#include "models.h" void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false); // MoE parameters ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert); ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used); ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false); ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false); ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); // Expert gating function (GLM-4.5 uses sigmoid) ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) { hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID; } // NextN/MTP parameters ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false); GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer"); // TODO: when MTP is implemented, this should probably be updated if needed hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers; switch (hparams.n_layer) { case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer) case 48: type = LLM_TYPE_102B_A12B; break; // Solar Open case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer) default: type = LLM_TYPE_UNKNOWN; } } void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) { LLAMA_LOAD_LOCALS; const int64_t n_expert_shared = hparams.n_expert_shared; GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers"); GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers"); tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); // if output is NULL, init from the input tok embed if (output == NULL) { output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED); } // Load ALL tensors including NextN layer to satisfy total tensor count // but only PROCESS up to last layer (skipping final NextN layer) in forward pass for (int i = 0; i < n_layer; ++i) { int flags = 0; if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { // skip all tensors in the NextN layers flags |= TENSOR_SKIP; } auto & layer = layers[i]; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags); // GLM-style attention with bias terms create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags); // K/Q norm tensors (optional for GLM-4.5 355B variant) layer.attn_q_norm = create_tensor( tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags); layer.attn_k_norm = create_tensor( tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags); // Check if this layer uses MoE or dense FFN based on n_layer_dense_lead // GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE const bool use_moe = (static_cast(i) >= hparams.n_layer_dense_lead); if (use_moe) { // MoE layers layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags); layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags); // MoE branch const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; layer.ffn_gate_exps = create_tensor( tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags); layer.ffn_down_exps = create_tensor( tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags); layer.ffn_up_exps = create_tensor( tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags); // Shared expert if (n_expert_shared > 0) { const int64_t n_ff_shexp = n_ff_exp * n_expert_shared; layer.ffn_gate_shexp = create_tensor( tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags); layer.ffn_down_shexp = create_tensor( tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags); layer.ffn_up_shexp = create_tensor( tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags); } } else { // Dense layers (first k layers) - GLM uses separate gate/up projections layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags); layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, flags); } // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags); layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags); layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags); // Optional tensors layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED); } } } std::unique_ptr llama_model_glm4_moe::build_arch_graph(const llm_graph_params & params) const { return std::make_unique(*this, params); } llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); int sections[4]; std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); ggml_tensor * cur; ggml_tensor * inpL; inpL = build_inp_embd(model.tok_embd); bool use_mrope = hparams.use_mrope(); if (ubatch.embd && !use_mrope) { // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results GGML_ABORT("This GGUF does not support multimodal. Please reconvert it."); } // inp_pos - contains the positions ggml_tensor * inp_pos = build_inp_pos(); auto * inp_attn = build_attn_inp_kv(); ggml_tensor * inp_out_ids = build_inp_out_ids(); // Only process up to last layer (skip final NextN layer) // Final layer tensors are loaded but not processed in forward pass const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; for (int il = 0; il < n_transformer_layers; ++il) { ggml_tensor * inpSA = inpL; // Pre-attention norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); // self-attention { auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il); // Apply Q/K norm if available (GLM-4.5 355B variant) if (model.layers[il].attn_q_norm) { Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); cb(Qcur, "Qcur_normed", il); } if (model.layers[il].attn_k_norm) { Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); cb(Kcur, "Kcur_normed", il); } if (use_mrope) { Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr, n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr, n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); } else { // Normal RoPE Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); } cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); cur = build_attn(inp_attn, model.layers[il].wo, NULL, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_transformer_layers - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); // Post-attention norm cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "post_attn_norm", il); // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense) if (static_cast(il) < hparams.n_layer_dense_lead) { // Dense FFN layer cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(cur, "ffn_out", il); } else { // Process routed experts using existing MoE infrastructure ggml_tensor * routed_out = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, hparams.expert_weights_norm, hparams.expert_weights_scale, (llama_expert_gating_func_type) hparams.expert_gating_func, il); cb(routed_out, "ffn_moe_out", il); // Process shared expert on original input ggml_tensor * shared_out = build_ffn(cur, model.layers[il].ffn_up_shexp, NULL, NULL, model.layers[il].ffn_gate_shexp, NULL, NULL, model.layers[il].ffn_down_shexp, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(shared_out, "ffn_shexp_out", il); // Final output: routed_output + shared_output cur = ggml_add(ctx0, routed_out, shared_out); cb(cur, "ffn_out", il); } cur = ggml_add(ctx0, cur, ffn_inp); cur = build_cvec(cur, il); cb(cur, "l_out", il); // input for next layer inpL = cur; } cur = inpL; cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); cb(cur, "result_norm", -1); res->t_embd = cur; // lm_head cur = build_lora_mm(model.output, cur, model.output_s); cb(cur, "result_output", -1); res->t_logits = cur; ggml_build_forward_expand(gf, cur); }