1#include "models.h"
  2
  3llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) {
  4    const int64_t n_embd_head = hparams.n_embd_head_v;
  5
  6    ggml_tensor * cur;
  7    ggml_tensor * inpL;
  8
  9    // {n_embd, n_tokens}
 10    inpL = build_inp_embd(model.tok_embd);
 11
 12    auto * inp_hybrid = build_inp_mem_hybrid();
 13
 14    ggml_tensor * inp_out_ids = build_inp_out_ids();
 15
 16    for (int il = 0; il < n_layer; ++il) {
 17        const int64_t n_head_kv = hparams.n_head_kv(il);
 18
 19        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 20        cb(cur, "attn_norm", il);
 21
 22        if (n_head_kv == 0) {
 23            cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);
 24        } else {
 25            // Attention
 26
 27            struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 28            struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 29            struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 30
 31            cb(Qcur, "Qcur", il);
 32            cb(Kcur, "Kcur", il);
 33            cb(Vcur, "Vcur", il);
 34
 35            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 36            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 37            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 38
 39            cb(Qcur, "Qcur", il);
 40            cb(Kcur, "Kcur", il);
 41            cb(Vcur, "Vcur", il);
 42
 43            // No RoPE :)
 44            cur = build_attn(inp_hybrid->get_attn(),
 45                    model.layers[il].wo, NULL,
 46                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);
 47        }
 48        if (il == n_layer - 1 && inp_out_ids) {
 49            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
 50            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
 51        }
 52        // residual
 53        struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur);
 54        cb(cur, "ffn_inp", il);
 55
 56        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 57        cb(cur, "ffn_norm", il);
 58
 59        // feed-forward network
 60        if (model.layers[il].ffn_gate_inp == nullptr) {
 61            // FFN
 62            cur = build_ffn(cur,
 63                    model.layers[il].ffn_up,   NULL, NULL,
 64                    model.layers[il].ffn_gate, NULL, NULL,
 65                    model.layers[il].ffn_down, NULL, NULL,
 66                    NULL,
 67                    LLM_FFN_SILU, LLM_FFN_PAR, il);
 68            cb(cur, "ffn_out", il);
 69        } else {
 70            // MoE branch
 71            cur = build_moe_ffn(cur,
 72                    model.layers[il].ffn_gate_inp,
 73                    model.layers[il].ffn_up_exps,
 74                    model.layers[il].ffn_gate_exps,
 75                    model.layers[il].ffn_down_exps,
 76                    nullptr,
 77                    n_expert, n_expert_used,
 78                    LLM_FFN_SILU, false,
 79                    false, 0.0,
 80                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
 81                    il);
 82            cb(cur, "ffn_moe_out", il);
 83        }
 84        // residual
 85        cur = ggml_add(ctx0, ffn_inp, cur);
 86
 87        cur = build_cvec(cur, il);
 88        cb(cur, "l_out", il);
 89
 90        // input for next layer
 91        inpL = cur;
 92    }
 93    // final rmsnorm
 94    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);
 95
 96    cb(cur, "result_norm", -1);
 97    res->t_embd = cur;
 98
 99    // lm_head
100    cur = build_lora_mm(model.output, cur);
101
102    cb(cur, "result_output", -1);
103    res->t_logits = cur;
104
105    ggml_build_forward_expand(gf, cur);
106}