llama.cpp/src/models/mistral3.cpp

#include "models.h"

void llama_model_mistral3::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);

    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast,    false);
    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
    ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, false);

    hparams.f_attn_temp_offset = 0.0f;

    // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
    if (hparams.f_attn_temp_scale != 0.0f) {
        hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
        if (hparams.n_attn_temp_floor_scale == 0) {
            throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
        }
    }

    switch (hparams.n_layer()) {
        case 26: type = LLM_TYPE_3B; break;
        case 34: type = LLM_TYPE_8B; break;
        case 40: type = LLM_TYPE_14B; break;
        default: type = LLM_TYPE_UNKNOWN;
    }
}

void llama_model_mistral3::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_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);

        // optional bias tensors
        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);

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

        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
        }
        else {
            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
        }

        if (n_expert == 0) {
            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);

            // optional MLP bias
            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
        } else {
            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, n_expert}, TENSOR_NOT_REQUIRED);
            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);

            // For Granite MoE Shared
            if (hparams.n_ff_shexp > 0) {
                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
            }
        }
    }
}

std::unique_ptr<llm_graph_context> llama_model_mistral3::build_arch_graph(const llm_graph_params & params) const {
    return std::make_unique<graph>(*this, params);
}

llama_model_mistral3::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_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();

    // (optional) temperature tuning
    ggml_tensor * inp_attn_scale = nullptr;
    if (hparams.f_attn_temp_scale != 0.0f) {
        inp_attn_scale = build_inp_attn_scale();
    }

    auto * 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) {
        ggml_tensor * inpSA = inpL;

        // 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);

            Qcur = ggml_rope_ext(
                    ctx0, Qcur, inp_pos, rope_factors,
                    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, rope_factors,
                    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);

            if (inp_attn_scale) {
                // apply llama 4 temperature scaling
                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
                cb(Qcur, "Qcur_attn_temp_scaled", 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,   model.layers[il].ffn_up_s,
                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
                    NULL,
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        } else {
            // MoE branch
            cur = build_norm(ffn_inp,
                    model.layers[il].ffn_norm, NULL,
                    LLM_NORM_RMS, il);
            cb(cur, "ffn_norm", il);

            cur = 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, true,
                    hparams.expert_weights_scale,
                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                    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(cur, "ffn_moe_out", 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);
}