koboldcpp/src/models/llama4.cpp

#include "models.h"

void llama_model_llama4::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
    ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);

    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
    if (found_swa && hparams.n_swa == 0) {
        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
        hparams.n_no_rope_layer_step = hparams.n_layer(); // always use rope
    } else {
        hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
        hparams.n_swa                   = 8192;
        hparams.n_attn_temp_floor_scale = 8192;
        hparams.f_attn_temp_scale       = 0.1f;
        hparams.f_attn_temp_offset      = 1.0f;

        uint32_t swa_period = 4; // pattern: 3 chunked - 1 full
        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
        hparams.set_swa_pattern(swa_period);

        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
    }

    switch (hparams.n_expert) {
        case 0: {
            // MobileLLM (no MoE)
            switch (hparams.n_embd) {
                case 2048: type = LLM_TYPE_140M; break;
                case 4096: type = LLM_TYPE_360M; break;
                case 6144: type = LLM_TYPE_950M; break;
                default:   type = LLM_TYPE_UNKNOWN;
            }
        } break;
        case 16:  type = LLM_TYPE_17B_16E; break;
        case 128: type = LLM_TYPE_17B_128E; break;
        default:  type = LLM_TYPE_UNKNOWN;
    }

    hparams.use_kq_norm = type != LLM_TYPE_17B_128E;
}

void llama_model_llama4::load_arch_tensors(llama_model_loader &) {
    LLAMA_LOAD_LOCALS;

    if (n_expert == 0) {
        throw std::runtime_error(arch_name() + " model cannot have zero experts");
    }
    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

    // output
    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);

    // if output is NULL, init from the input tok embed
    if (output == NULL) {
        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
    }

    for (int i = 0; i < n_layer; ++i) {
        const bool is_moe_layer = hparams.n_moe_layer_step > 0 && (i + 1) % hparams.n_moe_layer_step == 0;

        auto & layer = layers[i];

        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);

        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);

        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);

        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));

        if (is_moe_layer) {
            const int64_t n_ff_exp = hparams.n_ff_exp();

            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);
            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);

            // Shared expert
            const int64_t n_ff_shexp = n_ff_exp;
            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp}, 0);
            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd    }, 0);
            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);
        } else {
            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
        }
    }
}

std::unique_ptr<llm_graph_context> llama_model_llama4::build_arch_graph(const llm_graph_params & params) const {
    if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
        return std::make_unique<graph<false>>(*this, params);
    } else {
        return std::make_unique<graph<true>>(*this, params);
    }
}

template <bool iswa>
llama_model_llama4::graph<iswa>::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_ASSERT(n_embd_head == n_rot);

    ggml_tensor * cur;
    ggml_tensor * inpL;

    inpL = build_inp_embd(model.tok_embd);

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

    // temperature tuning
    ggml_tensor * inp_attn_scale = nullptr;
    inp_attn_scale = build_inp_attn_scale();

    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
    inp_attn_type * inp_attn = nullptr;

    if constexpr (iswa) {
        inp_attn = build_attn_inp_kv_iswa();
    } else {
        inp_attn = build_attn_inp_kv();
    }

    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);

        ggml_tensor * inpSA = inpL;

        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
        const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
                              (il + 1) % hparams.n_no_rope_layer_step != 0;

        // norm
        cur = build_norm(inpL,
                model.layers[il].attn_norm, NULL,
                LLM_NORM_RMS, il);
        cb(cur, "attn_norm", il);

        // self-attention
        {
            // rope freq factors for llama3; may return nullptr for llama2 and other models
            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);

            // compute Q and K and RoPE them
            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
                    n_embd_head, n_head, n_head_kv, il);

            if (use_rope) {
                Qcur = ggml_rope_ext(
                        ctx0, Qcur, inp_pos, rope_factors,
                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                        ext_factor, attn_factor, beta_fast, beta_slow
                        );

                Kcur = ggml_rope_ext(
                        ctx0, Kcur, inp_pos, rope_factors,
                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                        ext_factor, attn_factor, beta_fast, beta_slow
                        );
            } else if (inp_attn_scale) {
                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
            }
            cb(Qcur, "Qcur", il);
            cb(Kcur, "Kcur", il);
            cb(Vcur, "Vcur", il);

            if (use_rope && hparams.use_kq_norm) {
                // Llama4TextL2Norm
                Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps);
                Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps);
                cb(Qcur, "Qcur_normed", il);
                cb(Kcur, "Kcur_normed", 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, kq_scale, il);
            cb(cur, "attn_out", il);
        }
        if (il == n_layer - 1 && inp_out_ids) {
            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
        }
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);

        // feed-forward network (non-MoE)
        if (model.layers[il].ffn_gate_inp == nullptr) {
            cur = build_norm(ffn_inp,
                    model.layers[il].ffn_norm, NULL,
                    LLM_NORM_RMS, il);
            cb(cur, "ffn_norm", il);

            cur = build_ffn(cur,
                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                    NULL,
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        } else {
            ggml_tensor * ffn_inp_normed = build_norm(ffn_inp,
                    model.layers[il].ffn_norm, NULL,
                    LLM_NORM_RMS, il);
            cb(cur, "ffn_norm", il);

            ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed,
                    model.layers[il].ffn_gate_inp,
                    model.layers[il].ffn_up_exps,
                    model.layers[il].ffn_gate_exps,
                    model.layers[il].ffn_down_exps,
                    nullptr,
                    n_expert, n_expert_used,
                    LLM_FFN_SILU, false,
                    hparams.expert_weights_scale,
                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
                    il);

            // Shared experts
            ggml_tensor * shexp_out = build_ffn(ffn_inp_normed,
                model.layers[il].ffn_up_shexp,   NULL, NULL,
                model.layers[il].ffn_gate_shexp, NULL, NULL,
                model.layers[il].ffn_down_shexp, NULL, NULL,
                NULL,
                LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(shexp_out, "ffn_moe_shexp", il);

            cur = ggml_add(ctx0, moe_out, shexp_out);
            cb(cur, "ffn_moe_out_merged", il);
        }
        cur = ggml_add(ctx0, cur, ffn_inp);
        cb(cur, "ffn_out", il);

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

        // input for next layer
        inpL = cur;
    }
    cur = inpL;

    cur = build_norm(cur,
            model.output_norm, NULL,
            LLM_NORM_RMS, -1);

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

    // lm_head
    cur = build_lora_mm(model.output, cur, model.output_s);

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

    ggml_build_forward_expand(gf, cur);
}

// Explicit template instantiations
template struct llama_model_llama4::graph<false>;
template struct llama_model_llama4::graph<true>;