1#include "models.h"
 2
 3llm_build_rwkv7::llm_build_rwkv7(const llama_model & model, const llm_graph_params & params) :
 4    llm_build_rwkv7_base(model, params) {
 5    GGML_ASSERT(hparams.token_shift_count == 2);
 6
 7    ggml_tensor * cur;
 8    ggml_tensor * inpL;
 9    ggml_tensor * v_first = nullptr;
10
11    inpL = build_inp_embd(model.tok_embd);
12    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
13
14    auto * rs_inp = build_rs_inp();
15
16    const auto n_embd       = hparams.n_embd;
17    const auto n_seq_tokens = ubatch.n_seq_tokens;
18    const auto n_seqs       = ubatch.n_seqs;
19
20    ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22    for (int il = 0; il < n_layer; ++il) {
23        const llama_layer * layer = &model.layers[il];
24        inpL                      = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
25
26        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
27
28        ggml_tensor * att_shift =
29            ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
30        ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1],
31                                               token_shift->nb[2], n_embd * ggml_element_size(token_shift));
32
33        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
34        cb(att_norm, "attn_norm", il);
35
36        ggml_tensor * x_prev = ggml_concat(
37            ctx0, att_shift,
38            ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1);
39
40        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
41
42        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
43        cb(ffn_inp, "ffn_inp", il);
44
45        ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
46        cb(ffn_norm, "ffn_norm", il);
47
48        x_prev = ggml_concat(
49            ctx0, ffn_shift,
50            ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1);
51
52        token_shift = ggml_concat(ctx0,
53                                  ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2],
54                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)),
55                                  ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2],
56                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)),
57                                  1);
58        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
59
60        ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
61        ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
62        x_prev   = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens);
63
64        if (il == n_layer - 1 && inp_out_ids) {
65            ffn_inp  = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
66            ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
67            x_prev   = ggml_get_rows(ctx0, x_prev, inp_out_ids);
68        }
69        cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7);
70        cur = ggml_add(ctx0, cur, ffn_inp);
71
72        cur = build_cvec(cur, il);
73        cb(cur, "l_out", il);
74
75        // input for next layer
76        inpL = cur;
77    }
78    cur = inpL;
79    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
80
81    cb(cur, "result_norm", -1);
82    res->t_embd = cur;
83
84    cur = build_lora_mm(model.output, cur);
85
86    cb(cur, "result_output", -1);
87    res->t_logits = cur;
88
89    ggml_build_forward_expand(gf, cur);
90}