echodict/llama.cpp

1#include "models.h"2 3llm_build_deepseek::llm_build_deepseek(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    const float kq_scale =21        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;22 23    ggml_tensor * inp_out_ids = build_inp_out_ids();24 25    for (int il = 0; il < n_layer; ++il) {26        ggml_tensor * inpSA = inpL;27 28        // norm29        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);30        cb(cur, "attn_norm", il);31 32        // self-attention33        {34            // rope freq factors for llama3; may return nullptr for llama2 and other models35            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);36 37            // compute Q and K and RoPE them38            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,39                    n_embd_head, n_head, n_head_kv, il);40 41            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,42                                 ext_factor, attn_factor, beta_fast, beta_slow);43 44            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, 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, kq_scale, 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        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);63        cb(cur, "ffn_norm", il);64 65        if ((uint32_t) il < hparams.n_layer_dense_lead) {66            cur = build_ffn(cur,67                    model.layers[il].ffn_up, NULL, NULL,68                    model.layers[il].ffn_gate, NULL, NULL,69                    model.layers[il].ffn_down, NULL, NULL,70                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);71            cb(cur, "ffn_out", il);72        } else {73            // MoE branch74            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                nullptr,80                n_expert, n_expert_used,81                LLM_FFN_SILU, false,82                hparams.expert_weights_scale,83                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,84                il);85            cb(moe_out, "ffn_moe_out", il);86 87            // FFN shared expert88            {89                ggml_tensor * ffn_shexp =90                    build_ffn(cur,91                        model.layers[il].ffn_up_shexp, NULL, NULL,92                        model.layers[il].ffn_gate_shexp, NULL, NULL,93                        model.layers[il].ffn_down_shexp, NULL, NULL,94                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);95                cb(ffn_shexp, "ffn_shexp", il);96 97                cur = ggml_add(ctx0, moe_out, ffn_shexp);98                cb(cur, "ffn_out", il);99            }100        }101        cur = ggml_add(ctx0, cur, ffn_inp);102 103        cur = build_cvec(cur, il);104        cb(cur, "l_out", il);105 106        // input for next layer107        inpL = cur;108    }109    cur = inpL;110 111    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);112 113    cb(cur, "result_norm", -1);114    res->t_embd = cur;115 116    // lm_head117    cur = build_lora_mm(model.output, cur);118 119    cb(cur, "result_output", -1);120    res->t_logits = cur;121 122    ggml_build_forward_expand(gf, cur);123}124