#include "models.h" #include "llama-memory-recurrent.h" void llama_model_clover_moe::load_arch_hparams(llama_model_loader & ml) { if (hparams.n_embd_head_k() != hparams.n_embd_head_v()) { throw std::runtime_error("Clover requires equal key and value head sizes"); } 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); if (hparams.n_expert_shared > 0) { ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp); } ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) { uint32_t full_attn_interval; ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval); if (full_attn_interval == 0) { throw std::runtime_error("Clover requires a positive full attention interval"); } for (uint32_t i = 0; i < hparams.n_layer_all; ++i) { hparams.is_recr_impl[i] = ((i + 1) % full_attn_interval != 0); } } bool has_full = false; bool has_linear = false; for (uint32_t i = 0; i < hparams.n_layer_all; ++i) { has_full |= !hparams.is_recr(i); has_linear |= hparams.is_recr(i); } if (!has_full || !has_linear) { throw std::runtime_error("Clover requires both full and linear attention layers"); } if (hparams.ssm_dt_rank == 0 || hparams.ssm_n_group == 0 || hparams.ssm_d_state == 0 || hparams.ssm_dt_rank % hparams.ssm_n_group != 0 || hparams.ssm_d_inner != uint64_t(hparams.ssm_d_state) * hparams.ssm_dt_rank || hparams.ssm_d_conv == 0) { throw std::runtime_error("Clover has invalid gated delta net dimensions"); } switch (hparams.n_layer()) { case 40: type = LLM_TYPE_35B_A3B; break; default: type = LLM_TYPE_UNKNOWN; } } void llama_model_clover_moe::load_arch_tensors(llama_model_loader &) { LLAMA_LOAD_LOCALS; tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED); if (output == NULL) { output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED); } for (int il = 0; il < n_layer; ++il) { auto & layer = layers[il]; const int64_t n_ff_exp = hparams.n_ff_exp(il); const int64_t n_ff_shexp = hparams.n_ff_shexp; const int64_t head_k_dim = hparams.ssm_d_state; const int64_t head_v_dim = hparams.ssm_d_state; const int64_t n_k_heads = hparams.ssm_n_group; const int64_t n_v_heads = hparams.ssm_dt_rank; const int64_t key_dim = head_k_dim * n_k_heads; const int64_t value_dim = head_v_dim * n_v_heads; const int64_t conv_dim = key_dim * 2 + value_dim; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", il), { n_embd }, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, 0); if (!hparams.is_recr(il)) { create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, 0); layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, 0); } else { layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", il), { n_embd, key_dim * 2 + value_dim }, 0); layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", il), { n_embd, value_dim }, 0); layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", il), { hparams.ssm_d_conv, conv_dim }, 0); layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", il), { hparams.ssm_dt_rank }, 0); layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, il), { hparams.ssm_dt_rank }, 0); layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", il), { n_embd, n_v_heads }, 0); layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", il), { n_embd, n_v_heads }, 0); layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", il), { head_v_dim }, 0); layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", il), { value_dim, n_embd }, 0); } layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", il), { n_embd, n_expert }, 0); layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, 0); create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, 0); if (hparams.n_expert_shared > 0) { layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, 0); layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", il), { n_embd, n_ff_shexp }, 0); layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", il), { n_embd, n_ff_shexp }, 0); layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", il), { n_ff_shexp, n_embd }, 0); } } } std::unique_ptr llama_model_clover_moe::build_arch_graph(const llm_graph_params & params) const { return std::make_unique(*this, params); } llama_model_clover_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_delta_net_base(params), model(model) { 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; inpL = build_inp_embd(model.tok_embd); cb(inpL, "model.input_embed", -1); auto * inp = build_inp_mem_hybrid(); ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); for (int il = 0; il < n_layer; ++il) { res->t_layer_inp[il] = inpL; ggml_tensor * inpSA = inpL; cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); ggml_build_forward_expand(gf, cur); if (hparams.is_recr(il)) { cur = build_layer_attn_linear(inp->get_recr(), cur, il); } else { cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); } if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } cur = ggml_add(ctx0, cur, inpSA); cb(cur, "attn_residual", il); ggml_tensor * ffn_residual = cur; ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); cb(attn_post_norm, "attn_post_norm", il); cur = build_layer_ffn(attn_post_norm, il); cb(cur, "ffn_out", il); cur = ggml_add(ctx0, cur, ffn_residual); cb(cur, "post_moe", il); cur = build_cvec(cur, il); cb(cur, "l_out", il); inpL = cur; } cur = inpL; cur = build_norm(cur, model.output_norm, nullptr, 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::pair llama_model_clover_moe::graph::build_qkvz( ggml_tensor * input, int il) { const int64_t n_seqs = ubatch.n_seqs; const int64_t n_seq_tokens = ubatch.n_seq_tokens; ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input, model.layers[il].wqkv_s); qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs); cb(qkv_mixed, "linear_attn_qkv_mixed", il); ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input, model.layers[il].wqkv_gate_s); cb(z, "z", il); return { qkv_mixed, z }; } ggml_tensor * llama_model_clover_moe::graph::build_norm_gated( ggml_tensor * input, ggml_tensor * weights, ggml_tensor * gate, int il) { auto * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, il); return ggml_mul(ctx0, normalized, ggml_silu(ctx0, gate)); } ggml_tensor * llama_model_clover_moe::graph::build_layer_attn( llm_graph_input_attn_kv * inp, ggml_tensor * cur, ggml_tensor * inp_pos, int il) { const int64_t n_embd_head = hparams.n_embd_head_v(); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k()); auto [Qcur_full, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head * 2, n_head, n_embd_head, n_head_kv, n_embd_head, n_head_kv, il, false); cb(Qcur_full, "Qcur_full", il); ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, ggml_element_size(Qcur_full) * n_embd_head * 2, ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0); cb(Qcur, "Qcur_reshaped", il); Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); cb(Qcur, "Qcur_normed", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); cb(Kcur, "Kcur_normed", il); ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, ggml_element_size(Qcur_full) * n_embd_head * 2, ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, ggml_element_size(Qcur_full) * n_embd_head); gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); cb(gate, "gate_reshaped", il); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; cur = build_attn(inp, nullptr, nullptr, nullptr, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "attn_pregate", il); ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); cb(gate_sigmoid, "gate_sigmoid", il); cur = ggml_mul(ctx0, cur, gate_sigmoid); cb(cur, "attn_gated", il); cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s); cb(cur, "attn_output", il); return cur; } ggml_tensor * llama_model_clover_moe::graph::build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, int il) { const auto * mctx_cur = inp->mctx; const int64_t d_inner = hparams.ssm_d_inner; const int64_t n_seqs = ubatch.n_seqs; const int64_t head_k_dim = hparams.ssm_d_state; const int64_t num_k_heads = hparams.ssm_n_group; const int64_t num_v_heads = hparams.ssm_dt_rank; const int64_t head_v_dim = d_inner / num_v_heads; const int64_t n_seq_tokens = ubatch.n_seq_tokens; GGML_ASSERT(n_seqs != 0); GGML_ASSERT(ubatch.equal_seqs()); GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); auto qkvz = build_qkvz(cur, il); ggml_tensor * qkv_mixed = qkvz.first; ggml_tensor * z = qkvz.second; ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur, model.layers[il].ssm_beta_s); ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur, model.layers[il].ssm_alpha_s); beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs); cb(beta, "beta", il); beta = ggml_sigmoid(ctx0, beta); cb(beta, "beta_sigmoid", il); alpha = ggml_reshape_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs); cb(alpha, "alpha", il); ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); cb(alpha_softplus, "a_softplus", il); ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); cb(gate, "gate", il); gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs); ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; const int64_t conv_kernel_size = conv_kernel->ne[0]; const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; ggml_tensor * conv_input = build_conv_state(inp, conv_states_all, qkv_mixed, conv_kernel_size, conv_channels, il); ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs); cb(state, "state_predelta", il); ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); cb(conv_output_proper, "conv_output_raw", il); ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); cb(conv_output_silu, "conv_output_silu", il); ggml_tensor * conv_qkv_mix = conv_output_silu; int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, ggml_row_size(conv_qkv_mix->type, head_k_dim), nb1_qkv, nb1_qkv * n_seq_tokens, 0); ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, ggml_row_size(conv_qkv_mix->type, head_k_dim), nb1_qkv, nb1_qkv * n_seq_tokens, head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs, ggml_row_size(conv_qkv_mix->type, head_v_dim), nb1_qkv, nb1_qkv * n_seq_tokens, ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads)); cb(q_conv, "q_conv", il); cb(k_conv, "k_conv", il); cb(v_conv, "v_conv", il); q_conv = build_gdn_l2_norm(ctx0, q_conv, 1e-6f); k_conv = build_gdn_l2_norm(ctx0, k_conv, 1e-6f); if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) { q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); } ggml_tensor * output = build_recurrent_attn(inp, ssm_states_all, q_conv, k_conv, v_conv, gate, beta, state, il); ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il); ggml_tensor * final_output = ggml_reshape_2d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens * n_seqs); cb(final_output, "final_output", il); cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s); cb(cur, "linear_attn_out", il); return cur; } ggml_tensor * llama_model_clover_moe::graph::build_layer_ffn(ggml_tensor * cur, const int il) { GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr); ggml_tensor * moe_out = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, nullptr, n_expert, n_expert_used, LLM_FFN_SILU, true, hparams.expert_weights_scale, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il, nullptr, model.layers[il].ffn_gate_up_exps, model.layers[il].ffn_up_exps_s, model.layers[il].ffn_gate_exps_s, model.layers[il].ffn_down_exps_s); cb(moe_out, "ffn_moe_out", il); if (model.layers[il].ffn_up_shexp != nullptr) { ggml_tensor * ffn_shexp = build_ffn(cur, model.layers[il].ffn_up_shexp, NULL, model.layers[il].ffn_up_shexp_s, model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s, model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); cb(ffn_shexp, "ffn_shexp", il); ggml_tensor * shared_gate; shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); cb(shared_gate, "shared_expert_gate", il); shared_gate = ggml_sigmoid(ctx0, shared_gate); cb(shared_gate, "shared_expert_gate_sigmoid", il); ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); cb(ffn_shexp, "ffn_shexp_gated", il); cur = ggml_add(ctx0, moe_out, ffn_shexp); cb(cur, "ffn_out", il); } else { cur = moe_out; } return cur; }