Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_olmo2::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);7    if (found_swa && hparams.n_swa > 0) {8        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;9        uint32_t swa_period = 4;10        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);11        hparams.set_swa_pattern(swa_period);12 13        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;14        hparams.rope_freq_scale_train_swa = 1.0; // See olmo2.cpp15        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);16    } else {17        hparams.swa_type = LLAMA_SWA_TYPE_NONE;18    }19 20    switch (hparams.n_layer()) {21        case 16: type = LLM_TYPE_1B; break;22        case 32: type = LLM_TYPE_7B; break;23        case 40: type = LLM_TYPE_13B; break;24        case 64: type = LLM_TYPE_32B; break;25        default: type = LLM_TYPE_UNKNOWN;26    }27}28 29void llama_model_olmo2::load_arch_tensors(llama_model_loader &) {30    LLAMA_LOAD_LOCALS;31 32    const int64_t n_embd_head = n_embd / n_head;33 34    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);35 36    // output37    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);38    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);39 40    for (int i = 0; i < n_layer; ++i) {41        auto & layer = layers[i];42 43        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);44        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);45        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);46        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_head_kv * n_embd_head}, 0);47        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);48 49        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);50        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);51        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);52        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);53    }54}55 56std::unique_ptr<llm_graph_context> llama_model_olmo2::build_arch_graph(const llm_graph_params & params) const {57    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {58        return std::make_unique<graph<true>>(*this, params);59    } else {60        return std::make_unique<graph<false>>(*this, params);61    }62}63 64template <bool iswa>65llama_model_olmo2::graph<iswa>::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    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;80    inp_attn_type * inp_attn = nullptr;81 82    if constexpr (iswa) {83        inp_attn = build_attn_inp_kv_iswa();84    } else {85        inp_attn = build_attn_inp_kv();86    }87    ggml_tensor * inp_out_ids = build_inp_out_ids();88 89    for (int il = 0; il < n_layer; ++il) {90        ggml_tensor * inpSA = inpL;91 92        cur = inpL;93 94        // self_attention95        {96            // compute Q and K and RoPE them97            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);98            cb(Qcur, "Qcur", il);99 100            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);101            cb(Kcur, "Kcur", il);102 103            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);104            cb(Vcur, "Vcur", il);105 106            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,107                    LLM_NORM_RMS, il);108            cb(Qcur, "Qcur_normed", il);109 110            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,111                    LLM_NORM_RMS, il);112            cb(Kcur, "Kcur_normed", il);113 114            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);115            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);116            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);117 118            const bool is_swa = hparams.is_swa(il);119 120            if (is_swa) {121                // For sliding window layers, Olmo3 use regular rope with no yarn rope scaling.122                // This is achieved here by setting freq_scale and attn_factor to 1.123                // We also set ext_factor to 0 to avoid a few unnecessary computations.124                Qcur = ggml_rope_ext(125                    ctx0, Qcur, inp_pos, nullptr,126                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,127                    0.0, 1.0, beta_fast, beta_slow128                    );129 130                Kcur = ggml_rope_ext(131                    ctx0, Kcur, inp_pos, nullptr,132                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,133                    0.0, 1.0, beta_fast, beta_slow134                    );135            } else {136                Qcur = ggml_rope_ext(137                    ctx0, Qcur, inp_pos, nullptr,138                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,139                    ext_factor, attn_factor, beta_fast, beta_slow140                    );141 142                Kcur = ggml_rope_ext(143                    ctx0, Kcur, inp_pos, nullptr,144                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,145                    ext_factor, attn_factor, beta_fast, beta_slow146                    );147            }148            cb(Qcur, "Qcur", il);149            cb(Kcur, "Kcur", il);150            cb(Vcur, "Vcur", il);151 152            cur = build_attn(inp_attn,153                    model.layers[il].wo, NULL, model.layers[il].wo_s,154                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);155        }156        if (il == n_layer - 1 && inp_out_ids) {157            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);158            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);159        }160        cur = build_norm(cur,161                model.layers[il].attn_post_norm, NULL,162                LLM_NORM_RMS, il);163        cb(cur, "attn_post_norm", il);164 165        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);166        cb(ffn_inp, "ffn_inp", il);167 168        // feed-forward network169        cur = build_ffn(ffn_inp,170                model.layers[il].ffn_up,   NULL, NULL,171                model.layers[il].ffn_gate, NULL, NULL,172                model.layers[il].ffn_down, NULL, NULL,173                NULL,174                LLM_FFN_SILU, LLM_FFN_PAR, il);175        cb(cur, "ffn_out", il);176 177        cur = build_norm(cur,178                model.layers[il].ffn_post_norm, NULL,179                LLM_NORM_RMS, -1);180        cb(cur, "ffn_post_norm", -1);181 182        cur = ggml_add(ctx0, cur, ffn_inp);183        cb(cur, "ffn_out", il);184 185        cur = build_cvec(cur, il);186        cb(cur, "l_out", il);187 188        // input for next layer189        inpL = cur;190    }191    cur = inpL;192 193    cur = build_norm(cur,194            model.output_norm, NULL,195            LLM_NORM_RMS, -1);196 197    cb(cur, "result_norm", -1);198    res->t_embd = cur;199 200    // lm_head201    cur = build_lora_mm(model.output, cur, model.output_s);202 203    cb(cur, "result_output", -1);204    res->t_logits = cur;205 206    ggml_build_forward_expand(gf, cur);207}208 209// Explicit template instantiations210template struct llama_model_olmo2::graph<false>;211template struct llama_model_olmo2::graph<true>;212