1#include "models.h"
 2
 3llm_build_rwkv6::llm_build_rwkv6(const llama_model & model, const llm_graph_params & params) :
 4    llm_build_rwkv6_base(model, params) {
 5    GGML_ASSERT(hparams.token_shift_count == 2);
 6
 7    ggml_tensor * cur;
 8    ggml_tensor * inpL;
 9
10    inpL = build_inp_embd(model.tok_embd);
11    inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
12
13    auto * rs_inp = build_rs_inp();
14
15    const auto n_embd       = hparams.n_embd;
16    const auto n_seq_tokens = ubatch.n_seq_tokens;
17    const auto n_seqs       = ubatch.n_seqs;
18
19    ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21    for (int il = 0; il < n_layer; ++il) {
22        const llama_layer * layer = &model.layers[il];
23        inpL                      = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
24
25        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);
26
27        ggml_tensor * att_shift =
28            ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
29        ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1],
30                                               token_shift->nb[2], n_embd * ggml_element_size(token_shift));
31
32        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il);
33        cb(att_norm, "attn_norm", il);
34
35        ggml_tensor * x_prev = ggml_concat(
36            ctx0, att_shift,
37            ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1);
38
39        cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);
40
41        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
42        cb(ffn_inp, "ffn_inp", il);
43
44        ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il);
45        cb(ffn_norm, "ffn_norm", il);
46
47        x_prev = ggml_concat(
48            ctx0, ffn_shift,
49            ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1);
50
51        token_shift = ggml_concat(ctx0,
52                                  ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2],
53                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)),
54                                  ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2],
55                                               (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)),
56                                  1);
57        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
58
59        ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
60        ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
61        x_prev   = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens);
62        cur      = ggml_reshape_2d(ctx0, cur, 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            cur      = ggml_get_rows(ctx0, cur, inp_out_ids);
69        }
70        cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6);
71        cur = ggml_add(ctx0, cur, ffn_inp);
72
73        if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) {
74            cur = ggml_scale(ctx0, cur, 0.5F);
75        }
76        cur = build_cvec(cur, il);
77        cb(cur, "l_out", il);
78
79        // input for next layer
80        inpL = cur;
81    }
82    cur = inpL;
83    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1);
84
85    cb(cur, "result_norm", -1);
86    res->t_embd = cur;
87
88    cur = build_lora_mm(model.output, cur);
89
90    cb(cur, "result_output", -1);
91    res->t_logits = cur;
92
93    ggml_build_forward_expand(gf, cur);
94}