", il);48            cb(Kcur, "Kcur", il);49            cb(Vcur, "Vcur", il);50 51            cur = build_attn(inp_attn,52                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,53                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);54        }55        if (il == n_layer - 1 && inp_out_ids) {56            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);57            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);58        }59        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);60        cb(ffn_inp, "ffn_inp", il);61 62        // MoE branch63        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);64        cb(cur, "ffn_norm", il);65 66        if ((uint32_t) il < hparams.n_layer_dense_lead) {67            cur = build_ffn(cur,68                    model.layers[il].ffn_up, NULL, NULL,69                    model.layers[il].ffn_gate, NULL, NULL,70                    model.layers[il].ffn_down, NULL, NULL,71                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);72            cb(cur, "ffn_out", il);73        } else {74            ggml_tensor * moe_out = build_moe_ffn(cur,75                model.layers[il].ffn_gate_inp,76                model.layers[il].ffn_up_exps,77                model.layers[il].ffn_gate_exps,78                model.layers[il].ffn_down_exps,79                model.layers[il].ffn_exp_probs_b,80                n_expert, n_expert_used,81                LLM_FFN_SILU, hparams.expert_weights_norm,82                hparams.expert_weights_scale,83                (llama_expert_gating_func_type) hparams.expert_gating_func,84                il);85            cb(moe_out, "ffn_moe_out", il);86 87            {88                ggml_tensor * ffn_shexp =89                    build_ffn(cur,90                        model.layers[il].ffn_up_shexp, NULL, NULL,91                        model.layers[il].ffn_gate_shexp, NULL, NULL,92                        model.layers[il].ffn_down_shexp, NULL, NULL,93                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);94                cb(ffn_shexp, "ffn_shexp", il);95 96                cur = ggml_add(ctx0, moe_out, ffn_shexp);97                cb(cur, "ffn_out", il);98            }99        }100        cur = ggml_add(ctx0, cur, ffn_inp);101 102        cur = build_cvec(cur, il);103        cb(cur, "l_out", il);104 105        // input for next layer106        inpL = cur;107    }108    cur = inpL;109 110    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);111 112    cb(cur, "result_norm", -1);113    res->t_embd = cur;114 115    // lm_head116    cur = build_lora_mm(model.output, cur);117 118    cb(cur, "result_output", -1);119    res->t_logits = cur;120 121    ggml_build_forward_expand(gf, cur);122}123