#include "models.h" void llama_model_kolibri1::load_arch_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all); ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm); ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func); ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); if (hparams.n_swa == 0 || hparams.n_expert_shared != 1 || hparams.n_ff_shexp == 0) { throw std::runtime_error("Kolibri-1: invalid sliding window or shared expert configuration"); } hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl); hparams.rope_pattern = hparams.is_swa_impl; hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train; hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train; switch (hparams.n_layer()) { case 50: type = LLM_TYPE_78B_A3_5B; break; default: type = LLM_TYPE_UNKNOWN; } } void llama_model_kolibri1::load_arch_tensors(llama_model_loader &) { LLAMA_LOAD_LOCALS; const int64_t n_ff_exp = hparams.n_ff_exp(); const int64_t n_ff_shexp = hparams.n_ff_shexp; tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); // Routing is sigmoid, the selection bias is added to the logits before top-k. layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0); layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0); layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0); layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0); layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0); layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0); layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0); } } llama_model_kolibri1::graph::graph(const llm_graph_params & params) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); ggml_tensor * cur; ggml_tensor * inpL = build_inp_embd(model.tok_embd); ggml_tensor * inp_pos = build_inp_pos(); auto * inp_attn = build_attn_inp_kv_iswa(); ggml_tensor * inp_out_ids = build_inp_out_ids(); const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); for (int il = 0; il < n_layer; ++il) { const auto & layer = model.layers[il]; ggml_tensor * residual = inpL; cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); // QK-norm, then RoPE on the sliding layers only: the global layers are NoPE. auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur, n_embd_head, n_head, n_head_kv, il); Qcur = build_norm(Qcur, layer.attn_q_norm, NULL, LLM_NORM_RMS, il); Kcur = build_norm(Kcur, layer.attn_k_norm, NULL, LLM_NORM_RMS, il); cb(Qcur, "Qcur_normed", il); cb(Kcur, "Kcur_normed", il); Qcur = build_rope(Qcur, inp_pos, il); cb(Qcur, "Qcur_rope", il); Kcur = build_rope(Kcur, inp_pos, il); cb(Kcur, "Kcur_rope", il); cur = build_attn(inp_attn, layer.wo, NULL, layer.wo_s, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "attn_out", il); cur = build_norm(cur, layer.attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); residual = ggml_get_rows(ctx0, residual, inp_out_ids); } ggml_tensor * ffn_inp = ggml_add(ctx0, cur, residual); cb(ffn_inp, "ffn_inp", il); cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "ffn_norm", il); // The selection adds the bias to the logits, the weights are the sigmoid of the unbiased logits, so the same logits feed both. ggml_tensor * logits = build_lora_mm(layer.ffn_gate_inp, cur); cb(logits, "ffn_moe_logits", il); ggml_tensor * selection = ggml_add(ctx0, logits, layer.ffn_exp_probs_b); cb(selection, "ffn_moe_logits_biased", il); ggml_tensor * selected = ggml_argsort_top_k(ctx0, selection, n_expert_used); cb(selected, "ffn_moe_topk", il); ggml_tensor * moe_out = build_moe_ffn(cur, layer.ffn_gate_inp, layer.ffn_up_exps, layer.ffn_gate_exps, layer.ffn_down_exps, nullptr, n_expert, n_expert_used, LLM_FFN_SILU, hparams.expert_weights_norm, 1.0f, (llama_expert_gating_func_type) hparams.expert_gating_func, il, logits, nullptr, nullptr, nullptr, nullptr, selected); cb(moe_out, "ffn_moe_out", il); ggml_tensor * shared = build_ffn(cur, layer.ffn_up_shexp, nullptr, nullptr, layer.ffn_gate_shexp, nullptr, nullptr, layer.ffn_down_shexp, nullptr, nullptr, nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(shared, "ffn_shexp_out", il); cur = ggml_add(ctx0, moe_out, shared); cur = build_norm(cur, layer.ffn_post_norm, NULL, LLM_NORM_RMS, il); cur = ggml_add(ctx0, cur, ffn_inp); inpL = build_cvec(cur, il); cb(inpL, "l_out", il); } cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); cb(cur, "result_norm", -1); res->t_embd = cur; cur = build_lora_mm(model.output, cur, model.output_s); cb(cur, "result_output", -1); res->t_logits = cur; } std::unique_ptr llama_model_kolibri1::build_arch_graph(const llm_graph_params & params) const { return std::make_unique(params); }