backend/llama.cpp/src/models/glm4.cpp

#include "models.h"

void llama_model_glm4::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_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);

    // NextN/MTP parameters (GLM-OCR)
    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_impl");

    switch (hparams.n_layer()) {
        case 17: type = LLM_TYPE_1B; break; // GLM-OCR
        case 40: type = LLM_TYPE_9B; break;
        case 61: type = LLM_TYPE_32B; break;
        default: type = LLM_TYPE_UNKNOWN;
    }
}

void llama_model_glm4::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_all; ++i) {
        int flags = 0;
        if (i >= n_layer) {
            // skip all tensors in the NextN layers
            flags |= TENSOR_SKIP;
        }

        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_k_gqa, n_embd_v_gqa, flags);

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

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

        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, flags);

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

        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
        if (i >= n_layer) {
            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);

            // Optional tensors
            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
        }
    }
}

std::unique_ptr<llm_graph_context> llama_model_glm4::build_arch_graph(const llm_graph_params & params) const {
    return std::make_unique<graph>(*this, params);
}

llama_model_glm4::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());

    int sections[4];
    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);

    ggml_tensor * cur;
    ggml_tensor * inpL;

    inpL = build_inp_embd(model.tok_embd);

    bool use_mrope = hparams.use_mrope();
    if (ubatch.embd && !use_mrope) {
        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
    }

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

    auto * inp_attn = build_attn_inp_kv();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    // Only process up to last layer (skip final NextN layer)
    // Final layer tensors are loaded but not processed in forward pass
    for (int il = 0; il < n_layer; ++il) {
        ggml_tensor * inpSA = inpL;

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

        // self-attention
        {
            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
                    n_embd_head, n_head, n_head_kv, il);

            if (use_mrope) {
                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                            ext_factor, attn_factor, beta_fast, beta_slow);

                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                            ext_factor, attn_factor, beta_fast, beta_slow);
            } else {
                // Normal RoPE
                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, 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, nullptr, 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);

            cur = build_attn(inp_attn,
                    model.layers[il].wo, NULL, 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);
        }
        // Post-attention norm (new!)
        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "post_attn_norm", il);

        // Add the input (residual connection after post-attention norm)
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
        cb(ffn_inp, "ffn_inp", il);

        // FF
        {
            // Pre-MLP norm
            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
            cb(cur, "ffn_norm", il);

            // MLP
            cur = build_ffn(cur,
                    model.layers[il].ffn_up, NULL, NULL,
                    NULL, NULL, NULL,
                    model.layers[il].ffn_down, NULL, NULL,
                    NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
            cb(cur, "ffn_out", il);

            // Post-MLP norm
            cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
            cb(cur, "post_mlp_norm", il);
        }
        cur = ggml_add(ctx0, cur, ffn_inp);

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

        // input for next layer
        inpL = cur;
    }
    // Final norm
    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);

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

    // Output projection
    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);
}