echodict/llama.cpp

1#include "models.h"2 3llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {4    const int64_t n_embd_head = hparams.n_embd_head_v();5 6    ggml_tensor * cur;7    ggml_tensor * inpL;8 9    // {n_embd, n_tokens}10    inpL = build_inp_embd(model.tok_embd);11 12    auto * inp_hybrid = build_inp_mem_hybrid();13 14    ggml_tensor * inp_out_ids = build_inp_out_ids();15 16    for (int il = 0; il < n_layer; ++il) {17        const int64_t n_head_kv = hparams.n_head_kv(il);18 19        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);20        cb(cur, "attn_norm", il);21 22        if (n_head_kv == 0) {23            cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il);24        } else {25            // Attention26 27            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,28                    n_embd_head, n_head, n_head_kv, il);29 30            // No RoPE :)31            cur = build_attn(inp_hybrid->get_attn(),32                    model.layers[il].wo, NULL, model.layers[il].wo_s,33                    Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il);34        }35        if (il == n_layer - 1 && inp_out_ids) {36            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);37            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);38        }39        // residual40        struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur);41        cb(cur, "ffn_inp", il);42 43        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);44        cb(cur, "ffn_norm", il);45 46        // feed-forward network47        if (model.layers[il].ffn_gate_inp == nullptr) {48            // FFN49            cur = build_ffn(cur,50                    model.layers[il].ffn_up,   NULL, NULL,51                    model.layers[il].ffn_gate, NULL, NULL,52                    model.layers[il].ffn_down, NULL, NULL,53                    NULL,54                    LLM_FFN_SILU, LLM_FFN_PAR, il);55            cb(cur, "ffn_out", il);56        } else {57            // MoE branch58            cur = build_moe_ffn(cur,59                    model.layers[il].ffn_gate_inp,60                    model.layers[il].ffn_up_exps,61                    model.layers[il].ffn_gate_exps,62                    model.layers[il].ffn_down_exps,63                    nullptr,64                    n_expert, n_expert_used,65                    LLM_FFN_SILU, false,66                    hparams.expert_weights_scale,67                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,68                    il);69            cb(cur, "ffn_moe_out", il);70        }71        // residual72        cur = ggml_add(ctx0, ffn_inp, cur);73 74        cur = build_cvec(cur, il);75        cb(cur, "l_out", il);76 77        // input for next layer78        inpL = cur;79    }80    // final rmsnorm81    cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1);82 83    cb(cur, "result_norm", -1);84    res->t_embd = cur;85 86    // lm_head87    cur = build_lora_mm(model.output, cur);88 89    cb(cur, "result_output", -1);90    res->t_logits = cur;91 92    ggml_build_forward_expand(gf, cur);93}94