vshmidt/llama.cpp: LLM inference in C/C++

src/models/phi2.cpp
98 строк3 KB
Sigbjørn Skjæret
model : refactor bias tensor variable names (#22079)
18 апр 2026, 21:12
18 апр 2026, 21:124f02d47
#include "models.h"

llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
    const int64_t n_embd_head = hparams.n_embd_head_v();

    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

    ggml_tensor * cur;
    ggml_tensor * attn_norm_output;
    ggml_tensor * ffn_output;
    ggml_tensor * inpL;

    inpL = build_inp_embd(model.tok_embd);

    // inp_pos - contains the positions
    ggml_tensor * inp_pos = build_inp_pos();

    auto * inp_attn = build_attn_inp_kv();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        attn_norm_output = build_norm(inpL,
                model.layers[il].attn_norm,
                model.layers[il].attn_norm_b,
                LLM_NORM, il);
        cb(attn_norm_output, "attn_norm", il);

        // self-attention
        {
            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], attn_norm_output,
                    n_embd_head, n_head, n_head_kv, il);
            Qcur = ggml_rope_ext(
                    ctx0, Qcur, inp_pos, nullptr,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );

            Kcur = ggml_rope_ext(
                    ctx0, Kcur, inp_pos, nullptr,
                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                    ext_factor, attn_factor, beta_fast, beta_slow
                    );

            cb(Qcur, "Qcur", il);
            cb(Kcur, "Kcur", il);
            cb(Vcur, "Vcur", il);

            // with phi2, we scale the Q to avoid precision issues
            // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
            Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));

            cur = build_attn(inp_attn,
                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
        }
        if (il == n_layer - 1 && inp_out_ids) {
            cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
            inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
            attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
        }
        // FF
        {
            ffn_output = build_ffn(attn_norm_output,
                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                    NULL,                      NULL,                        NULL,
                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                    NULL,
                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
            cb(ffn_output, "ffn_out", il);
        }
        cur = ggml_add(ctx0, cur, ffn_output);
        cur = ggml_add(ctx0, cur, inpL);

        cur = build_cvec(cur, il);
        cb(cur, "l_out", il);

        // input for next layer
        inpL = cur;
    }
    cur = build_norm(inpL,
            model.output_norm,
            model.output_norm_b,
            LLM_NORM, -1);

    cb(cur, "result_norm", -1);
    res->t_embd = cur;

    cur = build_lora_mm(model.output, cur);
    cb(cur, "result_output_no_bias", -1);

    cur = ggml_add(ctx0, cur, model.output_b);

    cb(cur, "result_output", -1);
    res->t_logits = cur;

    ggml_build_forward_expand(gf, cur);
}