echodict/llama.cpp

1#include "models.h"2 3llm_build_granite::llm_build_granite(4    const llama_model & model,5    const llm_graph_params & params)6    : llm_graph_context(params) {7 8    const int64_t n_embd_head = hparams.n_embd_head_v();9 10    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());11    GGML_ASSERT(n_embd_head == n_rot);12 13    ggml_tensor * cur;14    ggml_tensor * inpL;15 16    inpL = build_inp_embd(model.tok_embd);17 18    // inp_pos - built only if rope enabled19    ggml_tensor * inp_pos = nullptr;20    if (hparams.rope_finetuned) {21        inp_pos = build_inp_pos();22    }23    auto * inp_attn = build_attn_inp_kv();24 25    ggml_tensor * inp_out_ids = build_inp_out_ids();26 27    for (int il = 0; il < n_layer; ++il) {28        ggml_tensor * inpSA = inpL;29 30        // norm31        cur = build_norm(inpL,32                model.layers[il].attn_norm, NULL,33                LLM_NORM_RMS, il);34        cb(cur, "attn_norm", il);35 36        // self-attention37        cur = build_attention_layer(38            cur, inp_pos, inp_attn,39            model, n_embd_head, il);40 41        if (il == n_layer - 1 && inp_out_ids) {42            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);43            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);44        }45        // ffn46        cur = build_layer_ffn(cur, inpSA, model, il);47 48        // input for next layer49        inpL = cur;50    }51    cur = inpL;52 53    cur = build_norm(cur,54            model.output_norm, NULL,55            LLM_NORM_RMS, -1);56 57    cb(cur, "result_norm", -1);58    res->t_embd = cur;59 60    // lm_head61    cur = build_lora_mm(model.output, cur);62 63    // For Granite architectures - scale logits64    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);65    cb(cur, "result_output", -1);66    res->t_logits = cur;67 68    ggml_build_forward_expand(gf, cur);69}70 71ggml_tensor * llm_build_granite::build_attention_layer(72          ggml_tensor             * cur,73          ggml_tensor             * inp_pos,74          llm_graph_input_attn_kv * inp_attn,75    const llama_model             & model,76    const int64_t                 n_embd_head,77    const int                     il) {78 79    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,80            n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);81 82    const bool use_rope = hparams.rope_finetuned;83    if (use_rope) {84        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);85        Qcur = ggml_rope_ext(86                ctx0, Qcur, inp_pos, rope_factors,87                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,88                ext_factor, attn_factor, beta_fast, beta_slow89                );90 91        Kcur = ggml_rope_ext(92                ctx0, Kcur, inp_pos, rope_factors,93                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,94                ext_factor, attn_factor, beta_fast, beta_slow95                );96    }97 98    cb(Qcur, "Qcur", il);99    cb(Kcur, "Kcur", il);100    cb(Vcur, "Vcur", il);101 102    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;103    cur = build_attn(inp_attn,104            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,105            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);106            cb(cur, "attn_out", il);107    return cur;108}109 110ggml_tensor * llm_build_granite::build_layer_ffn(111          ggml_tensor       * cur,112          ggml_tensor       * inpSA,113    const llama_model       & model,114    const int                 il) {115 116    // For Granite architectures - scale residual117    if (hparams.f_residual_scale) {118        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);119    }120    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);121    cb(ffn_inp, "ffn_inp", il);122 123    // feed-forward network (non-MoE)124    if (model.layers[il].ffn_gate_inp == nullptr) {125 126        cur = build_norm(ffn_inp,127                model.layers[il].ffn_norm, NULL,128                LLM_NORM_RMS, il);129                cb(cur, "ffn_norm", il);130 131        cur = build_ffn(cur,132                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,133                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,134                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,135                NULL,136                LLM_FFN_SILU, LLM_FFN_PAR, il);137                cb(cur, "ffn_out", il);138 139    } else {140        // MoE branch141        cur = build_norm(ffn_inp,142                model.layers[il].ffn_norm, NULL,143                LLM_NORM_RMS, il);144                cb(cur, "ffn_norm", il);145 146        ggml_tensor * moe_out = build_moe_ffn(cur,147                model.layers[il].ffn_gate_inp,148                model.layers[il].ffn_up_exps,149                model.layers[il].ffn_gate_exps,150                model.layers[il].ffn_down_exps,151                nullptr,152                n_expert, n_expert_used,153                LLM_FFN_SILU, true,154                hparams.expert_weights_scale,155                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,156                il);157        cb(moe_out, "ffn_moe_out", il);158 159        // For Granite MoE Shared160        if (hparams.n_ff_shexp > 0) {161            ggml_tensor * ffn_shexp = build_ffn(cur,162                model.layers[il].ffn_up_shexp,   NULL, NULL,163                model.layers[il].ffn_gate_shexp, NULL, NULL,164                model.layers[il].ffn_down_shexp, NULL, NULL,165                NULL,166                LLM_FFN_SILU, LLM_FFN_PAR, il);167            cb(ffn_shexp, "ffn_shexp", il);168 169            cur = ggml_add(ctx0, moe_out, ffn_shexp);170            cb(cur, "ffn_out", il);171        } else {172            cur = moe_out;173        }174    }175 176    // For Granite architectures - scale residual177    if (hparams.f_residual_scale) {178        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);179    }180    cur = ggml_add(ctx0, cur, ffn_inp);181    cb(cur, "ffn_out", il);182 183    cur = build_cvec(cur, il);184    cb(cur, "l_out", il);185 186    return cur;187}188