echodict/llama.cpp

1#include "models.h"2 3llm_build_stablelm::llm_build_stablelm(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    ggml_tensor * cur;9    ggml_tensor * inpL;10 11    inpL = build_inp_embd(model.tok_embd);12 13    // inp_pos - contains the positions14    ggml_tensor * inp_pos = build_inp_pos();15 16    auto * inp_attn = build_attn_inp_kv();17 18    ggml_tensor * inp_out_ids = build_inp_out_ids();19 20    for (int il = 0; il < n_layer; ++il) {21        // norm22        cur = build_norm(inpL,23                model.layers[il].attn_norm,24                model.layers[il].attn_norm_b,25                LLM_NORM, il);26        cb(cur, "attn_norm", il);27 28        ggml_tensor * inpSA = cur;29 30        // self-attention31        {32            // compute Q and K and RoPE them33            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,34                    n_embd_head, n_head, n_head_kv, il);35 36            if (model.layers[il].attn_q_norm) {37                Qcur = build_norm(Qcur,38                        model.layers[il].attn_q_norm,39                        NULL,40                        LLM_NORM, il);41                cb(Qcur, "Qcur", il);42            }43            if (model.layers[il].attn_k_norm) {44                Kcur = build_norm(Kcur,45                        model.layers[il].attn_k_norm,46                        NULL,47                        LLM_NORM, il);48                cb(Kcur, "Kcur", il);49            }50 51            Qcur = ggml_rope_ext(52                    ctx0, Qcur, inp_pos, nullptr,53                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,54                    ext_factor, attn_factor, beta_fast, beta_slow55                    );56 57            Kcur = ggml_rope_ext(58                    ctx0, Kcur, inp_pos, nullptr,59                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,60                    ext_factor, attn_factor, beta_fast, beta_slow61                    );62 63            cb(Qcur, "Qcur", il);64            cb(Kcur, "Kcur", il);65            cb(Vcur, "Vcur", il);66 67            cur = build_attn(inp_attn,68                    model.layers[il].wo, NULL, model.layers[il].wo_s,69                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);70        }71        if (il == n_layer - 1 && inp_out_ids) {72            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);73            inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);74            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);75        }76        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);77        cb(ffn_inp, "ffn_inp", il);78 79        // feed-forward network80        {81            if (model.layers[il].ffn_norm) {82                cur = build_norm(ffn_inp,83                        model.layers[il].ffn_norm,84                        model.layers[il].ffn_norm_b,85                        LLM_NORM, il);86                cb(cur, "ffn_norm", il);87            } else {88                // parallel residual89                cur = inpSA;90            }91            cur = build_ffn(cur,92                    model.layers[il].ffn_up,   NULL, NULL,93                    model.layers[il].ffn_gate, NULL, NULL,94                    model.layers[il].ffn_down, NULL, NULL,95                    NULL,96                    LLM_FFN_SILU, LLM_FFN_PAR, il);97            cb(cur, "ffn_out", il);98        }99        cur = ggml_add(ctx0, cur, ffn_inp);100 101        cur = build_cvec(cur, il);102        cb(cur, "l_out", il);103 104        // input for next layer105        inpL = cur;106    }107    cur = inpL;108 109    cur = build_norm(cur,110            model.output_norm,111            model.output_norm_b,112            LLM_NORM, -1);113 114    cb(cur, "result_norm", -1);115    res->t_embd = cur;116 117    // lm_head118    cur = build_lora_mm(model.output, cur);119 120    cb(cur, "result_output", -1);121    res->t_logits = cur;122 123    ggml_build_forward_expand(gf, cur);124}125