Skip to content

LLM inference in C/C++

C++ ggml
15.0 KB
Raw Blame
#include "models.h"

void llama_model_mellum::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_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);

    hparams.expert_weights_norm = true;
    hparams.expert_gating_func  = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false);
    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,  hparams.expert_gating_func,  false);

    if (hparams.n_swa > 0) {
        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;

        load_swa_pattern(ml, 4);

        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
        hparams.rope_freq_scale_train_swa = 1.0f;  // SWA layers use plain RoPE, no YaRN

        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
    } else {
        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
    }

    switch (hparams.n_layer()) {
        case 28: type = LLM_TYPE_12B_A2_5B; break;
        default: type = LLM_TYPE_UNKNOWN;
    }
}

void llama_model_mellum::load_arch_tensors(llama_model_loader & ml) {
    LLAMA_LOAD_LOCALS;

    const auto nf          = nextn_flags(ml);
    const int  trunk_flags = nf.trunk;
    const int  mtp_flags   = nf.mtp;

    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}, 0);

    if (n_expert == 0) {
        throw std::runtime_error("n_expert must be > 0 for Mellum");
    }
    if (n_expert_used == 0) {
        throw std::runtime_error("n_expert_used must be > 0 for Mellum");
    }

    auto load_block = [&](int i, int flags) {
        auto & layer = layers[i];

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

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

        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);

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

        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);

        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;

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

    for (int i = 0; i < n_layer; ++i) {
        load_block(i, trunk_flags);
    }

    // MTP block is a regular Mellum decoder block plus the NextN projection and norms
    for (int i = n_layer; i < n_layer_all; ++i) {
        auto & layer = layers[i];

        load_block(i, mtp_flags);

        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), {2 * n_embd, n_embd}, mtp_flags);
        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), {n_embd},             mtp_flags);
        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), {n_embd},             mtp_flags);
        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd},             mtp_flags);
    }
}

std::unique_ptr<llm_graph_context> llama_model_mellum::build_arch_graph(const llm_graph_params & params) const {
    const bool iswa = hparams.swa_type == LLAMA_SWA_TYPE_STANDARD;

    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
        if (iswa) {
            return std::make_unique<graph_mtp<true>>(*this, params);
        }
        return std::make_unique<graph_mtp<false>>(*this, params);
    }

    if (iswa) {
        return std::make_unique<graph<true>>(*this, params);
    }
    return std::make_unique<graph<false>>(*this, params);
}

template <bool iswa>
llama_model_mellum::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();

    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();
    }

    ggml_tensor * inp_out_ids = build_inp_out_ids();

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

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

        // self_attention
        {
            // 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);

            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
            cb(Qcur, "Qcur_normed", il);

            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
            cb(Kcur, "Kcur_normed", il);

            const bool is_swa = hparams.is_swa(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);
            const float ext_factor_l  = is_swa ? 0.0f : ext_factor;
            const float attn_factor_l = is_swa ? 1.0f : attn_factor;
            const float beta_fast_l   = is_swa ? 32.0f : beta_fast;
            const float beta_slow_l   = is_swa ? 1.0f  : beta_slow;

            Qcur = ggml_rope_ext(
                ctx0, Qcur, inp_pos, nullptr,
                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l
                );

            Kcur = ggml_rope_ext(
                ctx0, Kcur, inp_pos, nullptr,
                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l
                );

            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);
        }
        if (il == n_layer - 1 && crop_before_nextn(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);

        // MoE
        cur = build_norm(ffn_inp,
                model.layers[il].ffn_norm, nullptr,
                LLM_NORM_RMS, il);
        cb(cur, "ffn_norm", il);

        ggml_tensor * moe_out =
            build_moe_ffn(cur,
                    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, hparams.expert_weights_norm,
                    hparams.expert_weights_scale,
                    (llama_expert_gating_func_type) hparams.expert_gating_func,
                    il,
                    nullptr, nullptr,
                    model.layers[il].ffn_up_exps_s,
                    model.layers[il].ffn_gate_exps_s,
                    model.layers[il].ffn_down_exps_s);
        cb(moe_out, "ffn_moe_out", il);
        cur = moe_out;

        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, nullptr,
            LLM_NORM_RMS, -1);

    // post-norm hidden state feeds both the LM head and the MTP head
    cb(cur, "h_nextn", -1);
    res->t_h_nextn = cur;

    if (crop_after_nextn(inp_out_ids)) {
        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
    }

    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;
}

template <bool iswa>
llama_model_mellum::graph_mtp<iswa>::graph_mtp(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
    GGML_ASSERT(hparams.n_layer_nextn == 1 && "MELLUM MTP currently only supports a single MTP block");

    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);

    const int il = hparams.n_layer();
    const auto & layer = model.layers[il];

    GGML_ASSERT(layer.nextn.eh_proj          && "MTP block missing nextn.eh_proj");
    GGML_ASSERT(layer.nextn.enorm            && "MTP block missing nextn.enorm");
    GGML_ASSERT(layer.nextn.hnorm            && "MTP block missing nextn.hnorm");
    GGML_ASSERT(layer.nextn.shared_head_norm && "MTP block missing nextn.shared_head_norm");

    // TODO: extract in a common llm_graph_context::build_inp_embd_h()
    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd_inp(), hparams.n_embd);

    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
    ggml_set_input(inp->tokens);

    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
    ggml_set_input(inp->embd);

    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
    ggml_set_input(inp->h);
    ggml_set_name(inp->h, "mtp_h_input");

    // TODO: make static using `ggml_build_forward_select()`
    //       see llm_graph_context::build_inp_embd() for reference
    ggml_tensor * tok_embd;
    ASSERT_EMBD_OR_TOKEN(ubatch);
    if (ubatch.token) {
        tok_embd = ggml_get_rows(ctx0, model.tok_embd, inp->tokens);
    } else {
        tok_embd = inp->embd;
    }
    cb(tok_embd, "mtp_tok_embd", il);

    ggml_tensor * h_embd = inp->h;

    res->add_input(std::move(inp));

    ggml_tensor * inp_pos     = build_inp_pos();
    ggml_tensor * inp_out_ids = build_inp_out_ids();

    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();
    }

    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
    cb(e_norm, "mtp_enorm", il);

    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
    cb(h_norm, "mtp_hnorm", il);

    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0);
    cb(concat, "mtp_concat", il);

    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
    cb(cur, "mtp_eh_proj", il);

    ggml_tensor * inpSA = cur;

    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
    cb(cur, "mtp_attn_norm", il);

    // self_attention
    {
        auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il);

        Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
        cb(Qcur, "mtp_Qcur_normed", il);

        Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
        cb(Kcur, "mtp_Kcur_normed", il);

        const bool is_swa = hparams.is_swa(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);
        const float ext_factor_l  = is_swa ? 0.0f : ext_factor;
        const float attn_factor_l = is_swa ? 1.0f : attn_factor;
        const float beta_fast_l   = is_swa ? 32.0f : beta_fast;
        const float beta_slow_l   = is_swa ? 1.0f  : beta_slow;

        Qcur = ggml_rope_ext(
            ctx0, Qcur, inp_pos, nullptr,
            n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
            ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l
            );

        Kcur = ggml_rope_ext(
            ctx0, Kcur, inp_pos, nullptr,
            n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
            ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l
            );

        cb(Qcur, "mtp_Qcur", il);
        cb(Kcur, "mtp_Kcur", il);
        cb(Vcur, "mtp_Vcur", il);

        cur = build_attn(inp_attn,
                layer.wo, layer.wo_b, layer.wo_s,
                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
        cb(cur, "mtp_attn_out", il);
    }

    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
    cb(ffn_inp, "mtp_ffn_inp", il);

    cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);
    cb(cur, "mtp_ffn_norm", il);

    cur = build_moe_ffn(cur,
            layer.ffn_gate_inp,
            layer.ffn_up_exps,
            layer.ffn_gate_exps,
            layer.ffn_down_exps,
            nullptr,
            n_expert, n_expert_used,
            LLM_FFN_SILU, hparams.expert_weights_norm,
            hparams.expert_weights_scale,
            (llama_expert_gating_func_type) hparams.expert_gating_func,
            il,
            nullptr, nullptr,
            layer.ffn_up_exps_s,
            layer.ffn_gate_exps_s,
            layer.ffn_down_exps_s);
    cb(cur, "mtp_ffn_moe_out", il);

    cur = ggml_add(ctx0, cur, ffn_inp);
    cb(cur, "mtp_ffn_out", il);

    cur = build_norm(cur, layer.nextn.shared_head_norm, nullptr, LLM_NORM_RMS, -1);

    cb(cur, "h_nextn", -1);
    res->t_h_nextn = cur;

    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
    cb(cur, "mtp_shared_head_norm", -1);

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

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

template struct llama_model_mellum::graph<false>;
template struct llama_model_mellum::graph<true>;
template struct llama_model_mellum::graph_mtp<false>;
template struct llama_model_mellum::graph_mtp<true>;