llama.cpp/src/models/smallthinker.cpp

#include "models.h"

void llama_model_smallthinker::load_arch_hparams(llama_model_loader & ml) {
    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_STANDARD;
        hparams.n_swa       = 4096;
        load_swa_pattern(ml, 4, true);

        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);
    } else {
        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
        hparams.n_no_rope_layer_step = hparams.n_layer();
    }

    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);

    switch (hparams.n_layer()) {
        case 32: type = LLM_TYPE_4B;  break;
        case 52: type = LLM_TYPE_20B; break;
        default: type = LLM_TYPE_UNKNOWN;
    }
}

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

    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) {
        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_gqa, n_embd_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);

        GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for SMALLTHINKER");
        GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for SMALLTHINKER");

        // MoE branch
        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);
    }
}

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

template <bool iswa>
llama_model_smallthinker::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) {
        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 == n_layer ||
                              il % hparams.n_no_rope_layer_step != 0;

        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
        cb(probs, "ffn_moe_logits", il);

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

            if (use_rope) {
                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, 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, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
                                    ext_factor, attn_factor, beta_fast, beta_slow);
            }
            cb(Qcur, "Qcur", il);
            cb(Kcur, "Kcur", 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 && inp_out_ids) {
            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
            probs = ggml_get_rows(ctx0, probs, inp_out_ids);
        }
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);

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

        ggml_tensor * ffn_out =
            build_moe_ffn(cur,
                    nullptr,
                    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_RELU, true,
                    hparams.expert_weights_scale,
                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
                    il, probs);

        cb(ffn_out, "ffn_out", il);
        cur = ffn_out;

        cur = ggml_add(ctx0, cur, ffn_inp);

        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_smallthinker::graph<false>;
template struct llama_model_smallthinker::graph<true>;