koboldcpp/src/models/granite-swa.cpp

#include "models.h"

#include <sstream>

void llama_model_granite_swa::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_LOGIT_SCALE,                 hparams.f_logit_scale);
    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);

    // MoE expert configuration
    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert, false);
    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used, false);

     // iSWA configuration
    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;

    // Granite4 Vision uses array deepstack_mapping
    ml.get_arr(LLM_KV_DEEPSTACK_MAPPING, hparams.deepstack_mapping_arr, false);

    // Count the unique deepstack input indices
    std::unordered_set<uint32_t> unique_deepstack_idxs;
    for (const auto val : hparams.deepstack_mapping_arr) {
        if (val >= 0) {
            unique_deepstack_idxs.insert(val);
        }
    }
    hparams.n_deepstack_layers = unique_deepstack_idxs.size();

    // Ensure all values are valid (avoid overflow attacks)
    for (const auto val : unique_deepstack_idxs) {
        if (val > hparams.n_deepstack_layers) {
            std::stringstream ss;
            ss << "Invalid deepstack index: " << val << " > " << hparams.n_deepstack_layers;
            throw std::runtime_error(ss.str());
        }
    }

    // Per-layer RoPE pattern (optional)
    ml.get_arr(LLM_KV_ATTENTION_ROPE_PATTERN, hparams.rope_pattern, false);

    switch (hparams.n_layer()) {
        case 32: type = LLM_TYPE_3B; break;
        case 40: type = LLM_TYPE_3B; break;
        // Add additional layer/vocab/etc checks here for other model sizes
        default: type = LLM_TYPE_UNKNOWN;
    }

    // For Granite MoE Shared
    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
}

void llama_model_granite_swa::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);

        // Per-layer attention sinks for iSWA
        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);

        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_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
            create_tensor_gate_up_exps(layer, i, n_embd, n_ff, n_expert, 0);

            // For Granite MoE Shared - gate+up kept fused in ffn_up_shexp (see LLM_FFN_SWIGLU below)
            if (hparams.n_ff_shexp > 0) {
                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, 2*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_granite_swa::build_arch_graph(const llm_graph_params & params) const {
    return std::make_unique<graph>(*this, params);
}

llama_model_granite_swa::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 - built only if rope enabled
    ggml_tensor * inp_pos = build_inp_pos();
    auto * inp_attn = build_attn_inp_kv_iswa();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {

        // Granite Vision 4.1 deepstack: inject the projector stream that
        // targets decoder layer `il` before the decoder runs.
        // NOTE: skip the first deepstack layer since that's inpL
        const auto & deepstack_emb_idx = hparams.deepstack_mapping_arr[il];
        if (il > 0 && deepstack_emb_idx >= 0) {
            ggml_tensor * ds = ggml_view_2d(ctx0,
                res->t_inp_embd, n_embd, n_tokens,
                res->t_inp_embd->nb[1],
                deepstack_emb_idx * n_embd * sizeof(float));
            inpL = ggml_add(ctx0, inpL, ds);
            cb(inpL, "deepstack_in", 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
        cur = build_attention_layer(
            cur, inp_pos, inp_attn,
            model, 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);
        }
        // ffn
        cur = build_layer_ffn(cur, inpSA, model, 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);

    // For Granite architectures - scale logits
    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
    cb(cur, "result_output", -1);
    res->t_logits = cur;

    ggml_build_forward_expand(gf, cur);
}

ggml_tensor * llama_model_granite_swa::graph::build_attention_layer(
          ggml_tensor                  * cur,
          ggml_tensor                  * inp_pos,
          llm_graph_input_attn_kv_iswa * inp_attn,
    const llama_model                  & model,
    const int64_t                        n_embd_head,
    const int                            il) {

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

    const bool use_rope = hparams.has_rope(il);
    if (use_rope) {
        ggml_tensor * rope_factors = model.get_rope_factors(cparams, 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);

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

    // Pass layer.attn_sinks to build_attn for sink-based attention modulation
    cur = build_attn(inp_attn,
            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
            Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, kq_scale, il);
    cb(cur, "attn_out", il);
    return cur;
}

ggml_tensor * llama_model_granite_swa::graph::build_layer_ffn(
          ggml_tensor * cur,
          ggml_tensor * inpSA,
    const llama_model & model,
    const int           il) {

    // For Granite architectures - scale residual
    if (hparams.f_residual_scale) {
        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
    }
    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 {
        // MoE branch
        cur = 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(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, model.layers[il].ffn_gate_up_exps);
        cb(moe_out, "ffn_moe_out", il);

        // For Granite MoE Shared - gate+up kept fused in ffn_up_shexp
        if (hparams.n_ff_shexp > 0) {
            ggml_tensor * ffn_shexp = build_ffn(cur,
                model.layers[il].ffn_up_shexp,   NULL, NULL,
                NULL,                            NULL, NULL,
                model.layers[il].ffn_down_shexp, NULL, NULL,
                NULL,
                LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
            cb(ffn_shexp, "ffn_shexp", il);

            cur = ggml_add(ctx0, moe_out, ffn_shexp);
            cb(cur, "ffn_out", il);
        } else {
            cur = moe_out;
        }
    }

    // For Granite architectures - scale residual
    if (hparams.f_residual_scale) {
        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
    }
    cur = ggml_add(ctx0, cur, ffn_inp);
    cb(cur, "ffn_out", il);

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

    return cur;
}