Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_cohere2::load_arch_hparams(llama_model_loader & ml) {4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;5    uint32_t swa_period = 4;6    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);7    hparams.set_swa_pattern(swa_period);8 9    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;10    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;11 12    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa, false);13    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);14    ml.get_key(LLM_KV_LOGIT_SCALE,              hparams.f_logit_scale);15    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);16 17    switch (hparams.n_layer()) {18        case 32: type = LLM_TYPE_8B; break;19        default: type = LLM_TYPE_UNKNOWN;20    }21}22 23void llama_model_cohere2::load_arch_tensors(llama_model_loader &) {24    LLAMA_LOAD_LOCALS;25 26    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);27 28    // output29    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);30    // init output from the input tok embed31    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab },32                                      TENSOR_DUPLICATED);33 34    for (int i = 0; i < n_layer; ++i) {35        auto & layer = layers[i];36 37        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);38 39        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);40        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);41 42        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);43        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);44        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);45    }46}47 48std::unique_ptr<llm_graph_context> llama_model_cohere2::build_arch_graph(const llm_graph_params & params) const {49    return std::make_unique<graph>(*this, params);50}51 52llama_model_cohere2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {53    const int64_t n_embd_head = hparams.n_embd_head_v();54 55    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());56 57    const float f_logit_scale = hparams.f_logit_scale;58 59    ggml_tensor * cur;60    ggml_tensor * inpL;61 62    inpL = build_inp_embd(model.tok_embd);63 64    // inp_pos - contains the positions65    ggml_tensor * inp_pos = build_inp_pos();66 67    auto * inp_attn = build_attn_inp_kv_iswa();68 69    ggml_tensor * inp_out_ids = build_inp_out_ids();70 71    for (int il = 0; il < n_layer; ++il) {72        const bool is_swa = hparams.is_swa(il);73        // UNUSED:74        // const float freq_base_l  = model.get_rope_freq_base (cparams, il);75        // const float freq_scale_l = model.get_rope_freq_scale(cparams, il);76 77        // norm78        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);79        cb(cur, "attn_norm", il);80        ggml_tensor * ffn_inp = cur;81 82        // self-attention83        {84            // rope freq factors for 128k context85            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);86 87            // compute Q and K and RoPE them88            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,89                    n_embd_head, n_head, n_head_kv, il);90 91            if (is_swa) {92                Qcur = ggml_rope_ext(93                        ctx0, Qcur, inp_pos, rope_factors,94                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,95                        ext_factor, attn_factor, beta_fast, beta_slow96                        );97 98                Kcur = ggml_rope_ext(99                        ctx0, Kcur, inp_pos, rope_factors,100                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,101                        ext_factor, attn_factor, beta_fast, beta_slow102                        );103            }104 105            cb(Qcur, "Qcur", il);106            cb(Kcur, "Kcur", il);107            cb(Vcur, "Vcur", il);108 109            cur = build_attn(inp_attn,110                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,111                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);112        }113 114        if (il == n_layer - 1 && inp_out_ids) {115            cur     = ggml_get_rows(ctx0, cur, inp_out_ids);116            inpL    = ggml_get_rows(ctx0, inpL, inp_out_ids);117            ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);118        }119 120        ggml_tensor * attn_out = cur;121 122        // feed-forward network123        {124            cur = build_ffn(ffn_inp,125                    model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,126                    model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,127                    model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,128                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);129            cb(cur, "ffn_out", il);130        }131 132        // add together residual + FFN + self-attention133        cur = ggml_add(ctx0, cur, inpL);134        cur = ggml_add(ctx0, cur, attn_out);135 136        cur = build_cvec(cur, il);137        cb(cur, "l_out", il);138 139        // input for next layer140        inpL = cur;141    }142 143    cur = inpL;144 145    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);146 147    cb(cur, "result_norm", -1);148    res->t_embd = cur;149 150    // lm_head151    cur = build_lora_mm(model.output, cur, model.output_s);152 153    if (f_logit_scale) {154        cur = ggml_scale(ctx0, cur, f_logit_scale);155    }156 157    cb(cur, "result_output", -1);158    res->t_logits = cur;159 160    ggml_build_forward_expand(gf, cur);161}162