1#include "models.h"
 2
 3
 4llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
 5    GGML_ASSERT(n_embd == hparams.n_embd_r());
 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
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_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);
28        cb(att_norm, "attn_norm", il);
29
30        ggml_tensor * x_prev = ggml_concat(
31                ctx0,
32                token_shift,
33                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),
34                1
35                );
36
37        cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il);
38
39        token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm));
40        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));
41
42        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
43        cb(ffn_inp, "ffn_inp", il);
44
45        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
46        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
47
48        if (il == n_layer - 1 && inp_out_ids) {
49            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
50            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
51        }
52        // feed-forward network
53        cur = build_norm(ffn_inp,
54                model.layers[il].ffn_norm, NULL,
55                LLM_NORM_RMS, il);
56        cb(cur, "ffn_norm", il);
57
58        cur = build_ffn(cur,
59                model.layers[il].ffn_up,   NULL, NULL,
60                model.layers[il].ffn_gate, NULL, NULL,
61                model.layers[il].ffn_down, NULL, NULL,
62                NULL,
63                LLM_FFN_SILU, LLM_FFN_PAR, il);
64        cb(cur, "ffn_out", il);
65
66        cur = ggml_add(ctx0, cur, ffn_inp);
67
68        cur = build_cvec(cur, il);
69        cb(cur, "l_out", il);
70
71        // input for next layer
72        inpL = cur;
73    }
74    cur = inpL;
75    cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1);
76
77    cb(cur, "result_norm", -1);
78    res->t_embd = cur;
79
80    cur = build_lora_mm(model.output, cur);
81
82    cb(cur, "result_output", -1);
83    res->t_logits = cur;
84
85    ggml_build_forward_expand(gf, cur);
86}