Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);6    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);7 8    // MoE parameters9    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);10    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);11    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);12    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);13    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);14    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);15 16    // Expert gating function (GLM-4.5 uses sigmoid)17    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);18    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {19        hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;20    }21 22    // NextN/MTP parameters23    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);24    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_impl");25 26    switch (hparams.n_layer()) {27        case 46: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air28        case 48: type = LLM_TYPE_102B_A12B; break; // Solar Open29        case 92: type = LLM_TYPE_355B_A32B; break; // GLM-4.530        default: type = LLM_TYPE_UNKNOWN;31    }32}33 34void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {35    LLAMA_LOAD_LOCALS;36    const int64_t n_expert_shared = hparams.n_expert_shared;37 38 39    GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");40    GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");41 42    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);43 44    // output45    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);46    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);47    // if output is NULL, init from the input tok embed48    if (output == NULL) {49        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);50    }51 52    // Load ALL tensors including NextN layer to satisfy total tensor count53    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass54    for (int i = 0; i < n_layer_all; ++i) {55        int flags = 0;56        if (i >= n_layer) {57            // skip all tensors in the NextN layers58            flags |= TENSOR_SKIP;59        }60 61        auto & layer = layers[i];62 63        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);64 65        // GLM-style attention with bias terms66        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);67 68        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);69 70        // K/Q norm tensors (optional for GLM-4.5 355B variant)71        layer.attn_q_norm = create_tensor(72            tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);73        layer.attn_k_norm = create_tensor(74            tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);75 76        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags);77 78        // Check if this layer uses MoE or dense FFN based on n_layer_dense_lead79        // GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE80        const bool use_moe = (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead);81 82        if (use_moe) {83            // MoE layers84            layer.ffn_gate_inp =85                create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags);86            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags);87 88            // MoE branch89            const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;90 91            layer.ffn_gate_exps = create_tensor(92                tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);93            layer.ffn_down_exps = create_tensor(94                tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags);95            layer.ffn_up_exps = create_tensor(96                tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);97 98            // Shared expert99            if (n_expert_shared > 0) {100                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;101                layer.ffn_gate_shexp = create_tensor(102                    tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);103                layer.ffn_down_shexp = create_tensor(104                    tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);105                layer.ffn_up_shexp = create_tensor(106                    tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);107            }108        } else {109            // Dense layers (first k layers) - GLM uses separate gate/up projections110            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags);111            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags);112            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);113        }114 115        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers116        if (i >= n_layer) {117            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);118            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);119            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);120 121            // Optional tensors122            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);123            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);124            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);125        }126    }127}128 129std::unique_ptr<llm_graph_context> llama_model_glm4_moe::build_arch_graph(const llm_graph_params & params) const {130    return std::make_unique<graph>(*this, params);131}132 133llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {134    const int64_t n_embd_head = hparams.n_embd_head_v();135 136    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());137 138    int sections[4];139    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);140 141    ggml_tensor * cur;142    ggml_tensor * inpL;143 144    inpL = build_inp_embd(model.tok_embd);145 146    bool use_mrope = hparams.use_mrope();147    if (ubatch.embd && !use_mrope) {148        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results149        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");150    }151 152    // inp_pos - contains the positions153    ggml_tensor * inp_pos = build_inp_pos();154 155    auto * inp_attn = build_attn_inp_kv();156 157    ggml_tensor * inp_out_ids = build_inp_out_ids();158 159    // Only process up to last layer (skip final NextN layer)160    // Final layer tensors are loaded but not processed in forward pass161    for (int il = 0; il < n_layer; ++il) {162        ggml_tensor * inpSA = inpL;163 164        // Pre-attention norm165        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);166        cb(cur, "attn_norm", il);167 168        // self-attention169        {170            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,171                    n_embd_head, n_head, n_head_kv, il);172 173            // Apply Q/K norm if available (GLM-4.5 355B variant)174            if (model.layers[il].attn_q_norm) {175                Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);176                cb(Qcur, "Qcur_normed", il);177            }178            if (model.layers[il].attn_k_norm) {179                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);180                cb(Kcur, "Kcur_normed", il);181            }182 183            if (use_mrope) {184                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,185                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,186                            ext_factor, attn_factor, beta_fast, beta_slow);187 188                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,189                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,190                            ext_factor, attn_factor, beta_fast, beta_slow);191            } else {192                // Normal RoPE193                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,194                                    rope_type, n_ctx_orig, freq_base, freq_scale,195                                    ext_factor, attn_factor, beta_fast, beta_slow);196 197                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,198                                    rope_type, n_ctx_orig, freq_base, freq_scale,199                                    ext_factor, attn_factor, beta_fast, beta_slow);200            }201 202            cb(Qcur, "Qcur", il);203            cb(Kcur, "Kcur", il);204            cb(Vcur, "Vcur", il);205 206            cur = build_attn(inp_attn,207                    model.layers[il].wo, NULL, model.layers[il].wo_s,208                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);209        }210        if (il == n_layer - 1 && inp_out_ids) {211            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);212            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);213        }214        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);215        cb(ffn_inp, "ffn_inp", il);216 217        // Post-attention norm218        cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);219        cb(cur, "post_attn_norm", il);220 221        // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)222        if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {223            // Dense FFN layer224            cur = build_ffn(cur,225                    model.layers[il].ffn_up,   NULL, NULL,226                    model.layers[il].ffn_gate, NULL, NULL,227                    model.layers[il].ffn_down, NULL, NULL,228                    NULL,229                    LLM_FFN_SILU, LLM_FFN_PAR, il);230            cb(cur, "ffn_out", il);231        } else {232            // Process routed experts using existing MoE infrastructure233            ggml_tensor * routed_out = build_moe_ffn(cur,234                    model.layers[il].ffn_gate_inp,235                    model.layers[il].ffn_up_exps,236                    model.layers[il].ffn_gate_exps,237                    model.layers[il].ffn_down_exps,238                    model.layers[il].ffn_exp_probs_b,239                    n_expert, n_expert_used,240                    LLM_FFN_SILU, hparams.expert_weights_norm,241                    hparams.expert_weights_scale,242                    (llama_expert_gating_func_type) hparams.expert_gating_func,243                    il);244            cb(routed_out, "ffn_moe_out", il);245 246            // Process shared expert on original input247            ggml_tensor * shared_out = build_ffn(cur,248                    model.layers[il].ffn_up_shexp,   NULL, NULL,249                    model.layers[il].ffn_gate_shexp, NULL, NULL,250                    model.layers[il].ffn_down_shexp, NULL, NULL,251                    NULL,252                    LLM_FFN_SILU, LLM_FFN_PAR, il);253            cb(shared_out, "ffn_shexp_out", il);254 255            // Final output: routed_output + shared_output256            cur = ggml_add(ctx0, routed_out, shared_out);257            cb(cur, "ffn_out", il);258        }259        cur = ggml_add(ctx0, cur, ffn_inp);260 261        cur = build_cvec(cur, il);262        cb(cur, "l_out", il);263 264        // input for next layer265        inpL = cur;266    }267    cur = inpL;268    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);269 270    cb(cur, "result_norm", -1);271    res->t_embd = cur;272 273    // lm_head274    cur = build_lora_mm(model.output, cur, model.output_s);275 276    cb(cur, "result_output", -1);277    res->t_logits = cur;278 279    ggml_build_forward_expand(gf, cur);280}281