echodict/llama.cpp

1#include "models.h"2 3llm_build_cohere2_iswa::llm_build_cohere2_iswa(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 8    const float f_logit_scale = hparams.f_logit_scale;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    auto * inp_attn = build_attn_inp_kv_iswa();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    for (int il = 0; il < n_layer; ++il) {23        const bool is_swa = hparams.is_swa(il);24        // UNUSED:25        // const float freq_base_l  = model.get_rope_freq_base (cparams, il);26        // const float freq_scale_l = model.get_rope_freq_scale(cparams, il);27 28        // norm29        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);30        cb(cur, "attn_norm", il);31        ggml_tensor * ffn_inp = cur;32 33        // self-attention34        {35            // rope freq factors for 128k context36            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);37 38            // compute Q and K and RoPE them39            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,40                    n_embd_head, n_head, n_head_kv, il);41 42            if (is_swa) {43                Qcur = ggml_rope_ext(44                        ctx0, Qcur, inp_pos, rope_factors,45                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,46                        ext_factor, attn_factor, beta_fast, beta_slow47                        );48 49                Kcur = ggml_rope_ext(50                        ctx0, Kcur, 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 56            cb(Qcur, "Qcur", il);57            cb(Kcur, "Kcur", il);58            cb(Vcur, "Vcur", il);59 60            cur = build_attn(inp_attn,61                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,62                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);63        }64 65        if (il == n_layer - 1 && inp_out_ids) {66            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);67            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);68            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);69        }70 71        ggml_tensor * attn_out = cur;72 73        // feed-forward network74        {75            cur = build_ffn(ffn_inp,76                    model.layers[il].ffn_up, NULL, NULL,77                    model.layers[il].ffn_gate, NULL, NULL,78                    model.layers[il].ffn_down, NULL, NULL,79                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);80            cb(cur, "ffn_out", il);81        }82 83        // add together residual + FFN + self-attention84        cur = ggml_add(ctx0, cur, inpL);85        cur = ggml_add(ctx0, cur, attn_out);86 87        cur = build_cvec(cur, il);88        cb(cur, "l_out", il);89 90        // input for next layer91        inpL = cur;92    }93 94    cur = inpL;95 96    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);97 98    cb(cur, "result_norm", -1);99    res->t_embd = cur;100 101    // lm_head102    cur = build_lora_mm(model.output, cur);103 104    if (f_logit_scale) {105        cur = ggml_scale(ctx0, cur, f_logit_scale);106    }107 108    cb(cur, "result_output", -1);109    res->t_logits = cur;110 111    ggml_build_forward_expand(gf, cur);112}113