char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

PikaPikachuconvert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)465e49b9c2026-09-07

#include "models.h"

void llama_model_jais2::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);

    switch (hparams.n_layer()) {
        case 32: type = LLM_TYPE_8B; break;
        case 68: type = LLM_TYPE_70B; break;
        default: type = LLM_TYPE_UNKNOWN;
    }
}

void llama_model_jais2::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_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
    if (!output) {
        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);
        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", 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);

        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);

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

        // Jais-2 uses simple MLP (no gate) with biases
        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
    }
}

std::unique_ptr<llm_graph_context> llama_model_jais2::build_arch_graph(const llm_graph_params & params) const {
    return std::make_unique<graph>(*this, params);
}

// JAIS-2 model graph builder
// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
llama_model_jais2::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();

    // KV input for attention
    auto * inp_attn = build_attn_inp_kv();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        // Pre-attention LayerNorm
        cur = build_norm(inpL,
                model.layers[il].attn_norm,
                model.layers[il].attn_norm_b,
                LLM_NORM, il);
        cb(cur, "attn_norm", il);

        // Self-attention with separate Q, K, V projections
        {
            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
                    n_embd_head, n_head, n_head_kv, il);

            // Apply 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_rope", il);
            cb(Kcur, "Kcur_rope", 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);
            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
        }

        // Residual connection
        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
        cb(ffn_inp, "ffn_inp", il);

        // Pre-FFN LayerNorm
        cur = build_norm(ffn_inp,
                model.layers[il].ffn_norm,
                model.layers[il].ffn_norm_b,
                LLM_NORM, il);
        cb(cur, "ffn_norm", il);

        // FFN with relu2 activation (ReLU squared) - no gate projection
        // up -> relu2 -> down
        cur = build_ffn(cur,
                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
                NULL, NULL, NULL,  // no gate
                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
                NULL,
                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
        cb(cur, "ffn_out", il);

        // Residual connection
        inpL = ggml_add(ctx0, cur, ffn_inp);
        inpL = build_cvec(inpL, il);
        cb(inpL, "l_out", il);
    }

    // Final LayerNorm
    cur = build_norm(inpL,
            model.output_norm,
            model.output_norm_b,
            LLM_NORM, -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);
}