Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);5    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);6    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);7    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);8    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);9 10    // NextN/MTP: optional draft head appended as extra trailing block(s)11    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);12    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");13 14    // A layer is recurrent IFF the n_head_kv value is set to 0 and15    // the n_ff value is set to 0. Appended MTP blocks are dense (non-recurrent)16    for (uint32_t i = 0; i < hparams.n_layer_all; ++i) {17        hparams.is_recr_impl[i] = i < hparams.n_layer() && hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0;18    }19 20    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);21    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,     hparams.f_norm_eps); // MTP head final_layernorm22 23    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);24    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);25    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);26    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);27    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);28    ml.get_key(LLM_KV_MOE_LATENT_SIZE,                   hparams.moe_latent_size, false);29 30    switch (hparams.n_layer()) {31        case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B32        case 56: type = LLM_TYPE_9B; break;33        case 88: type = LLM_TYPE_120B_A12B; break;34        default: type = LLM_TYPE_UNKNOWN;35    }36}37 38void llama_model_nemotron_h::load_arch_tensors(llama_model_loader & ml) {39    LLAMA_LOAD_LOCALS;40 41    const bool mtp_only    = hparams.n_layer_nextn > 0 && ml.get_weight("blk.0.attn_norm.weight") == nullptr;42    const int  trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;43    const int  mtp_flags   = !ml.load_mtp ? TENSOR_SKIP : 0;44 45    // mamba2 Mixer SSM params46    // NOTE: int64_t for tensor dimensions47    const int64_t d_conv     = hparams.ssm_d_conv;48    const int64_t d_inner    = hparams.ssm_d_inner;49    const int64_t d_state    = hparams.ssm_d_state;50    const int64_t n_ssm_head = hparams.ssm_dt_rank;51    const int64_t n_group    = hparams.ssm_n_group;52    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;53    const int64_t moe_n_embd = hparams.moe_latent_size > 0 ? hparams.moe_latent_size : n_embd;54 55    // embeddings56    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);57 58    // output59    {60        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);61        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);62        // if output is NULL, init from the input tok embed, duplicated to allow offloading63        if (output == NULL) {64            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);65        }66    }67 68    for (int i = 0; i < n_layer; ++i) {69        auto & layer = layers[i];70 71        // all blocks use the attn norm72        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, trunk_flags);73 74        if (hparams.is_recr(i)) {75            // ssm layers76            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, trunk_flags);77 78            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, trunk_flags);79            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);80 81            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, trunk_flags);82 83            // no "weight" suffix for these84            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, trunk_flags);85            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, trunk_flags);86 87            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, trunk_flags);88 89            // out_proj90            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, trunk_flags);91        } else if (hparams.n_ff(i) == 0) {92            // attention layers (with optional bias)93            const int64_t n_head_i = hparams.n_head(i);94            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);95            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);96            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, trunk_flags);97            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, trunk_flags);98            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);99        }  else {100            if (n_expert != 0) {101                const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;102                const int64_t n_ff_shexp = hparams.n_ff_shexp;103 104                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, trunk_flags);105                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, trunk_flags);106 107                // MoE branch108                layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);109                layer.ffn_latent_up   = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP,   "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);110 111                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, trunk_flags);112                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, trunk_flags);113 114                // Shared expert branch115                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, trunk_flags);116                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, trunk_flags);117 118            } else {119                // mlp layers120                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, trunk_flags);121                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, trunk_flags);122                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);123                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);124            }125        }126    }127 128    // NextN/MTP draft head: each predict layer folds an attention sub-layer and a MoE129    // sub-layer into a single trailing block130    for (int i = n_layer; i < n_layer_all; ++i) {131        auto & layer = layers[i];132 133        const int64_t n_head_i       = hparams.n_head(i);134        const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);135        const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);136        const int64_t n_ff_exp       = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;137        const int64_t n_ff_shexp     = hparams.n_ff_shexp;138 139        // NextN input-fusion tensors140        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", i), {n_embd}, mtp_flags);141        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", i), {n_embd}, mtp_flags);142        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", i), {2*n_embd, n_embd}, mtp_flags);143        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, mtp_flags);144 145        // attention sub-layer146        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, mtp_flags);147        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, mtp_flags);148        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, mtp_flags);149        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd}, mtp_flags | TENSOR_NOT_REQUIRED);150 151        // MoE sub-layer152        layer.attn_post_norm  = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM,  "weight", i), {n_embd}, mtp_flags);153        layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, mtp_flags);154        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, mtp_flags);155        layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,   "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, mtp_flags);156        layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,     "weight", i), {moe_n_embd, n_ff_exp,   n_expert}, mtp_flags);157        layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,  "weight", i), {n_ff_shexp, n_embd}, mtp_flags);158        layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,    "weight", i), {n_embd, n_ff_shexp}, mtp_flags);159    }160}161 162std::unique_ptr<llm_graph_context> llama_model_nemotron_h::build_arch_graph(const llm_graph_params & params) const {163    return std::make_unique<graph>(*this, params);164}165 166llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_params & params) :167    llm_build_mamba_base(params) {168    const int64_t n_embd_head = hparams.n_embd_head_v();169    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());170 171    ggml_tensor * cur;172    ggml_tensor * inpL;173 174    inpL = build_inp_embd(model.tok_embd);175    ggml_build_forward_expand(gf, inpL);176 177    auto * inp = build_inp_mem_hybrid();178 179    ggml_tensor * inp_out_ids = build_inp_out_ids();180 181    for (int il = 0; il < n_layer; ++il) {182        struct ggml_tensor * inpSA = inpL;183 184        // norm185        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);186        cb(cur, "attn_norm", il);187 188        if (hparams.is_recr(il)) {189            // ssm layer //190            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);191        } else if (hparams.n_ff(il) == 0) {192            // attention layer //193            cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il);194        } else {195            cur = build_ffn_layer(cur, model, il);196        }197 198        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {199            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);200            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);201        }202 203        // add residual204        cur = ggml_add(ctx0, cur, inpSA);205        cb(cur, "nemotron_h_block_out", il);206 207        // input for next layer208        inpL = cur;209    }210 211    cur = inpL;212 213    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);214 215    // seed for the MTP/NextN draft head216    cb(cur, "h_nextn", -1);217    res->t_h_nextn = cur;218 219    if (!cparams.embeddings_nextn_masked && inp_out_ids) {220        cur = ggml_get_rows(ctx0, cur, inp_out_ids);221    }222 223    cb(cur, "result_norm", -1);224    res->t_embd = cur;225 226    // lm_head227    cur = build_lora_mm(model.output, cur, model.output_s);228    cb(cur, "result_output", -1);229    res->t_logits = cur;230 231    ggml_build_forward_expand(gf, cur);232}233 234ggml_tensor * llama_model_nemotron_h::graph::build_attention_layer(ggml_tensor *             cur,235                                                          llm_graph_input_attn_kv * inp_attn,236                                                          const llama_model &       model,237                                                                int64_t             n_embd_head,238                                                                int                 il) {239    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);240 241    const float kq_scale =242        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;243    cur = build_attn(inp_attn,244            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,245            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);246    cb(cur, "attn_out", il);247    return cur;248}249 250ggml_tensor * llama_model_nemotron_h::graph::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {251    if (model.layers[il].ffn_gate_inp == nullptr) {252        cur = build_ffn(cur,253                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,254                NULL,                      NULL,                        NULL,255                model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,256                NULL,257                LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);258        cb(cur, "ffn_out", il);259    } else {260        ggml_tensor * inp_emb    = cur;261        ggml_tensor * inp_latent = cur;262 263        if (model.layers[il].ffn_latent_down) {264            inp_latent = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_down, cur);265        }266 267        ggml_tensor * router_logits = build_lora_mm(model.layers[il].ffn_gate_inp, cur);268        cb(router_logits, "ffn_moe_logits", il);269 270        ggml_tensor * moe_out =271            build_moe_ffn(inp_latent,272                    model.layers[il].ffn_gate_inp,273                    model.layers[il].ffn_up_exps,274                    nullptr, // no gate275                    model.layers[il].ffn_down_exps,276                    model.layers[il].ffn_exp_probs_b,277                    n_expert, n_expert_used,278                    LLM_FFN_RELU_SQR, hparams.expert_weights_norm,279                    hparams.expert_weights_scale,280                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,281                    il,282                    router_logits, nullptr,283                    model.layers[il].ffn_up_exps_s,284                    nullptr, // no gate285                    model.layers[il].ffn_down_exps_s);286        cb(moe_out, "ffn_moe_out", il);287 288        if (model.layers[il].ffn_latent_up) {289            moe_out = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_up, moe_out);290        }291 292        ggml_tensor * ffn_shexp = build_ffn(inp_emb,293                    model.layers[il].ffn_up_shexp,   NULL, model.layers[il].ffn_up_shexp_s,294                    NULL /* no gate */           ,   NULL, NULL,295                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,296                    NULL,297                    LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);298        cb(ffn_shexp, "ffn_shexp", il);299 300        cur = ggml_add(ctx0, moe_out, ffn_shexp);301        cb(cur, "ffn_out", il);302    }303 304    cur = build_cvec(cur, il);305    cb(cur, "l_out", il);306 307    return cur;308}309