Function llm_build_kqv

llama.cpp:4764–4850

if max_alibi_bias > 0 then apply ALiBi

// if max_alibi_bias > 0 then apply ALiBi
static struct ggml_tensor * llm_build_kqv(
        struct ggml_context * ctx,
        const llama_hparams & hparams,
       const llama_kv_cache & kv,
         struct ggml_tensor * wo,
         struct ggml_tensor * wo_b,
         struct ggml_tensor * q_cur,
         struct ggml_tensor * kq_scale,
         struct ggml_tensor * kq_mask,
                    int64_t   n_ctx,
                    int32_t   n_tokens,
                    int32_t   n_kv,
                    float     max_alibi_bias,
         const llm_build_cb & cb,
                    int       il) {
    const int64_t n_embd      = hparams.n_embd;
    const int64_t n_head      = hparams.n_head;
    const int64_t n_head_kv   = hparams.n_head_kv;
    const int64_t n_embd_head = hparams.n_embd_head();
    const int64_t n_embd_gqa  = hparams.n_embd_gqa();

    struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
    cb(q, "q", il);

    struct ggml_tensor * k =
        ggml_view_3d(ctx, kv.k,
                n_embd_head, n_kv, n_head_kv,
                ggml_element_size(kv.k)*n_embd_gqa,
                ggml_element_size(kv.k)*n_embd_head,
                ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il);
    cb(k, "k", il);
    if (k_cpy != nullptr) {
        k->src[1] = k_cpy;
    }

    struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
    cb(kq, "kq", il);

    kq = ggml_scale(ctx, kq, kq_scale);
    cb(kq, "kq_scaled", il);

    if (max_alibi_bias > 0.0f) {
        // TODO: n_head or n_head_kv
        // TODO: K-shift is likely not working
        // TODO: change to ggml_add
        kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
        cb(kq, "kq_scaled_alibi", il);
    }

    kq = ggml_add(ctx, kq, kq_mask);
    cb(kq, "kq_masked", il);

    kq = ggml_soft_max(ctx, kq);
    cb(kq, "kq_soft_max", il);

    // split cached v into n_head heads
    struct ggml_tensor * v =
        ggml_view_3d(ctx, kv.v,