#include "models.h" void llama_model_hyv3::load_arch_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, 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); // MTP / NextN head (optional). Absent in pre-MTP gguf -> n_layer_nextn stays 0 // and everything below (extra block load, MTP graph, MTP context) is skipped. ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false); if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) { hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID; } switch (hparams.n_layer()) { case 48: type = LLM_TYPE_30B_A3B; break; default: type = LLM_TYPE_UNKNOWN; } } void llama_model_hyv3::load_arch_tensors(llama_model_loader &) { LLAMA_LOAD_LOCALS; tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); 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 == NULL) { output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, TENSOR_NOT_REQUIRED); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED); layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED); layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, TENSOR_NOT_REQUIRED); create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, TENSOR_NOT_REQUIRED); layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); } // MTP / NextN block(s): loaded as extra decoder blocks beyond the main stack // (index range [n_layer, n_layer_all)). Skipped entirely for pre-MTP gguf // where n_layer_all == n_layer, so those models load unchanged. for (int i = n_layer; i < (int) hparams.n_layer_all; ++i) { auto & layer = layers[i]; const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp; // Standard hy_v3 MoE decoder block (same layout as a trunk sparse layer). layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0); create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0); layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}, TENSOR_NOT_REQUIRED); layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, TENSOR_NOT_REQUIRED); // NextN-specific tensors. eh_proj fuses [enorm(embed), hnorm(hidden)] (2*n_embd -> n_embd). // shared_head_head / embed_tokens are tied to the main lm_head / tok_embd in Hy3, // so they are optional and fall back in the graph. layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, 0); layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, 0); layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, 0); layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); } } std::unique_ptr llama_model_hyv3::build_arch_graph(const llm_graph_params & params) const { if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) { return std::make_unique(*this, params); } return std::make_unique(*this, params); } llama_model_hyv3::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()); GGML_ASSERT(n_embd_head == n_rot); ggml_tensor * cur; ggml_tensor * inpL; inpL = build_inp_embd(model.tok_embd); ggml_tensor * inp_pos = build_inp_pos(); auto * inp_attn = build_attn_inp_kv(); ggml_tensor * inp_out_ids = build_inp_out_ids(); const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); { ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il); Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, 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, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cur = build_attn(inp_attn, model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "attn_out", il); } if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) { 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); cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "ffn_norm", il); if (model.layers[il].ffn_gate_inp == nullptr) { cur = build_ffn(cur, model.layers[il].ffn_up, model.layers[il].ffn_up_b, model.layers[il].ffn_up_s, model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s, model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s, nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(cur, "ffn_dense_out", il); } else { ggml_tensor * moe_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, nullptr, model.layers[il].ffn_gate_up_exps, model.layers[il].ffn_up_exps_s, model.layers[il].ffn_gate_exps_s, model.layers[il].ffn_down_exps_s); cb(moe_out, "ffn_moe_out", il); ggml_tensor * sh_out = build_ffn(cur, model.layers[il].ffn_up_shexp, nullptr, model.layers[il].ffn_up_shexp_s, model.layers[il].ffn_gate_shexp, nullptr, model.layers[il].ffn_gate_shexp_s, model.layers[il].ffn_down_shexp, nullptr, model.layers[il].ffn_down_shexp_s, nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(sh_out, "ffn_shared_out", il); cur = ggml_add(ctx0, moe_out, sh_out); cb(cur, "ffn_out", il); } cur = ggml_add(ctx0, cur, ffn_inp); cur = build_cvec(cur, il); cb(cur, "l_out", il); inpL = cur; } cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1); // post-norm hidden state feeds both the LM head and the MTP seed (t_h_nextn). // When masking is off, the MTP path needs the full-width hidden, so defer the // output-id gather until after capturing t_h_nextn (matches qwen35moe). cb(cur, "h_nextn", -1); res->t_h_nextn = cur; if (!cparams.embeddings_nextn_masked && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); } cb(cur, "result_norm", -1); res->t_embd = cur; 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); } // LLM_GRAPH_TYPE_DECODER_MTP draft head for Hy3 MoE. // Mirrors vLLM HYV3MultiTokenPredictorLayer.forward: // e = enorm(inputs_embeds); h = hnorm(previous_hidden); // x = eh_proj(cat([e, h])); x = mtp_block(x); x += residual; x = final_ln(x) // logits = lm_head(x) (lm_head/embed tied to the main model) // Differences vs qwen35moe MTP: default RoPE (not mrope), no attention gate, // no shared-expert gate, sigmoid MoE gating with correction bias. llama_model_hyv3::graph_mtp::graph_mtp(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()); GGML_ASSERT(n_embd_head == n_rot); GGML_ASSERT(hparams.n_layer_nextn > 0 && "HYV3 MTP requires n_layer_nextn > 0"); GGML_ASSERT(hparams.n_layer_nextn == 1 && "HYV3 MTP currently only supports a single MTP block"); const int il = hparams.n_layer(); const auto & layer = model.layers[il]; GGML_ASSERT(layer.nextn.eh_proj && "HYV3 MTP block missing nextn.eh_proj"); GGML_ASSERT(layer.nextn.enorm && "HYV3 MTP block missing nextn.enorm"); GGML_ASSERT(layer.nextn.hnorm && "HYV3 MTP block missing nextn.hnorm"); GGML_ASSERT(layer.ffn_gate_inp && "HYV3 MTP block missing ffn_gate_inp"); const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); // Inputs: token ids (-> embedding), and the previous-step hidden state h. auto inp = std::make_unique(hparams.n_embd); inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); ggml_set_input(inp->tokens); inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens); ggml_set_input(inp->embd); ggml_tensor * tok_embd; if (ubatch.token) { ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd; tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens); } else { tok_embd = inp->embd; } cb(tok_embd, "mtp_tok_embd", il); inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens); ggml_set_input(inp->h); ggml_set_name(inp->h, "mtp_h_input"); ggml_tensor * h_embd = inp->h; res->add_input(std::move(inp)); ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); auto * inp_attn = build_attn_inp_kv(); // e = enorm(embed), h = hnorm(hidden); fuse via eh_proj: cat([e, h]) -> n_embd ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il); cb(e_norm, "mtp_enorm", il); ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il); cb(h_norm, "mtp_hnorm", il); ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0); cb(concat, "mtp_concat", il); ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s); cb(cur, "mtp_eh_proj", il); // --- one standard hy_v3 decoder block on the fused hidden --- ggml_tensor * inpSA = cur; cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "mtp_attn_norm", il); { ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il); Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il); Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il); Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, 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, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cur = build_attn(inp_attn, layer.wo, layer.wo_b, layer.wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "mtp_attn_out", il); } ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "mtp_ffn_inp", il); cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "mtp_ffn_norm", il); ggml_tensor * moe_out = build_moe_ffn(cur, layer.ffn_gate_inp, layer.ffn_up_exps, layer.ffn_gate_exps, layer.ffn_down_exps, layer.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, nullptr, layer.ffn_gate_up_exps, layer.ffn_up_exps_s, layer.ffn_gate_exps_s, layer.ffn_down_exps_s); cb(moe_out, "mtp_ffn_moe_out", il); ggml_tensor * sh_out = build_ffn(cur, layer.ffn_up_shexp, nullptr, layer.ffn_up_shexp_s, layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s, layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s, nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(sh_out, "mtp_ffn_shared_out", il); cur = ggml_add(ctx0, moe_out, sh_out); cb(cur, "mtp_ffn_out", il); cur = ggml_add(ctx0, cur, ffn_inp); cb(cur, "mtp_post_ffn", il); // final_layernorm then LM head (both fall back to the main model when tied) ggml_tensor * head_norm_w = layer.nextn.shared_head_norm ? layer.nextn.shared_head_norm : model.output_norm; GGML_ASSERT(head_norm_w && "HYV3 MTP: missing both nextn.shared_head_norm and output_norm"); cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1); cb(cur, "h_nextn", -1); res->t_h_nextn = cur; cur = ggml_get_rows(ctx0, cur, inp_out_ids); cb(cur, "mtp_final_norm", -1); ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output; ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s; GGML_ASSERT(head_w && "HYV3 MTP: missing LM head (nextn.shared_head_head or model.output)"); cur = build_lora_mm(head_w, cur, head_s); cb(cur, "result_output", -1); res->t_logits = cur; ggml_build_forward_expand(gf, cur); }