echodict/llama.cpp

1#include "models.h"2 3llm_build_arcee::llm_build_arcee(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 = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;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            Kcur = ggml_rope_ext(48                    ctx0, Kcur, 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            cb(Qcur, "Qcur", il);54            cb(Kcur, "Kcur", il);55            cb(Vcur, "Vcur", il);56 57            cur = build_attn(inp_attn,58                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,59                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);60            cb(cur, "attn_out", il);61        }62 63        if (il == n_layer - 1 && inp_out_ids) {64            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);65            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);66        }67 68        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);69        cb(ffn_inp, "ffn_inp", il);70 71        // feed-forward network72        // ARCEE uses relu^2 instead of silu73        cur = build_norm(ffn_inp,74                model.layers[il].ffn_norm, NULL,75                LLM_NORM_RMS, il);76        cb(cur, "ffn_norm", il);77 78        cur = build_ffn(cur,79                model.layers[il].ffn_up,   NULL, NULL,80                NULL,                      NULL, NULL,81                model.layers[il].ffn_down, NULL, NULL,82                NULL,83                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);84        cb(cur, "ffn_out", il);85 86        cur = ggml_add(ctx0, cur, ffn_inp);87        cb(cur, "ffn_out", il);88 89        cur = build_cvec(cur, il);90        cb(cur, "l_out", il);91 92        // input for next layer93        inpL = cur;94    }95 96    cur = inpL;97 98    cur = build_norm(cur,99            model.output_norm, NULL,100            LLM_NORM_RMS, -1);101 102    cb(cur, "result_norm", -1);103    res->t_embd = cur;104 105    // lm_head106    cur = build_lora_mm(model.output, cur);107 108    cb(cur, "result_output", -1);109    res->t_logits = cur;110 111    ggml_build_forward_expand(gf, cur);112}113