1#include "models.h"
  2
  3llm_build_olmoe::llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
  4    const int64_t n_embd_head = hparams.n_embd_head_v;
  5
  6    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  7    GGML_ASSERT(n_embd_head == hparams.n_rot);
  8
  9    ggml_tensor * cur;
 10    ggml_tensor * inpL;
 11
 12    inpL = build_inp_embd(model.tok_embd);
 13
 14    // inp_pos - contains the positions
 15    ggml_tensor * inp_pos = build_inp_pos();
 16
 17    auto * inp_attn = build_attn_inp_kv();
 18
 19    ggml_tensor * inp_out_ids = build_inp_out_ids();
 20
 21    for (int il = 0; il < n_layer; ++il) {
 22        ggml_tensor * inpSA = inpL;
 23
 24        // norm
 25        cur = build_norm(inpL,
 26                model.layers[il].attn_norm, NULL,
 27                LLM_NORM_RMS, il);
 28        cb(cur, "attn_norm", il);
 29
 30        // self_attention
 31        {
 32            // compute Q and K and RoPE them
 33            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 34            cb(Qcur, "Qcur", il);
 35
 36            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 37            cb(Kcur, "Kcur", il);
 38
 39            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 40            cb(Vcur, "Vcur", il);
 41
 42            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
 43                    LLM_NORM_RMS, il);
 44            cb(Qcur, "Qcur_normed", il);
 45
 46            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
 47                    LLM_NORM_RMS, il);
 48            cb(Kcur, "Kcur_normed", il);
 49
 50            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
 51            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 52            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 53
 54            Qcur = ggml_rope_ext(
 55                    ctx0, Qcur, inp_pos, nullptr,
 56                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 57                    ext_factor, attn_factor, beta_fast, beta_slow
 58                    );
 59
 60            Kcur = ggml_rope_ext(
 61                    ctx0, Kcur, inp_pos, nullptr,
 62                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
 63                    ext_factor, attn_factor, beta_fast, beta_slow
 64                    );
 65
 66            cb(Qcur, "Qcur", il);
 67            cb(Kcur, "Kcur", il);
 68            cb(Vcur, "Vcur", il);
 69
 70            cur = build_attn(inp_attn,
 71                    model.layers[il].wo, NULL,
 72                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 73        }
 74        if (il == n_layer - 1 && inp_out_ids) {
 75            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
 76            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 77        }
 78        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 79        cb(ffn_inp, "ffn_inp", il);
 80
 81        // MoE branch
 82        cur = build_norm(ffn_inp,
 83                model.layers[il].ffn_norm, NULL,
 84                LLM_NORM_RMS, il);
 85        cb(cur, "ffn_norm", il);
 86
 87        cur = build_moe_ffn(cur,
 88                model.layers[il].ffn_gate_inp,
 89                model.layers[il].ffn_up_exps,
 90                model.layers[il].ffn_gate_exps,
 91                model.layers[il].ffn_down_exps,
 92                nullptr,
 93                n_expert, n_expert_used,
 94                LLM_FFN_SILU, false,
 95                false, 0.0,
 96                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
 97                il);
 98        cb(cur, "ffn_moe_out", il);
 99
100        cur = ggml_add(ctx0, cur, ffn_inp);
101
102        cur = build_cvec(cur, il);
103        cb(cur, "l_out", il);
104
105        // input for next layer
106        inpL = cur;
107    }
108    cur = inpL;
109
110    cur = build_norm(cur,
111            model.output_norm, NULL,
112            LLM_NORM_RMS, -1);
113
114    cb(cur, "result_norm", -1);
115    res->t_embd = cur;
116
117    // lm_head
118    cur = build_lora_mm(model.output, cur);
119
120    cb(cur, "result_output", -1);
121    res->t_logits = cur;
122
123    ggml_build_forward_expand(gf, cur);
124}