echodict/llama.cpp

1#include "models.h"2 3llm_build_grok::llm_build_grok(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    // inp_pos - contains the positions15    ggml_tensor * inp_pos = build_inp_pos();16 17    auto * inp_attn = build_attn_inp_kv();18 19    ggml_tensor * inp_out_ids = build_inp_out_ids();20 21    for (int il = 0; il < n_layer; ++il) {22        ggml_tensor * inpSA = inpL;23 24        // norm25        cur = build_norm(inpL,26                model.layers[il].attn_norm, NULL,27                LLM_NORM_RMS, il);28        cb(cur, "attn_norm", il);29 30        // self-attention31        {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 = ggml_rope_ext(37                    ctx0, Qcur, inp_pos, nullptr,38                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,39                    ext_factor, attn_factor, beta_fast, beta_slow40                    );41 42            Kcur = ggml_rope_ext(43                    ctx0, Kcur, inp_pos, nullptr,44                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,45                    ext_factor, attn_factor, beta_fast, beta_slow46                    );47 48            cb(Qcur, "Qcur", il);49            cb(Kcur, "Kcur", il);50            cb(Vcur, "Vcur", 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, 1.0f, il);55        }56        if (il == n_layer - 1 && inp_out_ids) {57            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);58            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);59        }60        cur = build_norm(cur,61                model.layers[il].attn_out_norm, NULL,62                LLM_NORM_RMS, il);63        cb(cur, "attn_out_norm", il);64 65        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);66        cb(ffn_inp, "ffn_inp", il);67 68        // feed-forward network69        cur = build_norm(ffn_inp,70                model.layers[il].ffn_norm, NULL,71                LLM_NORM_RMS, il);72        cb(cur, "ffn_norm", il);73 74        // MoE branch75        ggml_tensor * moe_out = build_moe_ffn(cur,76                model.layers[il].ffn_gate_inp,77                model.layers[il].ffn_up_exps,78                model.layers[il].ffn_gate_exps,79                model.layers[il].ffn_down_exps,80                nullptr,81                n_expert, n_expert_used,82                LLM_FFN_GELU, true,83                hparams.expert_weights_scale,84                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,85                il);86        cb(moe_out, "ffn_moe_out", il);87 88        if (model.layers[il].ffn_up) {89            ggml_tensor * ffn_out = build_ffn(cur,90                    model.layers[il].ffn_up,   NULL, NULL,91                    model.layers[il].ffn_gate, NULL, NULL,92                    model.layers[il].ffn_down, NULL, NULL,93                    NULL,94                    LLM_FFN_GELU, LLM_FFN_PAR, il);95            cb(ffn_out, "ffn_out", il);96 97            cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);98            cb(cur, "ffn_out", il);99        } else {100            cur = moe_out;101        }102        cur = build_norm(cur,103                model.layers[il].ffn_post_norm, NULL,104                LLM_NORM_RMS, il);105        cb(cur, "ffn_post_norm", il);106 107        cur = ggml_add(ctx0, cur, ffn_inp);108        cb(cur, "ffn_out", il);109 110        cur = build_cvec(cur, il);111        cb(cur, "l_out", il);112 113        // input for next layer114        inpL = cur;115    }116    cur = inpL;117 118    cur = build_norm(cur,119            model.output_norm, NULL,120            LLM_NORM_RMS, -1);121 122    cb(cur, "result_norm", -1);123    res->t_embd = cur;124 125    // lm_head126    cur = build_lora_mm(model.output, cur);127 128    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);129 130    // final logit soft-capping131    if (hparams.f_final_logit_softcapping) {132        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);133        cur = ggml_tanh(ctx0, cur);134        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);135    }136    cb(cur, "result_output", -1);137    res->t_logits = cur;138 139    ggml_build_forward_expand(gf, cur);140}141