llama.cpp/src/models/bert.cpp

#include "models.h"

void llama_model_bert::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);

    switch (hparams.n_layer()) {
        case 3:
            type = LLM_TYPE_17M; break; // bge-micro
        case 6:
            type = LLM_TYPE_22M; break; // MiniLM-L6
        case 12:
            switch (hparams.n_embd) {
                case 384: type = LLM_TYPE_33M; break; // MiniLM-L12, bge-small
                case 768: type = LLM_TYPE_109M; break; // bge-base
                default: type = LLM_TYPE_UNKNOWN;
            } break;
        case 24:
            type = LLM_TYPE_335M; break; // bge-large
        default: type = LLM_TYPE_UNKNOWN;
    }
}

void llama_model_bert::load_arch_tensors(llama_model_loader &) {
    LLAMA_LOAD_LOCALS;

    if (n_token_types == 0) {
        throw std::runtime_error(arch_name() + " model needs to define token type count");
    }
    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);

    if (arch == LLM_ARCH_BERT) {
        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);

        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);

        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
    }

    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);

    for (int i = 0; i < n_layer; ++i) {
        auto & layer = layers[i];

        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);

        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);

        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);

        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
        } else {
            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);

            if (arch == LLM_ARCH_NOMIC_BERT) {
                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
            }
        }

        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
    }
}

std::unique_ptr<llm_graph_context> llama_model_bert::build_arch_graph(const llm_graph_params & params) const {
    return std::make_unique<graph>(*this, params);
}

llama_model_bert::graph::graph(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 * inpL;
    ggml_tensor * inp_pos = nullptr;

    if (model.arch != LLM_ARCH_JINA_BERT_V2) {
        inp_pos = build_inp_pos();
    }

    // construct input embeddings (token, type, position)
    inpL = build_inp_embd(model.tok_embd);

    // token types are hardcoded to zero ("Sentence A")
    if (model.type_embd) {
        ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
        inpL                    = ggml_add(ctx0, inpL, type_row0);
    }
    if (model.arch == LLM_ARCH_BERT) {
        inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
    }
    cb(inpL, "inp_embd", -1);

    // embed layer norm
    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
    cb(inpL, "inp_norm", 0);

    auto * inp_attn = build_attn_inp_no_cache();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        ggml_tensor * cur = inpL;

        {
            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
                    n_embd_head, n_head, n_head_kv, il);

            if (model.layers[il].attn_q_norm) {
                Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);

                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);

                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
            }

            if (model.layers[il].attn_k_norm) {
                Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);

                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);

                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
            }

            // RoPE
            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
                model.arch == LLM_ARCH_JINA_BERT_V3) {
                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);

            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 / sqrtf(float(n_embd_head)), il);
            cb(cur, "kqv_out", 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);
        }

        // re-add the layer input
        cur = ggml_add(ctx0, cur, inpL);

        // attention layer norm
        cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);

        if (model.layers[il].attn_norm_2 != nullptr) {
            cur = ggml_add(ctx0, cur, inpL);  // re-add the layer input
            cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
        }

        ggml_tensor * ffn_inp = cur;
        cb(ffn_inp, "ffn_inp", il);

        // feed-forward network
        if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
            // MoE branch
            cur = build_moe_ffn(cur,
                    model.layers[il].ffn_gate_inp,
                    model.layers[il].ffn_up_exps,
                    nullptr,
                    model.layers[il].ffn_down_exps,
                    nullptr,
                    hparams.n_expert, hparams.n_expert_used,
                    LLM_FFN_GELU, false,
                    hparams.expert_weights_scale,
                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                    il);
            cb(cur, "ffn_moe_out", il);
        } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
                   model.arch == LLM_ARCH_JINA_BERT_V3) {
            cur = build_ffn(cur,
                    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(cur, "ffn_out", il);
        } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
            const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();
            auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;
            cur = build_ffn(cur,
                    model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
                    model.layers[il].ffn_gate, NULL, NULL,
                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
                    type_op, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        } else {
            cur = build_ffn(cur,
                model.layers[il].ffn_up, NULL, NULL,
                model.layers[il].ffn_gate, NULL, NULL,
                model.layers[il].ffn_down, NULL, NULL,
                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        }

        // attentions bypass the intermediate layer
        cur = ggml_add(ctx0, cur, ffn_inp);

        // output layer norm
        cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);

        // input for next layer
        inpL = cur;
    }

    cur = inpL;

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

    ggml_build_forward_expand(gf, cur);
}