1#include "models.h"
  2
  3template <bool iswa>
  4llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){
  5    const int64_t n_embd_head = hparams.n_embd_head_v;
  6
  7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
  8    GGML_ASSERT(n_embd_head == hparams.n_rot);
  9
 10    ggml_tensor * cur;
 11    ggml_tensor * inpL;
 12
 13    inpL = build_inp_embd(model.tok_embd);
 14
 15    // inp_pos - contains the positions
 16    ggml_tensor * inp_pos = build_inp_pos();
 17
 18    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
 19    inp_attn_type * inp_attn = nullptr;
 20
 21    if constexpr (iswa) {
 22        inp_attn = build_attn_inp_kv_iswa();
 23    } else {
 24        inp_attn = build_attn_inp_kv();
 25    }
 26    ggml_tensor * inp_out_ids = build_inp_out_ids();
 27
 28    for (int il = 0; il < n_layer; ++il) {
 29        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
 30        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
 31
 32        ggml_tensor * inpSA  = inpL;
 33
 34        // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
 35        const bool use_rope = hparams.n_no_rope_layer_step == n_layer ||
 36                              il % hparams.n_no_rope_layer_step != 0;
 37
 38        ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL);  // [n_expert, n_tokens]
 39        cb(probs, "ffn_moe_logits", il);
 40
 41        // norm
 42        cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
 43        cb(cur, "attn_norm", il);
 44
 45        // self_attention
 46        {
 47            // compute Q and K and RoPE them
 48            struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
 49            cb(Qcur, "Qcur", il);
 50
 51            struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
 52            cb(Kcur, "Kcur", il);
 53
 54            struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
 55            cb(Vcur, "Vcur", il);
 56
 57            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 58            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
 59            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
 60
 61            if (use_rope) {
 62                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
 63                                    ext_factor, attn_factor, beta_fast, beta_slow);
 64
 65                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
 66                                    ext_factor, attn_factor, beta_fast, beta_slow);
 67            }
 68            cb(Qcur, "Qcur", il);
 69            cb(Kcur, "Kcur", il);
 70
 71            cur = build_attn(inp_attn,
 72                    model.layers[il].wo, model.layers[il].bo,
 73                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
 74        }
 75        if (il == n_layer - 1 && inp_out_ids) {
 76            cur = ggml_get_rows(ctx0, cur, inp_out_ids);
 77            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
 78            probs = ggml_get_rows(ctx0, probs, inp_out_ids);
 79        }
 80        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
 81        cb(ffn_inp, "ffn_inp", il);
 82
 83        // MoE branch
 84        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
 85        cb(cur, "ffn_norm", il);
 86
 87        ggml_tensor * ffn_out =
 88            build_moe_ffn(cur,
 89                    nullptr,
 90                    model.layers[il].ffn_up_exps,
 91                    model.layers[il].ffn_gate_exps,
 92                    model.layers[il].ffn_down_exps,
 93                    nullptr,
 94                    n_expert, n_expert_used,
 95                    LLM_FFN_RELU, true,
 96                    false, 0.0,
 97                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
 98                    il, probs);
 99
100        cb(ffn_out, "ffn_out", il);
101        cur = ffn_out;
102
103        cur = ggml_add(ctx0, cur, ffn_inp);
104        cur = build_cvec(cur, il);
105        cb(cur, "l_out", il);
106
107        // input for next layer
108        inpL = cur;
109    }
110    cur = inpL;
111
112    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
113    cb(cur, "result_norm", -1);
114    res->t_embd = cur;
115
116    // lm_head
117    cur = build_lora_mm(model.output, cur);
118    cb(cur, "result_output", -1);
119    res->t_logits = cur;
120
121    ggml_build_forward_expand(gf, cur);
122}
123
124// Explicit template instantiations
125template struct llm_build_smallthinker<false>;
126template struct llm_build_smallthinker<true>;