Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_muse_glimmer::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);6    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);7    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);8 9    hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train;10    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);11 12    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;13    load_swa_pattern(ml, 4);14 15    switch (hparams.n_layer()) {16        case 52: type = LLM_TYPE_30B; break;17        default: type = LLM_TYPE_UNKNOWN;18    }19}20 21void llama_model_muse_glimmer::load_arch_tensors(llama_model_loader &) {22    LLAMA_LOAD_LOCALS;23 24    tok_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);25    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);26    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);27 28    for (int i = 0; i < n_layer; ++i) {29        auto & layer = layers[i];30 31        // Pre/post-attention norms (Muse Glimmer's `weight + 1` applied at conversion time).32        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);33        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);34 35        // Q/K/V/O projections.36        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);37        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);38 39        // QK-norm. Weights are synthesized at conversion time to absorb `qk_scale_factor`.40        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);41        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);42 43        // Attention output gate: sigmoid(gate) * attn_out before o_proj (same as afmoe).44        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);45 46        // Pre/post-FFN norms (FFN_PRE_NORM is aliased to LLM_TENSOR_FFN_NORM).47        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);48        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);49 50        // Dense FFN (unlike afmoe, no MoE branches).51        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);52        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);53        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);54    }55}56 57llama_model_muse_glimmer::graph::graph(const llama_model & model, const llm_graph_params & params)58    : llm_graph_context(params) {59    const int64_t n_embd_head = hparams.n_embd_head_v();60    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());61 62    // Different to f_norm_rms_eps for post-attn / post-FFN norms63    const float post_norm_eps = 1e-8f;64 65    ggml_tensor * cur;66    ggml_tensor * inpL;67 68    inpL = build_inp_embd(model.tok_embd);69    inpL = build_norm(inpL, nullptr, nullptr, LLM_NORM_RMS, -1);70    cb(inpL, "embd_norm", -1);71 72    ggml_tensor * inp_pos = build_inp_pos();73    auto * inp_attn = build_attn_inp_kv_iswa();74    ggml_tensor * inp_out_ids = build_inp_out_ids();75 76    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));77 78    for (int il = 0; il < n_layer; ++il) {79        // expose per-layer residual for speculative drafts (see LLM_KV_TARGET_LAYERS).80        res->t_layer_inp[il] = inpL;81 82        const float freq_base_l  = model.get_rope_freq_base (cparams, il);83        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);84 85        ggml_tensor * inpSA = inpL;86 87        // RoPE runs on the SWA layers, NoPE on full ones.88        const bool use_rope = hparams.is_swa(il);89 90        // pre-attention norm (weight+1 folded at conversion time)91        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);92        cb(cur, "attn_norm", il);93 94        // self-attention: attention output gate around SDPA (afmoe.cpp:147-191)95        {96            ggml_tensor * attn_inp = cur;  // save input for gate computation97 98            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,99                    n_embd_head, n_head, n_head_kv, il);100 101            // gate = wqkv_gate @ attn_inp (from pre-attn hidden state)102            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);103            cb(gate, "attn_gate_proj", il);104 105            // QK-norm. attn_q_norm weight was synthesized at conversion to broadcast106            // qk_scale_factor across head_dim; attn_k_norm is identity (ones).107            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);108            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);109            cb(Qcur, "Qcur_normed", il);110            cb(Kcur, "Kcur_normed", il);111 112            if (use_rope) {113                Qcur = ggml_rope_ext(114                        ctx0, Qcur, inp_pos, nullptr,115                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,116                        ext_factor, attn_factor, beta_fast, beta_slow);117                cb(Qcur, "Qcur_rope", il);118 119                Kcur = ggml_rope_ext(120                        ctx0, Kcur, inp_pos, nullptr,121                        n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,122                        ext_factor, attn_factor, beta_fast, beta_slow);123                cb(Kcur, "Kcur_rope", il);124            }125 126            // SDPA. wo is deferred; the gate goes between attn_out and o_proj.127            cur = build_attn(inp_attn,128                    NULL, NULL, NULL,129                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);130            cb(cur, "attn_out", il);131 132            gate = ggml_sigmoid(ctx0, gate);133            cb(gate, "attn_gate_sig", il);134            cur = ggml_mul(ctx0, cur, gate);135            cb(cur, "attn_gated", il);136 137            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);138            cb(cur, "attn_o_proj", il);139        }140 141        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);142        cur = ggml_mul(ctx0, cur, model.layers[il].attn_post_norm);143        cb(cur, "attn_post_norm", il);144 145        if (il == n_layer - 1 && inp_out_ids) {146            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);147            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);148        }149 150        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);151        cb(ffn_inp, "ffn_inp", il);152 153        // pre-FFN norm154        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);155        cb(cur, "ffn_norm", il);156 157        // SwiGLU dense FFN158        cur = build_ffn(cur,159                model.layers[il].ffn_up,   NULL, NULL,160                model.layers[il].ffn_gate, NULL, NULL,161                model.layers[il].ffn_down, NULL, NULL,162                NULL,163                LLM_FFN_SILU, LLM_FFN_PAR, il);164        cb(cur, "ffn_out", il);165 166        cur = ggml_rms_norm(ctx0, cur, post_norm_eps);167        cur = ggml_mul(ctx0, cur, model.layers[il].ffn_post_norm);168        cb(cur, "ffn_post_norm", il);169 170        cur = ggml_add(ctx0, cur, ffn_inp);171        cur = build_cvec(cur, il);172        cb(cur, "l_out", il);173 174        inpL = cur;175    }176 177    cur = inpL;178 179    // final norm180    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);181    cb(cur, "result_norm", -1);182    res->t_embd = cur;183 184    // lm_head, followed by output multiplier185    cur = build_lora_mm(model.output, cur, model.output_s);186    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);187 188    // Final logit tanh softcap (from gemma3.cpp).189    if (hparams.f_final_logit_softcapping) {190        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);191        cur = ggml_tanh(ctx0, cur);192        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);193    }194 195    cb(cur, "result_output", -1);196    res->t_logits = cur;197 198    ggml_build_forward_expand(gf, cur);199}200 201std::unique_ptr<llm_graph_context> llama_model_muse_glimmer::build_arch_graph(const llm_graph_params & params) const {202    return std::make_unique<graph>(*this, params);203}204