echodict/llama.cpp

1#include "models.h"2 3llm_build_hunyuan_moe::llm_build_hunyuan_moe(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    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));20 21    ggml_tensor * inp_out_ids = build_inp_out_ids();22 23    for (int il = 0; il < n_layer; ++il) {24        ggml_tensor * inpSA = inpL;25 26        // norm27        cur = build_norm(inpL,28                model.layers[il].attn_norm, NULL,29                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(42                    ctx0, Qcur, inp_pos, rope_factors,43                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,44                    ext_factor, attn_factor, beta_fast, beta_slow45                    );46 47            cb(Qcur, "Qcur", il);48            cb(Kcur, "Kcur", il);49            cb(Vcur, "Vcur", il);50 51            Kcur = ggml_rope_ext(52                    ctx0, Kcur, inp_pos, rope_factors,53                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,54                    ext_factor, attn_factor, beta_fast, beta_slow55                    );56 57            Kcur = build_norm(Kcur,58                    model.layers[il].attn_k_norm, nullptr,59                    LLM_NORM_RMS, il);60            cb(Kcur, "Kcur_norm", il);61 62            Qcur = build_norm(Qcur,63                    model.layers[il].attn_q_norm, nullptr,64                    LLM_NORM_RMS, il);65            cb(Qcur, "Qcur_norm", il);66 67            cur = build_attn(inp_attn,68                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,69                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);70            cb(cur, "attn_out", il);71        }72        if (il == n_layer - 1 && inp_out_ids) {73            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);74            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);75        }76        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);77        cb(ffn_inp, "ffn_inp", il);78 79        cur = build_norm(ffn_inp,80            model.layers[il].ffn_norm, NULL,81            LLM_NORM_RMS, il);82        cb(cur, "ffn_norm", il);83 84        // feed-forward network (non-MoE)85        ggml_tensor * cur_mlp = build_ffn(cur,86                model.layers[il].ffn_up_shexp,   NULL, NULL,87                model.layers[il].ffn_gate_shexp, NULL, NULL,88                model.layers[il].ffn_down_shexp, NULL, NULL,89                NULL,90                LLM_FFN_SILU, LLM_FFN_PAR, il);91        cb(cur_mlp, "ffn_mlp", il);92 93        // MoE branch94        ggml_tensor * cur_moe = build_moe_ffn(cur,95                model.layers[il].ffn_gate_inp,96                model.layers[il].ffn_up_exps,97                model.layers[il].ffn_gate_exps,98                model.layers[il].ffn_down_exps,99                nullptr,100                n_expert, n_expert_used,101                LLM_FFN_SILU,102                true, // norm_topk_prob103                hparams.expert_weights_scale,104                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,105                il);106        cb(cur_moe, "ffn_moe_out", il);107 108        ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);109        cb(ffn_out, "ffn_out", il);110 111        cur = ggml_add(ctx0, ffn_out, ffn_inp);112 113        cur = build_cvec(cur, il);114        cb(cur, "l_out", il);115 116        // input for next layer117        inpL = cur;118    }119    cur = inpL;120 121    cur = build_norm(cur,122            model.output_norm, NULL,123            LLM_NORM_RMS, -1);124 125    cb(cur, "result_norm", -1);126    res->t_embd = cur;127 128    // lm_head129    cur = build_lora_mm(model.output, cur);130    cb(cur, "result_output", -1);131    res->t_logits = cur;132 133    ggml_build_forward_expand(gf, cur);134}135