#pragma once #include "llama-model.h" #include "llama-graph.h" #include "llama-model-loader.h" // note: almost all graphs require at least sqrtf, so include cmath globally #include #include class llama_memory_hybrid_idx_context; // ref: https://github.com/ggml-org/llama.cpp/pull/28068 static inline ggml_tensor * build_gdn_l2_norm(ggml_context * ctx, ggml_tensor * x, float eps) { const float n = x->ne[0]; return ggml_scale(ctx, ggml_rms_norm(ctx, x, eps/n), 1.0f/sqrtf(n)); } // // base classes // struct llm_build_mamba_base : public llm_graph_context { llm_build_mamba_base(const llm_graph_params & params); virtual ~llm_build_mamba_base() = default; ggml_tensor * build_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_ubatch & ubatch, int il); ggml_tensor * build_mamba2_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_ubatch & ubatch, int il) const; }; struct llm_build_delta_net_base : public llm_graph_context { llm_build_delta_net_base(const llm_graph_params & params); virtual ~llm_build_delta_net_base() = default; // returns pair of output and new state std::pair build_delta_net( ggml_tensor * q, ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, ggml_tensor * b, ggml_tensor * s); // read conv state from cache, concat with qkv_mixed, write back (single slot or per-token) // qkv_mixed: (qkv_dim, n_seq_tokens, n_seqs); returns conv_input: (kernel_size + n_seq_tokens - 1, channels, n_seqs) ggml_tensor * build_conv_state( llm_graph_input_rs * inp, ggml_tensor * conv_states_all, ggml_tensor * qkv_mixed, int64_t conv_kernel_size, int64_t conv_channels, int il); // run delta-net attention and write the new recurrent state(s) back to ssm_states_all // s: (head_v_dim, head_v_dim, num_v_heads, n_seqs); returns output: (head_v_dim, num_v_heads, n_seq_tokens, n_seqs) ggml_tensor * build_recurrent_attn( llm_graph_input_rs * inp, ggml_tensor * ssm_states_all, ggml_tensor * q, ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, ggml_tensor * b, ggml_tensor * s, int il); }; struct llm_build_lfm_base : public llm_graph_context { llm_build_lfm_base(const llm_graph_params & params); virtual ~llm_build_lfm_base() = default; ggml_tensor * build_lfm_shortconv(const llama_memory_recurrent_context * mctx_recr, llm_graph_input_rs * inp_recr, ggml_tensor * cur, int il); // no recurrent state (decision graphs): causal over the ubatch from a zero state ggml_tensor * build_lfm_shortconv(ggml_tensor * cur, int il); ggml_tensor * build_lfm_dense_ffn(ggml_tensor * cur, int il); ggml_tensor * build_lfm_moe_ffn(ggml_tensor * cur, int il); }; struct llm_build_rwkv6_base : public llm_graph_context { llm_build_rwkv6_base(const llm_graph_params & params); virtual ~llm_build_rwkv6_base() = default; ggml_tensor * build_rwkv6_channel_mix(const llama_layer * layer, ggml_tensor * cur, ggml_tensor * x_prev, llm_arch arch) const; ggml_tensor * build_rwkv6_time_mix(llm_graph_input_rs * inp, ggml_tensor * cur, ggml_tensor * x_prev, const llama_ubatch & ubatch, int il) const; }; // Base class for RWKV7-related models struct llm_build_rwkv7_base : public llm_graph_context { llm_build_rwkv7_base(const llm_graph_params & params); virtual ~llm_build_rwkv7_base() = default; // RWKV7-specific graph building methods ggml_tensor * build_rwkv7_channel_mix(const llama_layer * layer, ggml_tensor * cur, ggml_tensor * x_prev, llm_arch arch) const; ggml_tensor * build_rwkv7_time_mix(llm_graph_input_rs * inp, ggml_tensor * cur, ggml_tensor * x_prev, ggml_tensor *& first_layer_value, const llama_ubatch & ubatch, int il) const; }; // // models // struct llama_model_llama : public llama_model_base { llama_model_llama(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_llama4 : public llama_model_base { llama_model_llama4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_llama_embed : public llama_model_llama { llama_model_llama_embed(const struct llama_model_params & params) : llama_model_llama(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_llama template using graph = llama_model_llama::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_maincoder : public llama_model_base { llama_model_maincoder(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_maion_coder : public llama_model_base { llama_model_maion_coder(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_talkie : public llama_model_base { llama_model_talkie(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deci : public llama_model_base { llama_model_deci(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_baichuan : public llama_model_base { llama_model_baichuan(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_falcon : public llama_model_base { llama_model_falcon(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_grok : public llama_model_base { llama_model_grok(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_starcoder : public llama_model_base { llama_model_starcoder(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_refact : public llama_model_base { llama_model_refact(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bert : public llama_model_base { llama_model_bert(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_jina_bert_v2 : public llama_model_base { llama_model_jina_bert_v2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_bert::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_jina_bert_v3 : public llama_model_base { llama_model_jina_bert_v3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_bert::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nomic_bert : public llama_model_base { llama_model_nomic_bert(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_bert::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nomic_bert_moe : public llama_model_base { llama_model_nomic_bert_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_bert::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_modern_bert : public llama_model_base { llama_model_modern_bert(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_decision_head( ggml_tensor * inp, llm_graph_input_attn_no_cache * inp_attn, ggml_tensor * inp_out_ids); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_neo_bert : public llama_model_base { llama_model_neo_bert(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_eurobert : public llama_model_base { llama_model_eurobert(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bloom : public llama_model_base { llama_model_bloom(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; // Quant-only stub for mmproj GGUFs // none of these are ever called, they only exist to satisfy the llama_model_base interface struct llama_model_clip : public llama_model_base { llama_model_clip(const struct llama_model_params & params) : llama_model_base(params) {} [[noreturn]] void load_arch_hparams(llama_model_loader & ml) override; [[noreturn]] void load_arch_tensors(llama_model_loader & ml) override; [[noreturn]] std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mpt : public llama_model_base { llama_model_mpt(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_stablelm : public llama_model_base { llama_model_stablelm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mellum : public llama_model_base { llama_model_mellum(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; template struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nanbeige : public llama_model_base { llama_model_nanbeige(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; int n_loops = 1; int n_layer_phys = 0; bool skip_loop_final_norm = false; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen : public llama_model_base { llama_model_qwen(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen2 : public llama_model_base { llama_model_qwen2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_dream : public llama_model_base { llama_model_dream(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_llada : public llama_model_base { llama_model_llada(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_llada_moe : public llama_model_base { llama_model_llada_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_rnd1 : public llama_model_base { llama_model_rnd1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen2vl : public llama_model_base { llama_model_qwen2vl(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen2moe : public llama_model_base { llama_model_qwen2moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_sensenova_u1 : public llama_model_base { llama_model_sensenova_u1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; float rope_freq_base_spatial = 10000.0f; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3 : public llama_model_base { llama_model_qwen3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); protected: // the helpers alone, a derived graph builds its own body struct no_build {}; graph(const llm_graph_params & params, no_build) : llm_graph_context(params) {} // the layers and the final norm: [n_embd, n_outputs] ggml_tensor * build_backbone(); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; // Qwen3 with the decision head of vllm-sr Decision 2.0: one score per option struct llama_model_vllm_decision2 : public llama_model_qwen3 { llama_model_vllm_decision2(const struct llama_model_params & params) : llama_model_qwen3(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; ggml_tensor * option_norm = nullptr; ggml_tensor * option_norm_b = nullptr; ggml_tensor * query_norm = nullptr; ggml_tensor * query_norm_b = nullptr; ggml_tensor * key = nullptr; ggml_tensor * query = nullptr; ggml_tensor * option_mlp = nullptr; ggml_tensor * option_mlp_b = nullptr; ggml_tensor * query_mlp = nullptr; ggml_tensor * scorer_out = nullptr; class input_decision; struct graph : public llama_model_qwen3::graph { graph(const llm_graph_params & params); private: // hidden: [n_embd, n_outputs], returns one score per option: [1, n_options] ggml_tensor * build_head(ggml_tensor * hidden, input_decision * inp); const llama_model_vllm_decision2 & model; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3moe : public llama_model_base { llama_model_qwen3moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_flint : public llama_model_base { llama_model_flint(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3vl : public llama_model_base { llama_model_qwen3vl(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3vlmoe : public llama_model_base { llama_model_qwen3vlmoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3tts : public llama_model_qwen3vl { llama_model_qwen3tts(const struct llama_model_params & params) : llama_model_qwen3vl(params) {} }; struct llama_model_phi2 : public llama_model_base { llama_model_phi2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_phi3 : public llama_model_base { llama_model_phi3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_phimoe : public llama_model_base { llama_model_phimoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template using graph = llama_model_phi3::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_plamo : public llama_model_base { llama_model_plamo(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_plamo2 : public llama_model_base { llama_model_plamo2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); private: ggml_tensor * build_plamo2_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_ubatch & ubatch, int il); ggml_tensor * build_plamo2_attn_layer(llm_graph_input_attn_kv * inp, ggml_tensor * inp_pos, ggml_tensor * cur, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_plamo3 : public llama_model_base { llama_model_plamo3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gpt2 : public llama_model_base { llama_model_gpt2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_pockettts : public llama_model_base { llama_model_pockettts(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_kokoro : public llama_model_base { llama_model_kokoro(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; // token + position embeddings are narrower than n_embd, then projected up ggml_tensor * tok_embd_proj = nullptr; ggml_tensor * tok_embd_proj_b = nullptr; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_tiny_voca : public llama_model_base { llama_model_tiny_voca(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_zonos3 : public llama_model_base { llama_model_zonos3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: // Full AttnRes bank: embedding + every sublayer output so far, [n_embd, n_src, n_tokens] ggml_tensor * bank = nullptr; void attn_res_push(ggml_tensor * cur); ggml_tensor * attn_res_route(ggml_tensor * query, int il); ggml_tensor * build_layer_attn(llm_graph_input_attn_kv * inp, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_gdn(llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn(ggml_tensor * cur, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_codeshell : public llama_model_base { llama_model_codeshell(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_orion : public llama_model_base { llama_model_orion(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_internlm2 : public llama_model_base { llama_model_internlm2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_minicpm3 : public llama_model_base { llama_model_minicpm3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma : public llama_model_base { llama_model_gemma(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma2 : public llama_model_base { llama_model_gemma2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma3 : public llama_model_base { llama_model_gemma3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma3n : public llama_model_base { llama_model_gemma3n(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { const int64_t n_embd_head; const int64_t n_embd_altup; const int64_t n_altup; const int i_altup_act; const int n_layer_sparsity = 10; // number of layers using activation sparsity const float f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95) graph(const llm_graph_params & params); ggml_tensor * calc_magnitude(ggml_tensor * x); // TODO: refactor in common "per-layer" functionality [TAG_PER_LAYER] ggml_tensor * build_inp_per_layer(); ggml_tensor * project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer); ggml_tensor * gaussian_topk(ggml_tensor * x); ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il); ggml_tensor * altup_predict(ggml_tensor * cur, int il); ggml_tensor * laurel(ggml_tensor * cur, int il); ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma4 : public llama_model_base { llama_model_gemma4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { const int64_t n_embd_per_layer; graph(const llm_graph_params & params); // TODO: refactor in common "per-layer" functionality [TAG_PER_LAYER] ggml_tensor * build_inp_per_layer(); ggml_tensor * project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma4_assistant : public llama_model_base { llama_model_gemma4_assistant(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma_embedding : public llama_model_base { llama_model_gemma_embedding(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gemma_embedding2 : public llama_model_base { llama_model_gemma_embedding2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_inp_per_layer(ggml_tensor * inpL); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_starcoder2 : public llama_model_base { llama_model_starcoder2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mamba : public llama_model_base { llama_model_mamba(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mamba2 : public llama_model_base { llama_model_mamba2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_mamba::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_maple : public llama_model_base { llama_model_maple(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_jamba : public llama_model_base { llama_model_jamba(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_xverse : public llama_model_base { llama_model_xverse(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_command_r : public llama_model_base { llama_model_command_r(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_cohere2 : public llama_model_base { llama_model_cohere2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_cohere2moe : public llama_model_base { llama_model_cohere2moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_dbrx : public llama_model_base { llama_model_dbrx(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_olmo : public llama_model_base { llama_model_olmo(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_olmo2 : public llama_model_base { llama_model_olmo2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_olmoe : public llama_model_base { llama_model_olmoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_muse_spark : public llama_model_base { llama_model_muse_spark(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_muse_glimmer : public llama_model_base { llama_model_muse_glimmer(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_openelm : public llama_model_base { llama_model_openelm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_gptneox : public llama_model_base { llama_model_gptneox(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_arctic : public llama_model_base { llama_model_arctic(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deepseek : public llama_model_base { llama_model_deepseek(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deepseek2 : public llama_model_base { llama_model_deepseek2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deepseek32 : public llama_model_base { llama_model_deepseek32(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_dots3note : public llama_model_base { llama_model_dots3note(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; // YaRN attention factor that cancels ggml's rope mscale (the DeepSeek-V4 reference applies none) static inline float dsv4_rope_attn_factor(float freq_scale, float ext_factor) { if (ext_factor == 0.0f) { return 1.0f; } return 1.0f / (1.0f + 0.1f*logf(1.0f/freq_scale)); } struct llama_model_deepseek4 : public llama_model_base { llama_model_deepseek4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { struct no_build {}; graph(const llm_graph_params & params, no_build) : llm_graph_context(params) {} graph(const llm_graph_params & params); // collapse the hc streams with per-stream weights ggml_tensor * build_hc_pre( ggml_tensor * x, ggml_tensor * weights, int il) const; // mean over the hyper-connection streams: [n_embd, hc, n_tokens] -> [n_embd, n_tokens] ggml_tensor * build_hc_mean(ggml_tensor * x) const; // all outputs of the hc mix builders: // build_hc_pre - cur is the collapsed input, post / comb feed build_hc_post // build_hc_mixes - pre is the pre weights, post / comb feed build_hc_post struct hc_pre { ggml_tensor * cur; ggml_tensor * post; ggml_tensor * comb; }; struct hc_mixes { ggml_tensor * pre; ggml_tensor * post; ggml_tensor * comb; }; hc_pre build_hc_pre( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base, int il) const; hc_mixes build_hc_mixes( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base, int il) const; ggml_tensor * build_hc_post( ggml_tensor * x, ggml_tensor * residual, ggml_tensor * post, ggml_tensor * comb, int il) const; ggml_tensor * build_hc_sinkhorn( ggml_tensor * comb, int il) const; ggml_tensor * build_hc_head( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base) const; ggml_tensor * build_attention( llm_graph_input_dsv4 * inp_dsv4, ggml_tensor * cur, ggml_tensor * inp_pos, int il) const; ggml_tensor * build_attention( llm_graph_input_attn_k_iswa * inp_mtp, ggml_tensor * cur, ggml_tensor * inp_pos, int il) const; ggml_tensor * build_attention_impl( llm_graph_input_dsv4 * inp_dsv4, llm_graph_input_attn_k_iswa * inp_mtp, ggml_tensor * cur, ggml_tensor * inp_pos, int il) const; ggml_tensor * build_hca_compressed_kv_from_state( ggml_tensor * kv_state, ggml_tensor * score_state, ggml_tensor * state_read_idxs, ggml_tensor * comp_pos, ggml_tensor * norm, int64_t n_embd_head, const char * name, int il) const; ggml_tensor * build_overlap_compressed_kv_from_state( ggml_tensor * kv_state, ggml_tensor * score_state, ggml_tensor * state_read_idxs, ggml_tensor * comp_pos, ggml_tensor * norm, int64_t ratio, int64_t n_embd_head, const char * name, int il) const; ggml_tensor * build_lid_top_k( llm_graph_input_dsv4 * inp_dsv4, ggml_tensor * qr, ggml_tensor * cur, ggml_tensor * inp_pos, int il) const; ggml_tensor * build_top_k_mask( ggml_tensor * kq_mask, ggml_tensor * top_k, const char * name, int il) const; ggml_tensor * build_csa_lid_attention( llm_graph_input_dsv4 * inp_dsv4, llm_graph_input_dsv4_raw * inp_attn, ggml_tensor * q, ggml_tensor * kv, ggml_tensor * qr, ggml_tensor * cur, ggml_tensor * inp_pos, ggml_tensor * sinks, float kq_scale, int il) const; ggml_tensor * build_hca_attention( llm_graph_input_dsv4 * inp_dsv4, llm_graph_input_dsv4_raw * inp_attn, ggml_tensor * q, ggml_tensor * kv, ggml_tensor * sinks, float kq_scale, int il) const; ggml_tensor * build_raw_attention( llm_graph_input_dsv4_raw * inp_attn, ggml_tensor * q, ggml_tensor * kv, ggml_tensor * sinks, float kq_scale, int il) const; }; struct graph_mtp : public graph { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_xing4_0 : public llama_model_base { llama_model_xing4_0(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llama_model & model, const llm_graph_params & params); ggml_tensor * build_hc_pre( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base, ggml_tensor ** post, ggml_tensor ** comb, int il) const; ggml_tensor * build_hc_pre( ggml_tensor * x, ggml_tensor * weights, int il) const; ggml_tensor * build_hc_sinkhorn( ggml_tensor * comb, int il) const; ggml_tensor * build_hc_post( ggml_tensor * x, ggml_tensor * residual, ggml_tensor * post, ggml_tensor * comb, int il) const; }; struct graph_mtp : public llm_graph_context { graph_mtp(const llama_model & model, const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deepseek41 : public llama_model_deepseek4 { llama_model_deepseek41(const struct llama_model_params & params) : llama_model_deepseek4(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; // Engram: the hash constants are hparams.ple_* (one module per Engram layer), hashed on the host (ggml has no int64 mul/xor/mod) int engram_index(int il) const; void engram_hash(int e, const llama_ubatch & ubatch, const std::vector & prev, std::vector & out) const; struct graph : public llama_model_deepseek4::graph { graph(const llm_graph_params & params); // helpers only (no graph built): base of graph_dsv41 graph(const llm_graph_params & params, no_build) : llama_model_deepseek4::graph(params, no_build{}) {} ggml_tensor * build_engram( const llama_model_deepseek41 & model, ggml_tensor * x, ggml_tensor * rows, ggml_tensor * keep, int il) const; ggml_tensor * build_attention41( const llama_model_deepseek41 & model, llm_graph_input_dsv4 * inp_dsv4, ggml_tensor * cur, ggml_tensor * inp_pos, ggml_tensor * cand_pin, int il); std::vector topk_of; // top-k of each index source layer std::vector comp_row_of; // [n_layer] new compressed rows of each kv source (for mirrors) ggml_tensor * cand_sel = nullptr; // candidate blocks of the candidate source layer int64_t cand_n_comp = 0; // compressed rows the candidate source scored // sparse path (LLAMA_DSV41_SPARSE=0 disables): fused indexer, FA bounded to window + top-k, decode gather bool sparse = true; // gather path, shared by layers with the same top-k: selected K rows per (storage layer, top-k source), // the sorted top-k and the mask per top-k source (one input copy per split) std::map, ggml_tensor *> gather_k_of; std::map gather_sel_of; std::map gather_mask_of; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_deepseek2ocr : public llama_model_base { llama_model_deepseek2ocr(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_deepseek2::graph; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_glm_dsa : public llama_model_base { llama_model_glm_dsa(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; llama_rope_params get_rope_params_indexer(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_eagle3 : public llama_model_base { llama_model_eagle3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_inp_embd_enc() const; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_dflash : public llama_model_base { llama_model_dflash(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; enum variant_type { VARIANT_DFLASH, VARIANT_DFLASH2, VARIANT_DSPARK, VARIANT_LILICORR, }; enum backbone_type { BACKBONE_BASE, BACKBONE_DSV4, BACKBONE_DSV41, BACKBONE_MISTRAL_LARGE4, }; variant_type variant = VARIANT_DFLASH; backbone_type backbone = BACKBONE_BASE; void load_hparams_lilicorr(llama_model_loader & ml); void load_hparams_dsv4(llama_model_loader & ml); void load_hparams_mistral_large4(llama_model_loader & ml); void load_tensors_dflash2(); void load_tensors_dflash2_conv(); void load_tensors_dspark(int64_t n_vocab_draft); void load_tensors_lilicorr(); void load_tensors_layers_base(); void load_tensors_layers_dsv4(); void load_tensors_layers_mistral_large4(); template struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_inp_embd_enc() const; }; // inputs and stages shared by the decoder graphs of the heads struct graph_base : public llm_graph_context { int64_t n_embd_head; bool use_iswa; int sections[4]; ggml_tensor * inp_pos; llm_graph_input_attn_kv * inp_attn = nullptr; llm_graph_input_attn_kv_iswa * inp_attn_iswa = nullptr; graph_base(const llm_graph_params & params); ggml_tensor * build_inject(); ggml_tensor * build_inp_tokens(); ggml_tensor * build_layers(ggml_tensor * inp_tokens); ggml_tensor * build_logits(ggml_tensor * cur); ggml_tensor * build_rope(ggml_tensor * cur, ggml_tensor * pos, int il); }; struct graph_dflash : public graph_base { graph_dflash(const llm_graph_params & params); }; struct graph_dflash2 : public graph_base { graph_dflash2(const llm_graph_params & params); }; struct graph_dspark : public graph_base { graph_dspark(const llm_graph_params & params); }; struct graph_lilicorr : public graph_base { graph_lilicorr(const llm_graph_params & params); }; struct graph_dsv4 : public llama_model_deepseek4::graph { graph_dsv4(const llm_graph_params & params); }; // DeepSeek-V4.1 DSpark: single-pass mHC (lagged pre-mix, no hc_head), no per-head q norm struct graph_dsv41 : public llama_model_deepseek41::graph { graph_dsv41(const llm_graph_params & params); }; // Mistral Large 4 DSpark: dense MLA stages, non-causal, rnn markov head struct graph_mistral_large4 : public llm_graph_context { graph_mistral_large4(const llm_graph_params & params); }; // graph helpers shared by the backbones static ggml_tensor * build_inject_inp(llm_graph_context & g, const llama_model & model); static ggml_tensor * get_tok_embd(const llm_graph_context & g, const llama_model & model); static std::pair get_lm_head(const llm_graph_context & g, const llama_model & model); // DFlash2 local conv, run by the base layers of DFlash2 and LiliCorr drafts static ggml_tensor * build_dflash2_conv(llm_graph_context & g, ggml_tensor * hidden, ggml_tensor * dynamic, ggml_tensor * base, int side); std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mistral_large4 : public llama_model_base { llama_model_mistral_large4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params_indexer(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mistral4 : public llama_model_deepseek2 { llama_model_mistral4(const struct llama_model_params & params) : llama_model_deepseek2(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_deepseek2 using graph = llama_model_deepseek2::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_chatglm : public llama_model_base { llama_model_chatglm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_glm4 : public llama_model_base { llama_model_glm4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_glm4_moe : public llama_model_base { llama_model_glm4_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bitnet : public llama_model_base { llama_model_bitnet(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_t5 : public llama_model_base { llama_model_t5(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_t5encoder : public llama_model_base { llama_model_t5encoder(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_t5::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_jais : public llama_model_base { llama_model_jais(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_jais2 : public llama_model_base { llama_model_jais2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nemotron : public llama_model_base { llama_model_nemotron(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nemotron_h : public llama_model_base { llama_model_nemotron_h(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); ggml_tensor * build_ffn_layer(ggml_tensor * cur, int il); ggml_tensor * build_attention_layer(ggml_tensor * cur, llm_graph_input_attn_kv * inp_attn, int64_t n_embd_head, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_nemotron_h_moe : public llama_model_nemotron_h { llama_model_nemotron_h_moe(const struct llama_model_params & params) : llama_model_nemotron_h(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_nemotron_h using graph = llama_model_nemotron_h::graph; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_metis16 : public llama_model_base { llama_model_metis16(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; uint32_t n_pass = 1; uint32_t n_stream = 1; uint32_t n_mem = 0; uint32_t n_route = 0; uint32_t n_layer_phys = 0; float cont_limit = 0.0f; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); // memory bank of the current forward: keys [n_mem, n_entry, n_tokens], values [n_entry, n_mem, n_tokens] ggml_tensor * bank_k = nullptr; ggml_tensor * bank_vt = nullptr; ggml_tensor * stream_mean(ggml_tensor * x) const; ggml_tensor * hc_read(ggml_tensor * x, ggml_tensor * mix, ggml_tensor * read, ggml_tensor * norm, int p, ggml_tensor ** mixed, int il) const; ggml_tensor * hc_write(ggml_tensor * mixed, ggml_tensor * write, ggml_tensor * update, int p) const; ggml_tensor * mem_read(const llama_model_metis16 & model, ggml_tensor ** x) const; void mem_write(const llama_model_metis16 & model, ggml_tensor * x, ggml_tensor * summary, ggml_tensor * entropy, ggml_tensor * confidence, ggml_tensor * cont, int entry); ggml_tensor * route_features(const llama_model_metis16 & model, ggml_tensor * state, ggml_tensor * summary, ggml_tensor * diff, ggml_tensor * history) const; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_exaone : public llama_model_base { llama_model_exaone(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_exaone4 : public llama_model_base { llama_model_exaone4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_exaone_moe : public llama_model_base { llama_model_exaone_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_rwkv6 : public llama_model_base { llama_model_rwkv6(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_rwkv6_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_rwkv6qwen2 : public llama_model_base { llama_model_rwkv6qwen2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_rwkv6_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_rwkv7 : public llama_model_base { llama_model_rwkv7(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_rwkv7_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_arwkv7 : public llama_model_base { llama_model_arwkv7(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_rwkv7_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_granite : public llama_model_base { llama_model_granite(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); private: ggml_tensor * build_attention_layer( ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, const int64_t n_embd_head, const int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, ggml_tensor * inpSA, const int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_granite_moe : public llama_model_base { llama_model_granite_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_granite::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_granite_switch : public llama_model_base { llama_model_granite_switch(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; uint32_t n_adapters = 0; uint32_t max_lora_rank = 0; float router_gain = 15.0f; std::unordered_map adapter_token_to_slot; std::unordered_map adapter_token_to_substitute; struct graph : public llm_graph_context { graph(const llm_graph_params & params); private: ggml_tensor * build_switched_lora_delta( ggml_tensor * lora_a, ggml_tensor * lora_b, ggml_tensor * cur, ggml_tensor * ids); ggml_tensor * build_switched_lora_mm( ggml_tensor * w, ggml_tensor * lora_a, ggml_tensor * lora_b, ggml_tensor * cur, ggml_tensor * ids); ggml_tensor * build_attention_layer( ggml_tensor * cur, ggml_tensor * inp_pos, ggml_tensor * adapter_ids, llm_graph_input_attn_kv * inp_attn, const int64_t n_embd_head, const int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, ggml_tensor * inpSA, ggml_tensor * adapter_ids, const int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_minicpm : public llama_model_base { llama_model_minicpm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; using graph = llama_model_granite::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_granite_hybrid : public llama_model_base { llama_model_granite_hybrid(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); ggml_tensor * build_layer_ffn(ggml_tensor * cur, ggml_tensor * inpSA, const int il); ggml_tensor * build_attention_layer(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, const int64_t n_embd_head, const int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_granite_swa : public llama_model_base { llama_model_granite_swa(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); private: ggml_tensor * build_attention_layer( ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv_iswa * inp_attn, const int64_t n_embd_head, const int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, ggml_tensor * inpSA, const int il); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_chameleon : public llama_model_base { llama_model_chameleon(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_wavtokenizer_dec : public llama_model_base { llama_model_wavtokenizer_dec(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_plm : public llama_model_base { llama_model_plm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_limite : public llama_model_base { llama_model_limite(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hrm_text : public llama_model_base { llama_model_hrm_text(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_stack( llm_graph_input_attn_kv * inp_attn, ggml_tensor * inp_pos, ggml_tensor * cur, int slot_base) const; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bailingmoe : public llama_model_base { llama_model_bailingmoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bailingmoe2 : public llama_model_base { llama_model_bailingmoe2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_bailingmoe3 : public llama_model_base { llama_model_bailingmoe3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); // causal conv1d over one of Q/K/V, qkv selects which third of the conv state row is used ggml_tensor * build_causal_conv1d(llm_graph_input_rs * inp_rs, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head, int64_t n_seq_tokens, int64_t n_seqs); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_seed_oss : public llama_model_base { llama_model_seed_oss(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_dots1 : public llama_model_base { llama_model_dots1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_arcee : public llama_model_base { llama_model_arcee(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_afmoe : public llama_model_base { llama_model_afmoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_beam : public llama_model_base { llama_model_beam(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_ricursive : public llama_model_base { llama_model_ricursive(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_laguna : public llama_model_base { llama_model_laguna(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_ernie4_5 : public llama_model_base { llama_model_ernie4_5(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_ernie4_5_moe : public llama_model_ernie4_5 { llama_model_ernie4_5_moe(const struct llama_model_params & params) : llama_model_ernie4_5(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_ernie4_5 struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_paddleocr : public llama_model_ernie4_5 { llama_model_paddleocr(const struct llama_model_params & params) : llama_model_ernie4_5(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_ernie4_5 struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hunyuan_moe : public llama_model_base { llama_model_hunyuan_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hy_v3 : public llama_model_base { llama_model_hy_v3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hy_v4 : public llama_model_base { llama_model_hy_v4(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); // iHC (independent Hyper-Connections): pre reduces the hc streams to one and returns the // per-stream post gates, post writes the sublayer output back into the streams, head // collapses the streams before the final norm. struct hc_pre { ggml_tensor * cur; // the collapsed stream ggml_tensor * post; // per-stream post gates, feeds build_hc_post }; hc_pre build_hc_pre( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base, int il) const; ggml_tensor * build_hc_post( ggml_tensor * x, ggml_tensor * residual, ggml_tensor * post, int il) const; ggml_tensor * build_hc_head( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base) const; ggml_tensor * build_attention( llm_graph_input_attn_k * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, float kq_scale, int il) const; // DSA lightning indexer: top-k KV positions for this layer. Only "full" layers compute // it, "shared" layers reuse the last preceding full layer result through last_top_k. ggml_tensor * build_indexer_top_k( llm_graph_input_attn_k_dsa * inp_attn_dsa, ggml_tensor * cur, ggml_tensor * qr, ggml_tensor * inp_pos, int il) const; // the attention output and the top-k to carry to the following "shared" layers struct attn_dsa { ggml_tensor * cur; ggml_tensor * top_k; }; attn_dsa build_attention_dsa( llm_graph_input_attn_k_dsa * inp_attn_dsa, ggml_tensor * cur, ggml_tensor * inp_pos, ggml_tensor * last_top_k, float kq_scale, int il) const; }; llama_rope_params get_rope_params_indexer(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hunyuan_vl : public llama_model_base { llama_model_hunyuan_vl(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_hunyuan_dense : public llama_model_hunyuan_vl { llama_model_hunyuan_dense(const struct llama_model_params & params) : llama_model_hunyuan_vl(params) {} // reuse load_arch_hparams and load_arch_tensors from llama_model_hunyuan_vl using graph = llama_model_hunyuan_vl::graph; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_smollm3 : public llama_model_base { llama_model_smollm3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_openai_moe : public llama_model_base { llama_model_openai_moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_falcon_h1 : public llama_model_base { llama_model_falcon_h1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_mamba_base { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_inkling2 : public llama_model_base { llama_model_inkling2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * mix(const std::vector & blocks, ggml_tensor * cur, ggml_tensor * query); ggml_tensor * clip(ggml_tensor * cur, ggml_tensor * scale, float threshold); ggml_tensor * convolve(ggml_tensor * cur, ggml_tensor * weight, ggml_tensor * dynamic, ggml_tensor * state, ggml_tensor * history, llm_graph_input_rs * inp, int64_t offset); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_lfm2 : public llama_model_base { llama_model_lfm2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_build_lfm_base { graph(const llm_graph_params & params); }; // non-causal trunk without memory, then the decision head struct graph_decision : public llm_graph_context { graph_decision(const llm_graph_params & params); ggml_tensor * build_decision_head( ggml_tensor * inp, llm_graph_input_attn_no_cache * inp_attn, ggml_tensor * inp_out_ids); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_lfm2moe : public llama_model_base { llama_model_lfm2moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template using graph = llama_model_lfm2::graph; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_lfm3 : public llama_model_base { llama_model_lfm3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; // decision: d1, the prompt read in one batch, no memory template struct graph : public llm_build_lfm_base { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_zaya1 : public llama_model_base { llama_model_zaya1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llama_model & model, const llm_graph_params & params); ggml_tensor * build_cca_block(const llama_model & model, ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_mem_hybrid * inp, int il); ggml_tensor * build_router_moe(const llama_model & model, ggml_tensor * cur, ggml_tensor *& router_state, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_smallthinker : public llama_model_base { llama_model_smallthinker(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; template struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_grovemoe : public llama_model_base { llama_model_grovemoe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_apertus : public llama_model_base { llama_model_apertus(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_minimax_01 : public llama_model_base { llama_model_minimax_01(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_minimax_m2 : public llama_model_base { llama_model_minimax_m2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct msa_params { int blk; int topk_blocks; int local; }; struct llama_model_minimax_m3 : public llama_model_base { llama_model_minimax_m3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; msa_params msa_p; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_attn_msa_fa( ggml_tensor * q_cur, // [D, HQ, S] f32 ggml_tensor * k, // [D, n_keys, 1, C] C = HKV or HKV*n_stream ggml_tensor * v, // [D, n_keys, 1, C] ggml_tensor * mask, // [n_keys, R, 1, C] f16, R = HQ*T/(Gp*C) int64_t Gp, float kq_scale, int il) const; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_cogvlm : public llama_model_base { llama_model_cogvlm(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_pangu_embed : public llama_model_base { llama_model_pangu_embed(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen3next : public llama_model_base { llama_model_qwen3next(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: ggml_tensor * build_layer_attn( llm_graph_input_attn_kv * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, int il); ggml_tensor * build_norm_gated( ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int layer); // returns pair of qkv, z std::pair build_qkvz( ggml_tensor * input, int il); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen35 : public llama_model_base { llama_model_qwen35(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: ggml_tensor * build_layer_attn( llm_graph_input_attn_kv * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, int il); ggml_tensor * build_norm_gated( ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int layer); // returns pair of qkv, z std::pair build_qkvz( ggml_tensor * input, int il); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; // Qwen3.5 backbone evaluated without memory, with a joint decision head on top of its hidden states struct llama_model_clef : public llama_model_qwen35 { llama_model_clef(const struct llama_model_params & params) : llama_model_qwen35(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct attn { ggml_tensor * wq = nullptr; ggml_tensor * bq = nullptr; ggml_tensor * wk = nullptr; ggml_tensor * bk = nullptr; ggml_tensor * wv = nullptr; ggml_tensor * bv = nullptr; ggml_tensor * wo = nullptr; ggml_tensor * bo = nullptr; }; struct norm { ggml_tensor * w = nullptr; ggml_tensor * b = nullptr; }; // the first n_layer_routing blocks have no self attention struct head_layer { norm self_norm; attn self_attn; norm cross_norm; norm cross_norm_kv; // routing blocks only attn cross_attn; norm ffn_norm; ggml_tensor * ffn_up = nullptr; ggml_tensor * ffn_up_b = nullptr; ggml_tensor * ffn_down = nullptr; ggml_tensor * ffn_down_b = nullptr; }; uint32_t n_layer_routing = 0; uint32_t n_layer_joint = 0; uint32_t n_head_decision = 0; std::vector head_layers; // routing blocks, then joint blocks norm hidden_norm; norm option_summary_norm; norm field_norm; norm option_norm; ggml_tensor * proj_memory = nullptr; ggml_tensor * proj_question = nullptr; ggml_tensor * proj_option_question = nullptr; ggml_tensor * proj_global = nullptr; ggml_tensor * proj_option_context = nullptr; ggml_tensor * proj_option_lexical = nullptr; ggml_tensor * scales = nullptr; // prior scale, joint scale, residual gate ggml_tensor * scorer = nullptr; ggml_tensor * scorer_b = nullptr; ggml_tensor * scorer_out = nullptr; ggml_tensor * scorer_out_b = nullptr; class input_decision; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: llm_graph_input_attn_no_cache * build_attn_inp_causal(); ggml_tensor * build_layer_attn( llm_graph_input_attn_no_cache * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear( ggml_tensor * cur, int il); // hidden: [n_embd, n_tokens], returns one score per option: [1, n_options] ggml_tensor * build_head( ggml_tensor * hidden, input_decision * inp); ggml_tensor * build_head_norm(ggml_tensor * cur, const norm & n); ggml_tensor * build_head_attn(ggml_tensor * q, ggml_tensor * kv, const attn & a); ggml_tensor * build_head_ffn (ggml_tensor * cur, const head_layer & layer); const llama_model_clef & model; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen4exp : public llama_model_base { llama_model_qwen4exp(const struct llama_model_params & params) : llama_model_base(params) {} class llm_graph_input_kpool; void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); protected: // the helpers alone, graph_mtp builds its own body struct no_build {}; graph(const llm_graph_params & params, no_build) : llm_build_delta_net_base(params) {} // HC replaces every layer norm: residual is [n_embd, hc, n_tokens] struct hc_mix { ggml_tensor * cur; // the mixed [n_embd, T] stream ggml_tensor * inject; // [hc, T] scatter weights, null when the inject weights are null }; hc_mix build_hc_mix( ggml_tensor * x, ggml_tensor * w_norm, ggml_tensor * w_down, ggml_tensor * w_up, ggml_tensor * w_inject, ggml_tensor * w_down_inject, // [w_down | w_inject] in one weight, can be null int il); ggml_tensor * build_hc_combine( ggml_tensor * residual, ggml_tensor * block_out, ggml_tensor * inject, int il); ggml_tensor * build_layer_attn( llm_graph_input_attn_kv * inp_attn, const llama_memory_hybrid_idx_context * mctx_hyb, llm_graph_input_kpool * inp_kpool, ggml_tensor * cur, ggml_tensor * inp_pos, int il); // dense self-attention over the cells the QSA mask keeps ggml_tensor * build_attn_qsa( llm_graph_input_attn_kv * inp, ggml_tensor * q_cur, ggml_tensor * k_cur, ggml_tensor * v_cur, ggml_tensor * sel, int64_t n_sel, float kq_scale, int il); // the QSA layers share one set of k-pool inputs, see llama_memory_hybrid_idx llm_graph_input_kpool * build_inp_kpool(const llama_memory_hybrid_idx_context * mctx_hyb); // QSA: the additive mask [n_kv, n_tokens] of the top blocks and the tail ggml_tensor * build_qsa_sel( const llama_memory_hybrid_idx_context * mctx_hyb, llm_graph_input_kpool * inp_kpool, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, int il); ggml_tensor * build_norm_gated( ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int layer); // build_rs writes the state tensor in place, so one gather per cache tensor is reused std::map rs_rows; // one conv history per cache tensor: delta-net and PLE each have their own ggml_tensor * build_conv_state_at( llm_graph_input_rs * inp, ggml_tensor * conv_states_all, ggml_tensor * x, int64_t state_cols, int64_t channels, int il); ggml_tensor * build_inp_ple( const llama_memory_hybrid_idx_context * mctx_hyb); ggml_tensor * build_ple( llm_graph_input_rs * inp, ggml_tensor * emb, ggml_tensor * hidden, int il); // returns pair of qkv, z std::pair build_qkvz( ggml_tensor * input, int il); }; // MTP draft head: one QSA block after the trunk, fed by the trunk's hc-wide residual struct graph_mtp : public graph { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_qwen35moe : public llama_model_base { llama_model_qwen35moe(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: ggml_tensor * build_layer_attn( llm_graph_input_attn_kv * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn( ggml_tensor * cur, int il); ggml_tensor * build_norm_gated( ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int layer); // returns pair of qkv, z std::pair build_qkvz( ggml_tensor * input, int il); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_clover_moe : public llama_model_base { llama_model_clover_moe(const llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); private: ggml_tensor * build_layer_attn(llm_graph_input_attn_kv * inp_attn, ggml_tensor * cur, ggml_tensor * inp_pos, int il); ggml_tensor * build_layer_attn_linear(llm_graph_input_rs * inp, ggml_tensor * cur, int il); ggml_tensor * build_layer_ffn(ggml_tensor * cur, int il); ggml_tensor * build_norm_gated(ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int il); std::pair build_qkvz(ggml_tensor * input, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mistral3 : public llama_model_base { llama_model_mistral3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; class input_t_cond; // voxtral realtime: time conditioning of the ada rms norm struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_mimo2 : public llama_model_base { llama_model_mimo2(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_kimi_k3 : public llama_model_base { llama_model_kimi_k3(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); // Cross-layer residual attention (K3's `_apply_attn_res`). ggml_tensor * resi_stack = nullptr; void res_push(ggml_tensor * cur, int64_t n_embd, int64_t n_tokens); ggml_tensor * res_mix(ggml_tensor * cur, ggml_tensor * score_w, int64_t n_tokens, int il); ggml_tensor * build_kda_layer(ggml_tensor * cur, const llama_layer & layer, llm_graph_input_rs * inp_rs, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t d_inner, int64_t n_seq_tokens, int64_t n_seqs, int il); // causal conv1d over one of Q/K/V, qkv selects which third of the conv state row is used ggml_tensor * build_kda_conv1d(llm_graph_input_rs * inp_rs, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t n_seq_tokens, int64_t n_seqs); ggml_tensor * build_mla_layer(ggml_tensor * cur, const llama_layer & layer, llm_graph_input_attn_k * inp_attn_k, llm_graph_input_attn_kv * inp_attn_kv, int64_t n_embd_head_k_mla, int64_t n_embd_head_v_mla, int64_t kv_lora_rank, int64_t n_embd_head_qk_rope, int64_t n_embd_head_qk_nope, float kq_scale, int il); ggml_tensor * build_latent_moe(ggml_tensor * cur, const llama_layer & layer, int64_t n_embd_latent, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; // Kimi-K3.1: Kimi-K3 with a lightning indexer and top-k sparse attention on the full-attention layers struct llama_model_kimi_k31 : public llama_model_base { llama_model_kimi_k31(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); // Cross-layer residual attention (kimi-k3's `_apply_attn_res`). ggml_tensor * resi_stack = nullptr; void res_push(ggml_tensor * cur, int64_t n_embd, int64_t n_tokens); ggml_tensor * res_mix(ggml_tensor * cur, ggml_tensor * score_w, int64_t n_tokens, int il); ggml_tensor * build_kda_conv1d(llm_graph_input_rs * inp_rs, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t n_seq_tokens, int64_t n_seqs); ggml_tensor * build_kda_layer(ggml_tensor * cur, const llama_layer & layer, llm_graph_input_rs * inp_rs, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t d_inner, int64_t n_seq_tokens, int64_t n_seqs, int il); struct dsa_out { ggml_tensor * out; ggml_tensor * sel; }; dsa_out build_dsa_layer(ggml_tensor * cur, const llama_layer & layer, const llama_kv_cache_context * mctx_idx, llm_graph_input_attn_k * inp_attn, ggml_tensor * prev_sel, int64_t n_embd_head_k_mla, int64_t kv_lora_rank, int64_t n_embd_head_qk_rope, int64_t n_embd_head_qk_nope, float kq_scale, int il); ggml_tensor * build_latent_moe(ggml_tensor * cur, const llama_layer & layer, int64_t n_embd_latent, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_glm5_next : public llama_model_base { llama_model_glm5_next(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; // k-pool indexer inputs on top of the generic hybrid input class llm_graph_input_kpool; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); // build the helpers without the trunk, so graph_mtp can reuse them struct no_build {}; graph(const llm_graph_params & params, no_build); // collapse the hc streams with per-stream weights ggml_tensor * build_hc_pre( ggml_tensor * x, ggml_tensor * weights, int il) const; // mean over the hyper-connection streams: [n_embd, hc, n_tokens] -> [n_embd, n_tokens] ggml_tensor * build_hc_mean(ggml_tensor * x) const; // all outputs of the hc mix builder: cur is the collapsed input, post / comb feed build_hc_post struct hc_pre { ggml_tensor * cur; ggml_tensor * post; ggml_tensor * comb; }; hc_pre build_hc_pre( ggml_tensor * x, ggml_tensor * hc_fn, ggml_tensor * hc_scale, ggml_tensor * hc_base, int il) const; ggml_tensor * build_hc_post( ggml_tensor * x, ggml_tensor * residual, ggml_tensor * post, ggml_tensor * comb, int il) const; ggml_tensor * build_hc_sinkhorn( ggml_tensor * comb, int il) const; llm_graph_input_kpool * build_inp_kpool(const llama_memory_hybrid_idx_context * mctx_hyb); ggml_tensor * build_kda_layer(ggml_tensor * cur, const llama_layer & layer, llm_graph_input_rs * inp_rs, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t d_inner, int64_t n_seq_tokens, int64_t n_seqs, int il); // causal conv1d over one of Q/K/V, qkv selects which third of the conv state row is used ggml_tensor * build_kda_conv1d(llm_graph_input_rs * inp_rs, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head_kda, int64_t n_seq_tokens, int64_t n_seqs); ggml_tensor * build_kpool_select(ggml_tensor * cur, ggml_tensor * qr, const llama_layer & layer, const llama_memory_hybrid_idx_context * mctx_hyb, llm_graph_input_kpool * inp_kpool, int il); // attention output (cur) and the selection to carry to the next shared layer (sel) struct dsa_out { ggml_tensor * cur; ggml_tensor * sel; }; dsa_out build_dsa_layer(ggml_tensor * cur, const llama_layer & layer, const llama_memory_hybrid_idx_context * mctx_hyb, llm_graph_input_attn_k * inp_attn, llm_graph_input_kpool * inp_kpool, ggml_tensor * prev_sel, int il); }; // the NextN / MTP draft head: one DSA block appended after the trunk struct graph_mtp : public graph { graph_mtp(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_kimi_linear : public llama_model_base { llama_model_kimi_linear(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_build_delta_net_base { graph(const llm_graph_params & params); std::pair build_kda_autoregressive( ggml_tensor * q, ggml_tensor * k, ggml_tensor * v, ggml_tensor * gk, ggml_tensor * beta, ggml_tensor * state, int il); std::pair build_kda_chunking( ggml_tensor * q, ggml_tensor * k, ggml_tensor * v, ggml_tensor * gk, ggml_tensor * beta, ggml_tensor * state, ggml_tensor * causal_mask, ggml_tensor * identity, ggml_tensor * diag_mask, int il); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_step35 : public llama_model_base { llama_model_step35(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; struct graph_mtp : public llm_graph_context { graph_mtp(const llm_graph_params & params); }; llama_rope_params get_rope_params(const llama_cparams & cparams, int il) const override; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_spark2_5 : public llama_model_base { llama_model_spark2_5(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_k2_horizon : public llama_model_base { llama_model_k2_horizon(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); ggml_tensor * build_routed_value(const llama_layer & layer, ggml_tensor * cur, int il) const; }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; }; struct llama_model_kolibri1 : public llama_model_base { llama_model_kolibri1(const struct llama_model_params & params) : llama_model_base(params) {} void load_arch_hparams(llama_model_loader & ml) override; void load_arch_tensors(llama_model_loader & ml) override; struct graph : public llm_graph_context { graph(const llm_graph_params & params); }; std::unique_ptr build_arch_graph(const llm_graph_params & params) const override; };