Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_bailingmoe2::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);5    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);6    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);7    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);8    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);9    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);10    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);11    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);12    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.n_layer_nextn, false);13 14    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_impl");15 16    switch (hparams.n_layer()) {17        case 20: type = LLM_TYPE_16B_A1B; break;18        case 32: type = LLM_TYPE_100B_A6B; break;19        default: type = LLM_TYPE_UNKNOWN;20    }21}22 23void llama_model_bailingmoe2::load_arch_tensors(llama_model_loader &) {24    LLAMA_LOAD_LOCALS;25    const int64_t n_expert_shared = hparams.n_expert_shared;26 27    const int64_t n_ff_exp        = hparams.n_ff_exp;28 29    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);30 31    // output32    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);33    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);34 35    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for bailingmoe2");36    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for bailingmoe2");37 38    for (int i = 0; i < n_layer_all; ++i) {39        int flags = 0;40        if (i >= n_layer) {41            // skip all tensors in the NextN layers42            flags |= TENSOR_SKIP;43        }44 45        auto & layer = layers[i];46 47        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);48 49        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, flags);50        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);51 52        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);53        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);54 55        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);56 57        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers58            const int64_t n_ff_shexp = (hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp) * n_expert_shared;59 60            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);61            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);62 63            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);64            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);65            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);66 67            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);68            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);69            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);70        } else { // Dense layers71            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);72            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);73            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);74        }75 76        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers77        if (i >= n_layer) {78            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);79            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);80            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);81            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);82            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);83            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);84            layer.layer_out_norm         = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, flags);85        }86    }87}88 89std::unique_ptr<llm_graph_context> llama_model_bailingmoe2::build_arch_graph(const llm_graph_params & params) const {90    return std::make_unique<graph>(*this, params);91}92 93llama_model_bailingmoe2::graph::graph(const llama_model & model, const llm_graph_params & params) :94    llm_graph_context(params) {95    const int64_t n_embd_head = hparams.n_embd_head_v();96 97    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());98 99    ggml_tensor * cur;100    ggml_tensor * inpL;101 102    inpL = build_inp_embd(model.tok_embd);103 104    // inp_pos - contains the positions105    ggml_tensor * inp_pos = build_inp_pos();106 107    auto * inp_attn = build_attn_inp_kv();108 109    ggml_tensor * inp_out_ids = build_inp_out_ids();110 111    for (int il = 0; il < n_layer; ++il) {112        ggml_tensor * inpSA = inpL;113 114        // norm115        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);116        cb(cur, "attn_norm", il);117 118        // self_attention119        {120            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,121                    n_embd_head, n_head, n_head_kv, il);122 123            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);124            cb(Qcur, "Qcur_normed", il);125 126            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,127                                 ext_factor, attn_factor, beta_fast, beta_slow);128 129            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);130            cb(Kcur, "Kcur_normed", il);131 132            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,133                                 ext_factor, attn_factor, beta_fast, beta_slow);134 135            cb(Qcur, "Qcur", il);136            cb(Kcur, "Kcur", il);137            cb(Vcur, "Vcur", il);138 139            cur = build_attn(inp_attn,140                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,141                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);142        }143 144        if (il == n_layer - 1 && inp_out_ids) {145            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);146            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);147        }148 149        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA);150        cb(sa_out, "sa_out", il);151 152        // MoE branch153        cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);154        cb(cur, "ffn_norm", il);155 156        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {157            cur = build_ffn(cur,158                    model.layers[il].ffn_up, NULL, NULL,159                    model.layers[il].ffn_gate, NULL, NULL,160                    model.layers[il].ffn_down, NULL, NULL,161                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);162            cb(cur, "ffn_out", il);163        } else {164            ggml_tensor * moe_out = build_moe_ffn(cur,165                model.layers[il].ffn_gate_inp,166                model.layers[il].ffn_up_exps,167                model.layers[il].ffn_gate_exps,168                model.layers[il].ffn_down_exps,169                model.layers[il].ffn_exp_probs_b,170                n_expert, n_expert_used,171                LLM_FFN_SILU, hparams.expert_weights_norm,172                hparams.expert_weights_scale,173                (llama_expert_gating_func_type) hparams.expert_gating_func,174                il);175            cb(moe_out, "ffn_moe_out", il);176 177            {178                ggml_tensor * ffn_shexp =179                    build_ffn(cur,180                        model.layers[il].ffn_up_shexp, NULL, NULL,181                        model.layers[il].ffn_gate_shexp, NULL, NULL,182                        model.layers[il].ffn_down_shexp, NULL, NULL,183                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);184                cb(ffn_shexp, "ffn_shexp", il);185 186                cur = ggml_add(ctx0, moe_out, ffn_shexp);187                cb(cur, "ffn_out", il);188            }189        }190 191        cur = ggml_add(ctx0, cur, sa_out);192 193        cur = build_cvec(cur, il);194        cb(cur, "l_out", il);195 196        // input for next layer197        inpL = cur;198    }199 200    cur = inpL;201 202    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);203 204    cb(cur, "result_norm", -1);205    res->t_embd = cur;206 207    // lm_head208    cur = build_lora_mm(model.output, cur, model.output_s);209 210    cb(cur, "result_output", -1);211    res->t_logits = cur;212 213    ggml_build_forward_expand(gf, cur);214}215