Felipe97/llama-cpp-compiled

1#include "models.h"2 3void llama_model_rnd1::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);6 7    switch (hparams.n_layer()) {8        case 48: type = LLM_TYPE_30B_A3B; break;9        default: type = LLM_TYPE_UNKNOWN;10    }11 12    // Set non-causal attention for diffusion models13    hparams.causal_attn = false;14}15 16void llama_model_rnd1::load_arch_tensors(llama_model_loader &) {17    LLAMA_LOAD_LOCALS;18 19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);20 21    // output22    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);23    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);24    // if output is NULL, init from the input tok embed25    if (output == NULL) {26        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);27    }28 29    for (int i = 0; i < n_layer; ++i) {30        auto & layer = layers[i];31 32        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);33 34        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);35        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);36 37        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);38        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);39 40        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);41 42        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);43 44        if (n_expert == 0) {45            throw std::runtime_error("n_expert must be > 0 for QWEN3MOE");46        }47        if (n_expert_used == 0) {48            throw std::runtime_error("n_expert_used must be > 0 for QWEN3MOE");49        }50 51        // MoE branch52        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;53 54        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);55        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);56        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);57    }58}59 60std::unique_ptr<llm_graph_context> llama_model_rnd1::build_arch_graph(const llm_graph_params & params) const {61    return std::make_unique<graph>(*this, params);62}63 64// RND1 is a Qwen3Moe AR model converted to diffusion model.65llama_model_rnd1::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {66    const int64_t n_embd_head = hparams.n_embd_head_v();67 68    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());69    GGML_ASSERT(n_embd_head == n_rot);70 71    ggml_tensor * cur;72    ggml_tensor * inpL;73 74    inpL = build_inp_embd(model.tok_embd);75 76    // inp_pos - contains the positions77    ggml_tensor * inp_pos = build_inp_pos();78 79    // Non-causal attention for diffusion80    auto * inp_attn = build_attn_inp_no_cache();81 82    ggml_tensor * inp_out_ids = build_inp_out_ids();83 84    for (int il = 0; il < n_layer; ++il) {85        ggml_tensor * inpSA = inpL;86 87        // norm88        cur = build_norm(inpL,89                model.layers[il].attn_norm, NULL,90                LLM_NORM_RMS, il);91        cb(cur, "attn_norm", il);92 93        // self_attention94        {95            // compute Q and K and RoPE them96            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,97                    n_embd_head, n_head, n_head_kv, il);98 99            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);100            cb(Qcur, "Qcur_normed", il);101 102            Qcur = ggml_rope_ext(103                    ctx0, Qcur, inp_pos, nullptr,104                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,105                    ext_factor, attn_factor, beta_fast, beta_slow106                    );107 108            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);109            cb(Kcur, "Kcur_normed", il);110 111            Kcur = ggml_rope_ext(112                    ctx0, Kcur, inp_pos, nullptr,113                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,114                    ext_factor, attn_factor, beta_fast, beta_slow115                    );116 117            cb(Qcur, "Qcur", il);118            cb(Kcur, "Kcur", il);119            cb(Vcur, "Vcur", il);120 121            cur = build_attn(inp_attn,122                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,123                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);124        }125        if (il == n_layer - 1 && inp_out_ids) {126            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);127            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);128        }129        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);130        cb(ffn_inp, "ffn_inp", il);131 132        // MoE branch133        cur = build_norm(ffn_inp,134                model.layers[il].ffn_norm, NULL,135                LLM_NORM_RMS, il);136        cb(cur, "ffn_norm", il);137 138        ggml_tensor * moe_out =139            build_moe_ffn(cur,140                    model.layers[il].ffn_gate_inp,141                    model.layers[il].ffn_up_exps,142                    model.layers[il].ffn_gate_exps,143                    model.layers[il].ffn_down_exps,144                    nullptr,145                    n_expert, n_expert_used,146                    LLM_FFN_SILU, true,147                    hparams.expert_weights_scale,148                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,149                    il);150        cb(moe_out, "ffn_moe_out", il);151        cur = moe_out;152 153        cur = ggml_add(ctx0, cur, ffn_inp);154 155        cur = build_cvec(cur, il);156        cb(cur, "l_out", il);157 158        // input for next layer159        inpL = cur;160    }161    cur = inpL;162 163    cur = build_norm(cur,164            model.output_norm, NULL,165            LLM_NORM_RMS, -1);166 167    cb(cur, "result_norm", -1);168    res->t_embd = cur;169 170    // lm_head171    cur = build_lora_mm(model.output, cur, model.output_s);172 173    cb(cur, "result_output", -1);174    res->t_logits = cur;175 176    ggml_build_forward_expand(gf, cur);177}178