Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_granite_hybrid::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_LOGIT_SCALE,                 hparams.f_logit_scale, /* required */ false);6    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, /* required */ false);7    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, /* required */ false);8    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, /* required */ false);9 10    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);11    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);12    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);13    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);14    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);15 16    // Granite uses rope_finetuned as a switch for rope, so default to true17    bool rope_finetuned = true;18    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);19    hparams.rope_finetuned = rope_finetuned;20 21    // A layer is recurrent IFF the n_head_kv value is set to 022    for (uint32_t i = 0; i < hparams.n_layer(); ++i) {23        hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;24    }25 26    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);27 28    switch (hparams.n_embd) {29        case 768: type = LLM_TYPE_350M; break;30        case 1536: type = (hparams.n_ff() == 512 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break;31        case 2048: case 2560: type = LLM_TYPE_3B; break;32        case 4096: type = LLM_TYPE_32B; break;33        default: type = LLM_TYPE_UNKNOWN;34    }35 36    // For Granite MoE Shared37    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);38}39 40void llama_model_granite_hybrid::load_arch_tensors(llama_model_loader &) {41    LLAMA_LOAD_LOCALS;42 43    // mamba2 Mixer SSM params44    // NOTE: int64_t for tensor dimensions45    const int64_t d_conv     = hparams.ssm_d_conv;46    const int64_t d_inner    = hparams.ssm_d_inner;47    const int64_t d_state    = hparams.ssm_d_state;48    const int64_t n_ssm_head = hparams.ssm_dt_rank;49    const int64_t n_group    = hparams.ssm_n_group;50    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;51 52    // only an expansion factor of 2 is supported for now53    GGML_ASSERT(2 * n_embd == d_inner);54 55    // embeddings56    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);57 58    // output59    {60        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);61        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);62        // if output is NULL, init from the input tok embed, duplicated to allow offloading63        if (output == NULL) {64            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);65        }66    }67 68    for (int i = 0; i < n_layer; ++i) {69        auto & layer = layers[i];70 71        // norm72        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);73 74        if (hparams.is_recr(i)) {75            // ssm layers76            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);77 78            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);79            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);80 81            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);82 83            // no "weight" suffix for these84            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);85            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);86 87            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);88 89            // out_proj90            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);91        } else {92            // attention layers (with optional bias)93            const int64_t n_head_i = hparams.n_head(i);94            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);95            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);96            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);97            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);98            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);99        }100 101        // feed forward (w/ optional biases)102        if (n_expert > 0) {103            // MoE FFN104            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);105            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));106            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);107            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);108            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);109            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);110 111            // For Granite MoE Shared112            if (hparams.n_ff_shexp > 0) {113                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);114                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);115                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);116            }117        } else {118            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);119            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));120            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);121            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);122            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);123            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);124            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);125            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);126        }127    }128}129 130std::unique_ptr<llm_graph_context> llama_model_granite_hybrid::build_arch_graph(const llm_graph_params & params) const {131    return std::make_unique<graph>(*this, params);132}133 134llama_model_granite_hybrid::graph::graph(const llama_model & model, const llm_graph_params & params) :135    llm_build_mamba_base(params) {136    const int64_t n_embd_head = hparams.n_embd_head_v();137    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());138 139    ggml_tensor * cur;140    ggml_tensor * inpL;141 142    inpL = build_inp_embd(model.tok_embd);143 144    auto * inp = build_inp_mem_hybrid();145 146    ggml_tensor * inp_out_ids = build_inp_out_ids();147 148    // Positional embeddings populated if rope enabled149    ggml_tensor * inp_pos = nullptr;150    if (hparams.rope_finetuned) {151        inp_pos = build_inp_pos();152    }153 154    for (int il = 0; il < n_layer; ++il) {155        struct ggml_tensor * inpSA = inpL;156 157        // norm158        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);159        cb(cur, "attn_norm", il);160 161        if (hparams.is_recr(il)) {162            // ssm layer //163            cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il);164        } else {165            // attention layer //166            cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il);167        }168 169        if (il == n_layer - 1 && inp_out_ids) {170            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);171            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);172        }173 174        // ffn175        cur = build_layer_ffn(cur, inpSA, model, il);176 177        // input for next layer178        inpL = cur;179    }180 181    cur = inpL;182 183    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);184 185    cb(cur, "result_norm", -1);186    res->t_embd = cur;187 188    // lm_head189    cur = build_lora_mm(model.output, cur, model.output_s);190 191    // For Granite architectures - scale logits192    if (hparams.f_logit_scale) {193        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);194    }195    cb(cur, "result_output", -1);196    res->t_logits = cur;197 198    ggml_build_forward_expand(gf, cur);199}200 201ggml_tensor * llama_model_granite_hybrid::graph::build_attention_layer(ggml_tensor *             cur,202                                                              ggml_tensor *             inp_pos,203                                                              llm_graph_input_attn_kv * inp_attn,204                                                              const llama_model &       model,205                                                              const int64_t             n_embd_head,206                                                              const int                 il) {207    auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, hparams.n_head(il), hparams.n_head_kv(il), il);208 209    const bool use_rope = hparams.rope_finetuned;210    if (use_rope) {211        ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);212        Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,213                             ext_factor, attn_factor, beta_fast, beta_slow);214 215        Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,216                             ext_factor, attn_factor, beta_fast, beta_slow);217    }218 219    cb(Qcur, "Qcur", il);220    cb(Kcur, "Kcur", il);221    cb(Vcur, "Vcur", il);222 223    const float kq_scale =224        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;225    cur = build_attn(inp_attn,226            model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,227            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);228    cb(cur, "attn_out", il);229    return cur;230}231 232ggml_tensor * llama_model_granite_hybrid::graph::build_layer_ffn(ggml_tensor *       cur,233                                                        ggml_tensor *       inpSA,234                                                        const llama_model & model,235                                                        const int           il) {236    // For Granite architectures - scale residual237    if (hparams.f_residual_scale) {238        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);239    }240    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);241    cb(ffn_inp, "ffn_inp", il);242 243    // feed-forward network (non-MoE)244    if (model.layers[il].ffn_gate_inp == nullptr) {245        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);246        cb(cur, "ffn_norm", il);247 248        cur = build_ffn(cur,249                model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,250                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,251                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,252                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);253        cb(cur, "ffn_out", il);254 255    } else {256        // MoE branch257        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);258        cb(cur, "ffn_norm", il);259 260        ggml_tensor * moe_out =261            build_moe_ffn(cur,262                model.layers[il].ffn_gate_inp,263                model.layers[il].ffn_up_exps,264                model.layers[il].ffn_gate_exps,265                model.layers[il].ffn_down_exps,266                nullptr,267                n_expert, n_expert_used,268                LLM_FFN_SILU, true,269                hparams.expert_weights_scale,270                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,271                il);272        cb(moe_out, "ffn_moe_out", il);273 274        // For Granite MoE Shared275        if (hparams.n_ff_shexp > 0) {276            ggml_tensor * ffn_shexp =277                build_ffn(cur,278                    model.layers[il].ffn_up_shexp, NULL, NULL,279                    model.layers[il].ffn_gate_shexp, NULL, NULL,280                    model.layers[il].ffn_down_shexp, NULL, NULL,281                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);282            cb(ffn_shexp, "ffn_shexp", il);283 284            cur = ggml_add(ctx0, moe_out, ffn_shexp);285            cb(cur, "ffn_out", il);286        } else {287            cur = moe_out;288        }289    }290 291    // For Granite architectures - scale residual292    if (hparams.f_residual_scale) {293        cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);294    }295    cur = ggml_add(ctx0, cur, ffn_inp);296    cb(cur, "ffn_out", il);297 298    cur = build_cvec(cur, il);299    cb(cur, "l_out", il);300 301    return cur;302}303