echodict/llama.cpp

1#include "models.h"2 3llm_build_rwkv6qwen2::llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) {4    GGML_ASSERT(n_embd == hparams.n_embd_r());5 6    ggml_tensor * cur;7    ggml_tensor * inpL;8 9    inpL = build_inp_embd(model.tok_embd);10 11    auto * rs_inp = build_rs_inp();12 13    const auto n_embd = hparams.n_embd;14    const auto n_seq_tokens = ubatch.n_seq_tokens;15    const auto n_seqs = ubatch.n_seqs;16 17    ggml_tensor * inp_out_ids = build_inp_out_ids();18 19    for (int il = 0; il < n_layer; ++il) {20        const llama_layer * layer = &model.layers[il];21        inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);22 23        ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il);24 25        ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il);26        cb(att_norm, "attn_norm", il);27 28        ggml_tensor * x_prev = ggml_concat(29                ctx0,30                token_shift,31                ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0),32                133                );34 35        cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il);36 37        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));38        ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il));39 40        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);41        cb(ffn_inp, "ffn_inp", il);42 43        cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);44        ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);45 46        if (il == n_layer - 1 && inp_out_ids) {47            cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);48            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);49        }50 51        // feed-forward network52        cur = build_norm(ffn_inp,53                model.layers[il].ffn_norm, NULL,54                LLM_NORM_RMS, il);55        cb(cur, "ffn_norm", il);56 57        cur = build_ffn(cur,58                model.layers[il].ffn_up,   NULL, NULL,59                model.layers[il].ffn_gate, NULL, NULL,60                model.layers[il].ffn_down, NULL, NULL,61                NULL,62                LLM_FFN_SILU, LLM_FFN_PAR, il);63        cb(cur, "ffn_out", il);64 65        cur = ggml_add(ctx0, cur, ffn_inp);66 67        cur = build_cvec(cur, il);68        cb(cur, "l_out", il);69 70        // input for next layer71        inpL = cur;72    }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}87