echodict/llama.cpp

1#include "models.h"2 3template <bool embed>4llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_graph_params & params) : 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    using inp_attn_type = std::conditional_t<embed, llm_graph_input_attn_no_cache, llm_graph_input_attn_kv>;19 20    inp_attn_type * inp_attn = nullptr;21    if constexpr (embed) {22        inp_attn = build_attn_inp_no_cache();23    } else {24        inp_attn = build_attn_inp_kv();25    }26 27    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;28 29    ggml_tensor * inp_out_ids = build_inp_out_ids();30 31    for (int il = 0; il < n_layer; ++il) {32        ggml_tensor * inpSA = inpL;33 34        // norm35        cur = build_norm(inpL,36                model.layers[il].attn_norm, NULL,37                LLM_NORM_RMS, il);38        cb(cur, "attn_norm", il);39 40        // self-attention41        {42            // rope freq factors for llama3; may return nullptr for llama2 and other models43            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);44 45            // compute Q and K and RoPE them46            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,47                    n_embd_head, n_head, n_head_kv, il);48 49            Qcur = ggml_rope_ext(50                    ctx0, Qcur, inp_pos, rope_factors,51                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,52                    ext_factor, attn_factor, beta_fast, beta_slow53                    );54 55            Kcur = ggml_rope_ext(56                    ctx0, Kcur, inp_pos, rope_factors,57                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,58                    ext_factor, attn_factor, beta_fast, beta_slow59                    );60 61            cb(Qcur, "Qcur", il);62            cb(Kcur, "Kcur", il);63            cb(Vcur, "Vcur", il);64 65            if (hparams.use_kq_norm) {66                // Llama4TextL2Norm67                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);68                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);69                cb(Qcur, "Qcur_normed", il);70                cb(Kcur, "Kcur_normed", il);71            }72            cur = build_attn(inp_attn,73                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,74                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);75            if (model.layers[il].wo_s) {76                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);77            }78            cb(cur, "attn_out", il);79        }80        if (il == n_layer - 1 && inp_out_ids) {81            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);82            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);83        }84        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);85        cb(ffn_inp, "ffn_inp", il);86 87        // feed-forward network (non-MoE)88        if (model.layers[il].ffn_gate_inp == nullptr) {89 90            cur = build_norm(ffn_inp,91                    model.layers[il].ffn_norm, NULL,92                    LLM_NORM_RMS, il);93            cb(cur, "ffn_norm", il);94 95            cur = build_ffn(cur,96                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,97                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,98                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,99                    NULL,100                    LLM_FFN_SILU, LLM_FFN_PAR, il);101            cb(cur, "ffn_out", il);102        } else {103            // MoE branch104            cur = build_norm(ffn_inp,105                    model.layers[il].ffn_norm, NULL,106                    LLM_NORM_RMS, il);107            cb(cur, "ffn_norm", il);108 109            cur = build_moe_ffn(cur,110                    model.layers[il].ffn_gate_inp,111                    model.layers[il].ffn_up_exps,112                    model.layers[il].ffn_gate_exps,113                    model.layers[il].ffn_down_exps,114                    nullptr,115                    n_expert, n_expert_used,116                    LLM_FFN_SILU, true,117                    hparams.expert_weights_scale,118                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,119                    il,120                    nullptr, nullptr,121                    model.layers[il].ffn_up_exps_s,122                    model.layers[il].ffn_gate_exps_s,123                    model.layers[il].ffn_down_exps_s);124            cb(cur, "ffn_moe_out", il);125        }126        cur = ggml_add(ctx0, cur, ffn_inp);127        cb(cur, "ffn_out", il);128 129        cur = build_cvec(cur, il);130        cb(cur, "l_out", il);131 132        // input for next layer133        inpL = cur;134    }135    cur = inpL;136 137    cur = build_norm(cur,138            model.output_norm, NULL,139            LLM_NORM_RMS, -1);140 141    cb(cur, "result_norm", -1);142    res->t_embd = cur;143 144    if constexpr (!embed) {145        // lm_head146        cur = build_lora_mm(model.output, cur);147 148        cb(cur, "result_output", -1);149        res->t_logits = cur;150    }151 152    ggml_build_forward_expand(gf, cur);153}154 155template struct llm_build_llama<false>;156template struct llm_build_llama<true>;157