echodict/llama.cpp

1#include "models.h"2 3llm_build_mistral3::llm_build_mistral3(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    // (optional) temperature tuning18    ggml_tensor * inp_attn_scale = nullptr;19    if (hparams.f_attn_temp_scale != 0.0f) {20        inp_attn_scale = build_inp_attn_scale();21    }22 23    auto * inp_attn = build_attn_inp_kv();24 25    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;26 27    ggml_tensor * inp_out_ids = build_inp_out_ids();28 29    for (int il = 0; il < n_layer; ++il) {30        ggml_tensor * inpSA = inpL;31 32        // norm33        cur = build_norm(inpL,34                model.layers[il].attn_norm, NULL,35                LLM_NORM_RMS, il);36        cb(cur, "attn_norm", il);37 38        // self-attention39        {40            // rope freq factors for llama3; may return nullptr for llama2 and other models41            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);42 43            // compute Q and K and RoPE them44            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,45                    n_embd_head, n_head, n_head_kv, il);46 47            Qcur = ggml_rope_ext(48                    ctx0, Qcur, inp_pos, rope_factors,49                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,50                    ext_factor, attn_factor, beta_fast, beta_slow51                    );52 53            Kcur = ggml_rope_ext(54                    ctx0, Kcur, inp_pos, rope_factors,55                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,56                    ext_factor, attn_factor, beta_fast, beta_slow57                    );58 59            cb(Qcur, "Qcur", il);60            cb(Kcur, "Kcur", il);61            cb(Vcur, "Vcur", il);62 63            if (inp_attn_scale) {64                // apply llama 4 temperature scaling65                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);66                cb(Qcur, "Qcur_attn_temp_scaled", il);67            }68 69            cur = build_attn(inp_attn,70                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,71                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);72            cb(cur, "attn_out", il);73        }74        if (il == n_layer - 1 && inp_out_ids) {75            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);76            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);77        }78        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);79        cb(ffn_inp, "ffn_inp", il);80 81        // feed-forward network (non-MoE)82        if (model.layers[il].ffn_gate_inp == nullptr) {83 84            cur = build_norm(ffn_inp,85                    model.layers[il].ffn_norm, NULL,86                    LLM_NORM_RMS, il);87            cb(cur, "ffn_norm", il);88 89            cur = build_ffn(cur,90                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,91                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,92                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,93                    NULL,94                    LLM_FFN_SILU, LLM_FFN_PAR, il);95            cb(cur, "ffn_out", il);96        } else {97            // MoE branch98            cur = build_norm(ffn_inp,99                    model.layers[il].ffn_norm, NULL,100                    LLM_NORM_RMS, il);101            cb(cur, "ffn_norm", il);102 103            cur = build_moe_ffn(cur,104                    model.layers[il].ffn_gate_inp,105                    model.layers[il].ffn_up_exps,106                    model.layers[il].ffn_gate_exps,107                    model.layers[il].ffn_down_exps,108                    nullptr,109                    n_expert, n_expert_used,110                    LLM_FFN_SILU, true,111                    hparams.expert_weights_scale,112                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,113                    il);114            cb(cur, "ffn_moe_out", il);115        }116        cur = ggml_add(ctx0, cur, ffn_inp);117        cb(cur, "ffn_out", il);118 119        cur = build_cvec(cur, il);120        cb(cur, "l_out", il);121 122        // input for next layer123        inpL = cur;124    }125    cur = inpL;126 127    cur = build_norm(cur,128            model.output_norm, NULL,129            LLM_NORM_RMS, -1);130 131    cb(cur, "result_norm", -1);132    res->t_embd = cur;133 134    // lm_head135    cur = build_lora_mm(model.output, cur);136 137    cb(cur, "result_output", -1);138    res->t_logits = cur;139 140    ggml_build_forward_expand(gf, cur);141}142