echodict/llama.cpp

1#include "models.h"2 3llm_build_phi2::llm_build_phi2(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 * attn_norm_output;10    ggml_tensor * ffn_output;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();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    for (int il = 0; il < n_layer; ++il) {23        attn_norm_output = build_norm(inpL,24                model.layers[il].attn_norm,25                model.layers[il].attn_norm_b,26                LLM_NORM, il);27        cb(attn_norm_output, "attn_norm", il);28 29        // self-attention30        {31            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], attn_norm_output,32                    n_embd_head, n_head, n_head_kv, il);33            Qcur = ggml_rope_ext(34                    ctx0, Qcur, inp_pos, nullptr,35                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,36                    ext_factor, attn_factor, beta_fast, beta_slow37                    );38 39            Kcur = ggml_rope_ext(40                    ctx0, Kcur, inp_pos, nullptr,41                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,42                    ext_factor, attn_factor, beta_fast, beta_slow43                    );44 45            cb(Qcur, "Qcur", il);46            cb(Kcur, "Kcur", il);47            cb(Vcur, "Vcur", il);48 49            // with phi2, we scale the Q to avoid precision issues50            // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L6651            Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));52 53            cur = build_attn(inp_attn,54                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,55                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);56        }57        if (il == n_layer - 1 && inp_out_ids) {58            cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);59            inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);60            attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);61        }62        // FF63        {64            ffn_output = build_ffn(attn_norm_output,65                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,66                    NULL,                      NULL,                        NULL,67                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,68                    NULL,69                    LLM_FFN_GELU, LLM_FFN_SEQ, il);70            cb(ffn_output, "ffn_out", il);71        }72        cur = ggml_add(ctx0, cur, ffn_output);73        cur = ggml_add(ctx0, cur, inpL);74 75        cur = build_cvec(cur, il);76        cb(cur, "l_out", il);77 78        // input for next layer79        inpL = cur;80    }81    cur = build_norm(inpL,82            model.output_norm,83            model.output_norm_b,84            LLM_NORM, -1);85 86    cb(cur, "result_norm", -1);87    res->t_embd = cur;88 89    cur = build_lora_mm(model.output, cur);90    cb(cur, "result_output_no_bias", -1);91 92    cur = ggml_add(ctx0, cur, model.output_b);93 94    cb(cur, "result_output", -1);95    res->t_logits = cur;96 97    ggml_build_forward_expand(gf, cur);98}99