llama.cpp/src/models/jamba.cpp

#include "models.h"

void llama_model_jamba::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);

    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);

    for (uint32_t i = 0; i < hparams.n_layer(); ++i) {
        hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
    }

    switch (hparams.n_layer()) {
        // TODO: Jamba layers are a bit heterogeneous, so naming this is hard.
        case 12: // 900M  8x???M
        case 32: // 51B  16x?B
        default: type = LLM_TYPE_UNKNOWN;
    }
}

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

    const int64_t d_conv  = hparams.ssm_d_conv;
    const int64_t d_inner = hparams.ssm_d_inner;
    const int64_t d_state = hparams.ssm_d_state;
    const int64_t dt_rank = hparams.ssm_dt_rank;

    // only an expansion factor of 2 is supported for now
    GGML_ASSERT(2 * n_embd == d_inner);

    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, duplicated to allow offloading
        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) {
        const int64_t n_head_kv = hparams.n_head_kv(i);
        const int64_t n_embd_gqa = hparams.n_embd_v_gqa(i);

        auto & layer = layers[i];

        // norm
        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);

        if (n_head_kv == 0) {
            // Mamba layer
            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}, 0);

            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}, 0);
            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}, 0);

            layer.ssm_x = create_tensor(tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}, 0);

            layer.ssm_dt_norm = create_tensor(tn(LLM_TENSOR_SSM_DT_NORM, "weight", i), {dt_rank}, 0);

            layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}, 0);
            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);

            layer.ssm_b_norm = create_tensor(tn(LLM_TENSOR_SSM_B_NORM, "weight", i), {d_state}, 0);
            layer.ssm_c_norm = create_tensor(tn(LLM_TENSOR_SSM_C_NORM, "weight", i), {d_state}, 0);

            // no "weight" suffix for these
            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);

            // out_proj
            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
        } else {
            // Attention layers

            create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
        }

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

        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);

        if (layer.ffn_gate_inp) {
            // MoE
            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
            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);
        } else {
            // FFN (no MoE)
            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);
        }
    }
}

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

llama_model_jamba::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
    const int64_t n_embd_head = hparams.n_embd_head_v();

    ggml_tensor * cur;
    ggml_tensor * inpL;

    // {n_embd, n_tokens}
    inpL = build_inp_embd(model.tok_embd);

    auto * inp_hybrid = build_inp_mem_hybrid();

    ggml_tensor * inp_out_ids = build_inp_out_ids();

    for (int il = 0; il < n_layer; ++il) {
        const int64_t n_head_kv = hparams.n_head_kv(il);

        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
        cb(cur, "attn_norm", il);

        if (n_head_kv == 0) {
            cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
        } else {
            // Attention

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

            // No RoPE :)
            cur = build_attn(inp_hybrid->get_attn(),
                    model.layers[il].wo, NULL, model.layers[il].wo_s,
                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 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
        struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur);
        cb(cur, "ffn_inp", il);

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

        // feed-forward network
        if (model.layers[il].ffn_gate_inp == nullptr) {
            // FFN
            cur = build_ffn(cur,
                    model.layers[il].ffn_up,   NULL, NULL,
                    model.layers[il].ffn_gate, NULL, NULL,
                    model.layers[il].ffn_down, NULL, NULL,
                    NULL,
                    LLM_FFN_SILU, LLM_FFN_PAR, il);
            cb(cur, "ffn_out", il);
        } else {
            // MoE branch
            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, false,
                    hparams.expert_weights_scale,
                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                    il);
            cb(cur, "ffn_moe_out", il);
        }
        // residual
        cur = ggml_add(ctx0, ffn_inp, cur);

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

        // input for next layer
        inpL = cur;
    }
    // final rmsnorm
    cur = build_norm(inpL, 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);
}