echodict/llama.cpp

1#include "models.h"2 3 4llm_build_bitnet::llm_build_bitnet(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 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    ggml_tensor * inp_out_ids = build_inp_out_ids();20 21    for (int il = 0; il < n_layer; ++il) {22        ggml_tensor * inpSA = inpL;23 24        cur = build_norm(inpL,25                model.layers[il].attn_norm, NULL,26                LLM_NORM_RMS, il);27        cb(cur, "attn_norm", il);28 29        // self-attention30        {31            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,32                    n_embd_head, n_head, n_head_kv, il);33 34            Qcur = ggml_rope_ext(35                    ctx0, Qcur, inp_pos, nullptr,36                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,37                    ext_factor, attn_factor, beta_fast, beta_slow38                    );39 40            Kcur = ggml_rope_ext(41                    ctx0, Kcur, inp_pos, nullptr,42                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,43                    ext_factor, attn_factor, beta_fast, beta_slow44                    );45 46            cb(Qcur, "Qcur", il);47            cb(Kcur, "Kcur", il);48            cb(Vcur, "Vcur", il);49 50            cur = build_attn(inp_attn,51                    NULL, NULL, NULL,52                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);53 54            cur = build_norm(cur,55                    model.layers[il].attn_sub_norm, NULL,56                    LLM_NORM_RMS, il);57            cb(cur, "attn_sub_norm", il);58 59            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);60            if (model.layers[il].wo_b) {61                cur = ggml_add(ctx0, cur, model.layers[il].wo_b);62            }63            cb(cur, "attn_out", il);64        }65 66        if (il == n_layer - 1 && inp_out_ids) {67            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);68            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);69        }70 71        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);72        cb(ffn_inp, "ffn_inp", il);73 74        // feed-forward forward75        cur = build_norm(ffn_inp,76                model.layers[il].ffn_norm, NULL,77                LLM_NORM_RMS, il);78        cb(cur, "ffn_norm", il);79 80        cur = build_ffn(cur,81                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,82                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,83                NULL,                      NULL, NULL,84                NULL,85                LLM_FFN_SILU, LLM_FFN_PAR, il);86        cb(cur, "ffn_sub_out", il);87 88        cur = build_norm(cur,89                model.layers[il].ffn_sub_norm, NULL,90                LLM_NORM_RMS, il);91        cb(cur, "ffn_sub_norm", il);92 93        cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);94        cb(cur, "ffn_down", il);95 96        cur = ggml_add(ctx0, cur, ffn_inp);97        cb(cur, "l_out", il);98 99        cur = build_cvec(cur, il);100        cb(cur, "l_out", il);101 102        // input for next layer103        inpL = cur;104    }105 106    cur = inpL;107 108    cur = build_norm(cur,109            model.output_norm, NULL,110            LLM_NORM_RMS, -1);111 112    cb(cur, "result_norm", -1);113    res->t_embd = cur;114 115    // lm_head116    // FIXME: do not use model.tok_embd directly, duplicate as model.output117    cur = build_lora_mm(model.tok_embd, cur);118 119    cb(cur, "result_output", -1);120    res->t_logits = cur;121 122    ggml_build_forward_expand(gf, cur);123}124