Brunobkr/llama.cpp_AlgMor24_github

1// Laguna (poolside): sigmoid-routed MoE with a score-correction bias, one shared2// expert, a softplus attention output gate, QK-norm, and per-layer-type RoPE3// (YaRN on full-attention layers, plain RoPE on sliding-window layers). XS.2 is4// hybrid full/SWA with a per-head gate; M.1 is full-attention with a per-element5// gate. Shares the MoE/gate structure with afmoe.6 7#include "models.h"8 9void llama_model_laguna::load_arch_hparams(llama_model_loader & ml) {10    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);11    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);12    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);13    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);14    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);15    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);16 17    // Laguna ships one shared expert and stores its size directly (routed and18    // shared experts may differ), so read the size from expert_shared_feed_forward_length.19    // The count is not in the config; default to 1 but read the key if present.20    hparams.n_expert_shared = 1;21    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);22    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);23    if (hparams.n_ff_shexp == 0) {24        // Weightless fixtures (test-llama-archs) omit this key; derive a nonzero25        // size so the shared expert is still built. Real GGUFs always carry the26        // exact value (routed and shared FF lengths may differ).27        hparams.n_ff_shexp = hparams.n_ff_exp * hparams.n_expert_shared;28    }29 30    // Sliding-window attention is OPTIONAL. XS.2 is hybrid (full / SWA / SWA /31    // SWA repeating, period 4 starting with full); M.1 has no sliding window32    // (all layers full attention). When sliding_window is absent or zero we33    // leave swa_type = NONE and skip the SWA-specific per-layer-type RoPE.34    hparams.n_swa = 0;35    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);36    if (hparams.n_swa > 0) {37        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;38 39        uint32_t swa_period = 4;40        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);41        hparams.set_swa_pattern(swa_period, /*dense_first=*/true);  // XS.2: FULL at il%4==042 43        // Per-layer-type RoPE: full layers use YaRN θ=500000 over 64 dims;44        // SWA layers use default RoPE θ=10000 over 128 dims. Base load_hparams45        // already reads ROPE_FREQ_BASE and ROPE_DIMENSION_COUNT into the46        // non-SWA fields; we explicitly pull the SWA mirrors here.47        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;48        hparams.rope_freq_scale_train_swa = 1.0f;  // SWA uses plain RoPE (no YaRN scaling); do NOT inherit full layers 1/factor49        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);50        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT_SWA, hparams.n_rot_swa, false);51    }52 53    // Default the expert gating function to SIGMOID when the key is absent54    // (matches the HF reference).55    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {56        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;57    }58 59    switch (hparams.n_layer()) {60        case 40: type = LLM_TYPE_30B_A3B;   break;  // Laguna-XS.261        case 48: type = LLM_TYPE_118B_A8B;  break;  // Laguna-S.262        case 70: type = LLM_TYPE_230B_A10B; break;  // Laguna-M.163        default: type = LLM_TYPE_UNKNOWN;64    }65}66 67void llama_model_laguna::load_arch_tensors(llama_model_loader & ml) {68    LLAMA_LOAD_LOCALS;69 70    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);71 72    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);73    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);74    if (output == NULL) {75        // tied embeddings fallback76        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);77    }78 79    const int64_t n_ff_exp   = hparams.n_ff_exp;80    const int64_t n_ff_shexp = hparams.n_ff_shexp;81 82    for (int i = 0; i < n_layer; ++i) {83        auto & layer = layers[i];84 85        // Per-layer head count — Laguna varies n_head between full and SWA86        // layers (48 vs 64 in XS.2). KV head count is uniform.87        const int64_t n_head_il    = hparams.n_head(i);88        const int64_t n_head_kv_il = hparams.n_head_kv(i);89        const int64_t n_embd_q_il  = n_embd_head_k * n_head_il;90        const int64_t n_embd_k_il  = n_embd_head_k * n_head_kv_il;91        const int64_t n_embd_v_il  = n_embd_head_v * n_head_kv_il;92 93        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);94 95        create_tensor_qkv(layer, i, n_embd, n_embd_q_il, n_embd_k_il, n_embd_v_il, 0);96        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q_il, n_embd}, 0);97 98        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);99        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);100 101        // Attention output gate. XS.2 is per-head (g_proj -> n_head, one scalar102        // per head broadcast over head_dim at multiply time); M.1 is per-element103        // (g_proj -> n_head*head_dim, like afmoe). Detect from the stored tensor104        // shape so a single arch handles both; the graph mirrors this check.105        // Gate width selects per-head vs per-element. Real GGUFs always carry the106        // gate tensor, so read the width from it and require EXACTLY one of the two107        // valid widths -- never guess between them. Weightless fixtures108        // (test-llama-archs) have no gate tensor; fall back to the per-head layout so109        // the per-head reshape path is still exercised.110        const int64_t n_gate_per_head = n_head_il;111        const int64_t n_gate_per_elem = n_embd_head_k * n_head_il;112        const ggml_tensor * gate_meta = ml.get_tensor_meta(tn(LLM_TENSOR_ATTN_GATE, "weight", i).str().c_str());113        int64_t n_gate_out;114        if (gate_meta != nullptr) {115            n_gate_out = gate_meta->ne[1];116            if (n_gate_out != n_gate_per_head && n_gate_out != n_gate_per_elem) {117                GGML_ABORT("Laguna: unexpected attention gate width %lld at layer %d "118                           "(expected %lld per-head or %lld per-element)",119                           (long long) n_gate_out, i, (long long) n_gate_per_head, (long long) n_gate_per_elem);120            }121        } else {122            n_gate_out = n_gate_per_head;123        }124        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_gate_out}, 0);125 126        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);127 128        if ((uint32_t)i >= hparams.n_layer_dense_lead) {129            // MoE layer130            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, 0);131            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, 0);132 133            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);134            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);135            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd,   n_expert}, 0);136 137            // Always-on shared expert.138            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd,    n_ff_shexp}, 0);139            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd,    n_ff_shexp}, 0);140            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd},    0);141        } else {142            // Dense layer (the leading n_layer_dense_lead layers)143            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);144            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);145            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff,   n_embd}, 0);146        }147    }148}149 150std::unique_ptr<llm_graph_context> llama_model_laguna::build_arch_graph(const llm_graph_params & params) const {151    return std::make_unique<graph>(*this, params);152}153 154llama_model_laguna::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {155    const int64_t n_embd_head = hparams.n_embd_head_v();156    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());157 158    ggml_tensor * cur;159    ggml_tensor * inpL;160 161    inpL = build_inp_embd(model.tok_embd);162    // No MuP embedding scale (laguna omits this; afmoe scales by sqrt(hidden)).163 164    ggml_tensor * inp_pos = build_inp_pos();165    // XS.2 is hybrid SWA -> interleaved-SWA KV input; M.1 is all-full -> plain166    // KV input. Pick the matching input (and build_attn overload) per swa_type.167    const bool has_swa = hparams.swa_type != LLAMA_SWA_TYPE_NONE;168    llm_graph_input_attn_kv      * inp_attn_kv   = has_swa ? nullptr : build_attn_inp_kv();169    llm_graph_input_attn_kv_iswa * inp_attn_iswa = has_swa ? build_attn_inp_kv_iswa() : nullptr;170    ggml_tensor * inp_out_ids = build_inp_out_ids();171 172    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));173 174    for (int il = 0; il < n_layer; ++il) {175        const bool    is_swa_il   = hparams.is_swa(il);176        const int64_t n_head_il   = hparams.n_head(il);177        const int64_t n_head_kv_il = hparams.n_head_kv(il);178 179        // Per-layer-type RoPE config. SWA layers run plain rope (no YaRN),180        // achieved by zeroing the YaRN ext/beta params for those layers.181        const int   n_rot_l       = is_swa_il ? hparams.n_rot_swa : n_rot;182        const float freq_base_l   = is_swa_il ? hparams.rope_freq_base_train_swa : freq_base;183        const float freq_scale_l  = is_swa_il ? hparams.rope_freq_scale_train_swa : freq_scale;184        const float ext_factor_l  = is_swa_il ? 0.0f : ext_factor;185        // YaRN magnitude scaling (mscale) is already handled by the framework:186        // llama_context pre-divides cparams.yarn_attn_factor by (1 + 0.1*ln(factor))187        // to cancel ggml rope_yarn's internal mscale *= 1 + 0.1*ln(1/freq_scale).188        // Pass attn_factor straight through (like every other arch); SWA layers run189        // plain RoPE (ext_factor 0, no mscale) so force 1.0 there.190        const float attn_factor_l = is_swa_il ? 1.0f : attn_factor;191        const float beta_fast_l   = is_swa_il ? 0.0f : beta_fast;192        const float beta_slow_l   = is_swa_il ? 0.0f : beta_slow;193        const int   n_ctx_orig_l  = is_swa_il ? hparams.n_ctx_train : n_ctx_orig;194 195        ggml_tensor * inpSA = inpL;196 197        // Pre-norm198        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);199        cb(cur, "attn_norm", il);200 201        // Self-attention202        {203            ggml_tensor * attn_inp = cur;  // saved for the gate projection204 205            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,206                    n_embd_head, n_head_il, n_head_kv_il, il);207 208            // g_proj on the *pre-attention* hidden state (matches HF209            // reference: gate is computed from the same `hidden_states`210            // input as q/k/v, not from the attn output).211            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);212            cb(gate, "attn_gate_proj", il);213 214            // QK RMSNorm at head_dim level (Qwen3 style)215            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);216            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);217            cb(Qcur, "Qcur_normed", il);218            cb(Kcur, "Kcur_normed", il);219 220            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,221                    n_rot_l, rope_type, n_ctx_orig_l, freq_base_l, freq_scale_l,222                    ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);223            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,224                    n_rot_l, rope_type, n_ctx_orig_l, freq_base_l, freq_scale_l,225                    ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);226            cb(Qcur, "Qcur_rope", il);227            cb(Kcur, "Kcur_rope", il);228 229            cur = has_swa230                ? build_attn(inp_attn_iswa,231                        NULL, NULL, NULL,    // o_proj deferred until after gating232                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il)233                : build_attn(inp_attn_kv,234                        NULL, NULL, NULL,235                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);236            cb(cur, "attn_out", il);237 238            // Softplus output gate (the unary kernel computes softplus in fp32239            // and casts back). Two shapes, distinguished by the g_proj output240            // dim (matching the load-time detection):241            //   XS.2 per-head     : gate [n_head_il, n_tokens] -> reshape to242            //                       [1, n_head_il, n_tokens] and broadcast over243            //                       head_dim against cur [head_dim, n_head, T].244            //   M.1  per-element  : gate [n_head_il*head_dim, n_tokens] spans the245            //                       full attention output -> direct ggml_mul.246            gate = ggml_softplus(ctx0, gate);247            cb(gate, "attn_gate_softplus", il);248 249            const int64_t n_tokens = cur->ne[1];250            if (model.layers[il].wqkv_gate->ne[1] == n_head_il) {251                cur  = ggml_reshape_3d(ctx0, cur,  n_embd_head, n_head_il, n_tokens);252                gate = ggml_reshape_3d(ctx0, gate, 1,           n_head_il, n_tokens);253                cur  = ggml_mul(ctx0, cur, gate);254                cur  = ggml_reshape_2d(ctx0, cur, n_embd_head * n_head_il, n_tokens);255            } else {256                cur = ggml_mul(ctx0, cur, gate);257            }258            cb(cur, "attn_gated", il);259 260            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);261            cb(cur, "attn_o_proj", il);262        }263 264        if (il == n_layer - 1 && inp_out_ids) {265            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);266            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);267        }268 269        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);270        cb(ffn_inp, "ffn_inp", il);271 272        // Pre-norm only (no post-attn norm)273        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);274        cb(cur, "ffn_norm", il);275 276        if ((uint32_t)il >= hparams.n_layer_dense_lead) {277            // MoE: sigmoid routing + score-correction bias + sum-norm +278            // routed_scaling_factor (all handled by build_moe_ffn).279            ggml_tensor * moe_out = build_moe_ffn(cur,280                    model.layers[il].ffn_gate_inp,281                    model.layers[il].ffn_up_exps,282                    model.layers[il].ffn_gate_exps,283                    model.layers[il].ffn_down_exps,284                    model.layers[il].ffn_exp_probs_b,285                    n_expert, n_expert_used,286                    LLM_FFN_SILU,287                    hparams.expert_weights_norm,288                    hparams.expert_weights_scale,289                    (llama_expert_gating_func_type) hparams.expert_gating_func,290                    il);291            cb(moe_out, "ffn_moe_out", il);292 293            // Always-on shared expert, summed in parallel.294            ggml_tensor * ffn_shexp = build_ffn(cur,295                    model.layers[il].ffn_up_shexp,   NULL, NULL,296                    model.layers[il].ffn_gate_shexp, NULL, NULL,297                    model.layers[il].ffn_down_shexp, NULL, NULL,298                    NULL,299                    LLM_FFN_SILU, LLM_FFN_PAR, il);300            cb(ffn_shexp, "ffn_shexp", il);301 302            cur = ggml_add(ctx0, moe_out, ffn_shexp);303            cb(cur, "ffn_out", il);304        } else {305            // Dense FFN for the leading n_layer_dense_lead layers (XS.2: 1, M.1: 3)306            cur = build_ffn(cur,307                    model.layers[il].ffn_up,   NULL, NULL,308                    model.layers[il].ffn_gate, NULL, NULL,309                    model.layers[il].ffn_down, NULL, NULL,310                    NULL,311                    LLM_FFN_SILU, LLM_FFN_PAR, il);312            cb(cur, "ffn_out", il);313        }314 315        // No post-ffn norm316        cur = ggml_add(ctx0, cur, ffn_inp);317        cur = build_cvec(cur, il);318        cb(cur, "l_out", il);319 320        inpL = cur;321    }322 323    cur = inpL;324    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);325    cb(cur, "result_norm", -1);326    res->t_embd = cur;327 328    cur = build_lora_mm(model.output, cur);329    cb(cur, "result_output", -1);330    res->t_logits = cur;331 332    ggml_build_forward_expand(gf, cur);333}334