Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6    if (!ml.get_arr(LLM_KV_TARGET_LAYERS, target_layer_ids, false)) {7        throw std::runtime_error("EAGLE3 model requires 'extract_layers' in GGUF metadata");8    }9    if (target_layer_ids.size() != 3) {10        throw std::runtime_error("EAGLE3 requires exactly 3 entries in 'extract_layers'");11    }12    LLAMA_LOG_INFO("%s: EAGLE3 extract_layers = [%d, %d, %d]\n", __func__,13            target_layer_ids[0],14            target_layer_ids[1],15            target_layer_ids[2]);16 17    uint32_t n_embd_tgt = 0;18 19    ml.get_key(LLM_KV_TARGET_HIDDEN_SIZE, n_embd_tgt);20    LLAMA_LOG_INFO("%s: EAGLE3 n_embd_tgt = %u (draft n_embd = %u)\n", __func__, n_embd_tgt, hparams.n_embd);21 22    hparams.n_embd_inp_enc_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;23 24    // eagle3 norm_before_residual (optional, default false)25    // compatible with Readhat eagle3 speculator model26    ml.get_key(LLM_KV_NORM_BEFORE_RESIDUAL, hparams.norm_before_residual, false);27    if (hparams.norm_before_residual) {28        LLAMA_LOG_INFO("%s: EAGLE3gnorm_before_residual = true\n", __func__);29    }30 31    // eagle3 norm_before_fc (optional, default false)32    // compatible with eagle3.1 (e.g. nvidia/gpt-oss-120b-Eagle3-v3)33    ml.get_key(LLM_KV_NORM_BEFORE_FC, hparams.norm_before_fc, false);34 35    type = LLM_TYPE_UNKNOWN;36}37 38void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {39    LLAMA_LOAD_LOCALS;40 41    const int64_t n_embd_inp = hparams.n_embd_inp_enc();42    const int64_t n_embd_attn_input = 2 * n_embd;43 44    // Get vocab size from the d2t tensor in the GGUF file (optional - only needed if eagle3 has different vocab_size than target)45    // d2t: draft to target vocabulary mapping46    int64_t n_draft_vocab = n_vocab;  // Default: same as target vocab47    const struct ggml_tensor * d2t_meta = ml->get_tensor_meta("d2t");48    if (d2t_meta) {49        n_draft_vocab = d2t_meta->ne[0]; // update draft vocab size50        d2t = create_tensor(tn(LLM_TENSOR_D2T), {n_draft_vocab}, 0);51        LLAMA_LOG_INFO("%s: EAGLE3 using d2t mapping (draft_vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);52    } else {53        d2t = nullptr; // no d2t, use default vocab size54        LLAMA_LOG_INFO("%s: EAGLE3 without d2t - sharing same vocab_size with target (vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);55    }56 57    // Feature fusion layer: projects 3 target layers to draft hidden size58    fc = create_tensor(tn(LLM_TENSOR_FC, "weight"), {n_embd_inp, n_embd}, 0);59 60    // RMSNorm on the fused target features (input to fc), only when norm_before_fc is set.61    if (hparams.norm_before_fc) {62        output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd_inp}, 0);63    }64 65    // Output layer (uses draft vocab size)66    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);67    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_draft_vocab}, TENSOR_NOT_REQUIRED);68 69    // Token embeddings (optional - Llama 3.3 70B EAGLE3 has its own)70    const struct ggml_tensor * tok_embd_meta = ml->get_tensor_meta(tn(LLM_TENSOR_TOKEN_EMBD, "weight").str().c_str());71    if (tok_embd_meta) {72        const int64_t n_target_vocab = tok_embd_meta->ne[1];73        tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_target_vocab}, 0);74        LLAMA_LOG_INFO("%s: EAGLE3 using its own token_embd (vocab = %lld)\n", __func__, (long long)n_target_vocab);75    }76 77    // Single decoder layer78    for (int i = 0; i < n_layer; ++i) {79        auto & layer = layers[i];80 81        // input_layernorm: applied to token embeddings82        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);83 84        // eagle3 specific: hidden_norm applied to fused target features85        layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, 0);86 87        // Attention takes input_embeds_normed + fused_target_normed as input88        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd_attn_input, n_embd_head_k * n_head}, 0);89        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd_attn_input, n_embd_k_gqa}, 0);90        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd_attn_input, n_embd_v_gqa}, 0);91        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);92 93        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);94        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);95        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);96        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);97 98        // rope_freqs for llama3 rope scaling (optional - only if eagle3 config has rope_scaling)99        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED);100    }101}102 103std::unique_ptr<llm_graph_context> llama_model_eagle3::build_arch_graph(const llm_graph_params & params) const {104    switch (params.gtype) {105        case LLM_GRAPH_TYPE_ENCODER:106            return std::make_unique<graph<true>>(*this, params);107        case LLM_GRAPH_TYPE_DEFAULT:108        case LLM_GRAPH_TYPE_DECODER:109            return std::make_unique<graph<false>>(*this, params);110        default:111            GGML_ABORT("invalid graph type");112    };113}114 115template <>116ggml_tensor * llama_model_eagle3::graph<true>::build_inp_embd_enc() const {117    ggml_tensor * cur = nullptr;118 119    // Input: Target model features (3 layers concatenated: low, mid, high)120    // Data will be provided via ubatch->embd in encode_eagle3_features()121    auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp_enc());122    inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp_enc(), n_tokens);123    ggml_set_input(inp_target->embd);124 125    cur = inp_target->embd;126    cb(cur, "inp_embd", -1);127 128    res->add_input(std::move(inp_target));129 130    return cur;131}132 133// eagle3 Encoder: processes target model features through feature fusion layer134// Input: target_features e.g. [12288, n_tokens] from target model layers low, middle, high135// Output: g_embeddings e.g. [4096, n_tokens] stored in context136template <>137llama_model_eagle3::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {138    ggml_tensor * cur = nullptr;139 140    cur = build_inp_embd_enc();141 142    // RMSNorm on the fused target features before fc143    if (hparams.norm_before_fc) {144        cur = build_norm(cur, model.output_norm_enc, NULL, LLM_NORM_RMS, -1);145        cb(cur, "enc_input_norm", -1);146    }147 148    // Feature fusion layer149    cur = build_lora_mm(model.fc, cur);150    cb(cur, "fc_out", -1);151 152    // Output: g_embeddings e.g. [4096, n_tokens]153    // store in t_h_nextn (same as MTP) so can be read via llama_get_embeddings_nextn(ctx_dft)154    ggml_set_output(cur);155    res->t_h_nextn = cur;156 157    ggml_build_forward_expand(gf, cur);158}159 160// eagle3 Decoder: processes draft tokens using g_embeddings from encoder161// Input: draft tokens + g_embeddings from encoder162// Output: draft logits163template <>164llama_model_eagle3::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {165    const int64_t n_embd_head = hparams.n_embd_head_v();166 167    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());168    GGML_ASSERT(n_layer == 1);  // eagle3 has only one decoder layer169 170    ggml_tensor * cur;171    ggml_tensor * inpL;172 173    // eagle3 Decoder receives:174    // 1. Token embeddings (e.g.from eagle3's own tok_embd for Llama 3.3 70B, or target model for Llama 3.1 8B)175    // 2. g_embeddings from encoder176    auto * tok_embd = model.tok_embd;177    if (model.tok_embd == nullptr) {178        GGML_ASSERT(cparams.ctx_other != nullptr);179        const auto * model_other = llama_get_model(cparams.ctx_other);180 181        GGML_ASSERT(model_other->tok_embd != nullptr && "EAGLE3 decoder requires token embeddings (own or from target model)");182        tok_embd = model_other->tok_embd;183    }184 185    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);186 187    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);188    ggml_set_input(inp->tokens);189 190    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);191    ggml_set_input(inp->embd);192 193    ggml_tensor * inp_embd = ggml_get_rows(ctx0, tok_embd, inp->tokens);194    cb(inp_embd, "inp_embd", -1);195 196    ggml_tensor * inp_g = inp->embd;197    cb(inp_g, "inp_g_embeddings", -1);198 199    res->add_input(std::move(inp));200 201    inpL = inp_g;202 203    // inp_pos - contains the positions204    ggml_tensor * inp_pos = build_inp_pos();205 206    auto * inp_attn = build_attn_inp_kv();207 208    const float kq_scale = 1.0f/sqrtf(float(n_embd_head));209 210    // Single decoder layer (il = 0)211    const int il = 0;212    {213        // Apply input_layernorm to the token embeddings214        ggml_tensor * embd_norm = build_norm(inp_embd,215                model.layers[il].attn_norm, NULL,216                LLM_NORM_RMS, il);217        cb(embd_norm, "embd_norm", il);218 219        // Apply hidden_norm to inp_g220        ggml_tensor * g_norm = build_norm(inp_g,221                model.layers[il].attn_norm_2, NULL,222                LLM_NORM_RMS, -1);223        cb(g_norm, "g_norm", il);224 225        // norm_before_residual: determines what goes into the residual connection (compatible with Readhat eagle3 speculator model)226        // - false (default): use raw inp_g for residual227        // - true: use normalized g_norm for residual228        // inpL is the concatenated input (normalized inp_embd + normalized inp_g)229        ggml_tensor * inpSA = hparams.norm_before_residual ? g_norm : inpL;230 231        // Concatenate normalized inp_embd and normalized inp_g232        cur = ggml_concat(ctx0, embd_norm, g_norm, il);233        cb(cur, "concat_embd", il);234 235        // Self-attention with concatenated input236        ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);237        cb(Qcur, "Qcur", il);238 239        ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);240        cb(Kcur, "Kcur", il);241 242        ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);243        cb(Vcur, "Vcur", il);244 245        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);246        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);247        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);248 249        // rope freq factors, returns nullptr if not available250        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);251 252        // RoPE253        Qcur = ggml_rope_ext(254                ctx0, Qcur, inp_pos, rope_factors,255                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,256                ext_factor, attn_factor, beta_fast, beta_slow257                );258        Kcur = ggml_rope_ext(259                ctx0, Kcur, inp_pos, rope_factors,260                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,261                ext_factor, attn_factor, beta_fast, beta_slow262                );263 264        cb(Qcur, "Qcur_rope", il);265        cb(Kcur, "Kcur_rope", il);266 267        cur = build_attn(inp_attn,268                model.layers[il].wo, NULL, nullptr,269                Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);270 271        // Add residual and update it272        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);273        cb(ffn_inp, "ffn_inp", il);274 275        // Apply FFN norm to the sum276        cur = build_norm(ffn_inp,277                model.layers[il].ffn_norm, NULL,278                LLM_NORM_RMS, il);279        cb(cur, "post_attn_norm", il);280 281        cur = build_ffn(cur,282                model.layers[il].ffn_up,   NULL, NULL,283                model.layers[il].ffn_gate, NULL, NULL,284                model.layers[il].ffn_down, NULL, NULL,285                NULL,286                LLM_FFN_SILU, LLM_FFN_PAR, il);287        cb(cur, "ffn_out", il);288 289        // Output norm with residual290        cur = ggml_add(ctx0, cur, ffn_inp);291        cb(cur, "eagle3_prenorm", il);292 293        inpL = cur;294    }295 296    cur = inpL;297 298    // Output prenorm state (for next token's g_embeddings in autoregressive generation)299    ggml_set_output(cur);300    res->t_h_nextn = cur;301 302    cur = build_norm(cur,303            model.output_norm, NULL,304            LLM_NORM_RMS, -1);305    cb(cur, "result_norm", -1);306 307    // lm_head - projects to draft vocabulary308    // if the draft has no own output projection, inherit the target model's lm_head309    auto * output = model.output;310    if (output == nullptr) {311        GGML_ASSERT(cparams.ctx_other != nullptr);312        const auto * model_other = llama_get_model(cparams.ctx_other);313 314        GGML_ASSERT(model_other->output != nullptr && "EAGLE3 decoder requires an output projection (own or from target model)");315        output = model_other->output;316    }317    cur = build_lora_mm(output, cur);318 319    if (model.d2t) {320        const int64_t n_draft_vocab = cur->ne[0];321        const int64_t n_outputs     = cur->ne[1];322        const int64_t n_vocab       = (int64_t) model.vocab.n_tokens();323 324        GGML_ASSERT(model.d2t->type == GGML_TYPE_I64);325        GGML_ASSERT(model.d2t->ne[0] == n_draft_vocab);326 327        ggml_tensor * logits = ggml_fill(ctx0, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, n_vocab, n_outputs), -INFINITY);328        cur = ggml_set_rows(ctx0, logits,329                ggml_reshape_3d(ctx0, cur,       1,             n_draft_vocab, n_outputs),330                ggml_reshape_3d(ctx0, model.d2t, n_draft_vocab, 1,             1));331        cur = ggml_reshape_2d(ctx0, cur, n_vocab, n_outputs);332    }333 334    cb(cur, "result_output", -1);335    res->t_logits = cur;336 337    ggml_build_forward_expand(gf, cur);338}339