echodict/llama.cpp

1#include "models.h"2 3llm_build_deci::llm_build_deci(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 =20        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;21 22    ggml_tensor * inp_out_ids = build_inp_out_ids();23 24    for (int il = 0; il < n_layer; ++il) {25        ggml_tensor * inpSA     = inpL;26        const int64_t n_head_kv = hparams.n_head_kv(il);27        const int64_t n_head    = hparams.n_head(il);28        const int64_t n_ff      = hparams.n_ff(il);29 30        if (n_head == 0) {31            // attention-free layer of Llama-3_1-Nemotron-51B32            cur = inpL;33        } else {34            // norm35            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);36            cb(cur, "attn_norm", il);37        }38        if (n_head > 0 && n_head_kv == 0) {39            // "linear attention" of Llama-3_1-Nemotron-51B40            cur = build_lora_mm(model.layers[il].wo, cur);41            cb(cur, "wo", il);42        } else if (n_head > 0) {43            // self-attention44            // rope freq factors for llama3; may return nullptr for llama2 and other models45            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);46 47            // compute Q and K and RoPE them48            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,49                    n_embd_head, n_head, n_head_kv, il);50 51            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,52                                 ext_factor, attn_factor, beta_fast, beta_slow);53 54            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,55                                 ext_factor, attn_factor, beta_fast, beta_slow);56 57            cb(Qcur, "Qcur", il);58            cb(Kcur, "Kcur", il);59            cb(Vcur, "Vcur", il);60 61            cur = build_attn(inp_attn,62                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,63                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);64        }65        if (il == n_layer - 1 && inp_out_ids) {66            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);67            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);68        }69        // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B70        if (n_ff == 0) {71            continue;72        }73        // modified to support attention-free layer of Llama-3_1-Nemotron-51B74        ggml_tensor * ffn_inp = cur;75        if (n_head > 0) {76            ffn_inp = ggml_add(ctx0, cur, inpSA);77            cb(ffn_inp, "ffn_inp", il);78        }79        // feed-forward network80        if (model.layers[il].ffn_gate_inp == nullptr) {81            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);82            cb(cur, "ffn_norm", il);83 84            cur = build_ffn(cur,85                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,86                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,87                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,88                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);89            cb(cur, "ffn_out", il);90        }91        cur = ggml_add(ctx0, cur, ffn_inp);92        cb(cur, "ffn_out", il);93 94        cur = build_cvec(cur, il);95        cb(cur, "l_out", il);96 97        // input for next layer98        inpL = cur;99    }100    cur = inpL;101 102    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);103 104    cb(cur, "result_norm", -1);105    res->t_embd = cur;106 107    // lm_head108    cur = build_lora_mm(model.output, cur);109 110    cb(cur, "result_output", -1);111    res->t_logits = cur;112 113    ggml_build_forward_expand(gf, cur);114}115