Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_olmoe::load_arch_hparams(llama_model_loader & ml) {4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);5 6    switch (hparams.n_layer()) {7        case 16: type = LLM_TYPE_A1_7B; break;8        default: type = LLM_TYPE_UNKNOWN;9    }10}11 12void llama_model_olmoe::load_arch_tensors(llama_model_loader &) {13    LLAMA_LOAD_LOCALS;14 15    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);16 17    // output18    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);19    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);20 21    for (int i = 0; i < n_layer; ++i) {22        auto & layer = layers[i];23 24        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);25 26        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);27        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);28        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);29        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0);30 31        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);32 33        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);34 35        if (n_expert == 0) {36            throw std::runtime_error("n_expert must be > 0");37        }38        if (n_expert_used == 0) {39            throw std::runtime_error("n_expert_used must be > 0");40        }41 42        // MoE branch43        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);44        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);45        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);46    }47}48 49std::unique_ptr<llm_graph_context> llama_model_olmoe::build_arch_graph(const llm_graph_params & params) const {50    return std::make_unique<graph>(*this, params);51}52 53llama_model_olmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {54    const int64_t n_embd_head = hparams.n_embd_head_v();55 56    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());57    GGML_ASSERT(n_embd_head == n_rot);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();68 69    ggml_tensor * inp_out_ids = build_inp_out_ids();70 71    for (int il = 0; il < n_layer; ++il) {72        ggml_tensor * inpSA = inpL;73 74        // norm75        cur = build_norm(inpL,76                model.layers[il].attn_norm, NULL,77                LLM_NORM_RMS, il);78        cb(cur, "attn_norm", il);79 80        // self_attention81        {82            // compute Q and K and RoPE them83            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);84            cb(Qcur, "Qcur", il);85 86            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);87            cb(Kcur, "Kcur", il);88 89            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);90            cb(Vcur, "Vcur", il);91 92            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,93                    LLM_NORM_RMS, il);94            cb(Qcur, "Qcur_normed", il);95 96            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,97                    LLM_NORM_RMS, il);98            cb(Kcur, "Kcur_normed", il);99 100            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);101            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);102            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);103 104            Qcur = ggml_rope_ext(105                    ctx0, Qcur, inp_pos, nullptr,106                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,107                    ext_factor, attn_factor, beta_fast, beta_slow108                    );109 110            Kcur = ggml_rope_ext(111                    ctx0, Kcur, inp_pos, nullptr,112                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,113                    ext_factor, attn_factor, beta_fast, beta_slow114                    );115 116            cb(Qcur, "Qcur", il);117            cb(Kcur, "Kcur", il);118            cb(Vcur, "Vcur", il);119 120            cur = build_attn(inp_attn,121                    model.layers[il].wo, NULL, model.layers[il].wo_s,122                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);123        }124        if (il == n_layer - 1 && inp_out_ids) {125            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);126            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);127        }128        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);129        cb(ffn_inp, "ffn_inp", il);130 131        // MoE branch132        cur = build_norm(ffn_inp,133                model.layers[il].ffn_norm, NULL,134                LLM_NORM_RMS, il);135        cb(cur, "ffn_norm", il);136 137        cur = build_moe_ffn(cur,138                model.layers[il].ffn_gate_inp,139                model.layers[il].ffn_up_exps,140                model.layers[il].ffn_gate_exps,141                model.layers[il].ffn_down_exps,142                nullptr,143                n_expert, n_expert_used,144                LLM_FFN_SILU, false,145                hparams.expert_weights_scale,146                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,147                il);148        cb(cur, "ffn_moe_out", il);149 150        cur = ggml_add(ctx0, cur, ffn_inp);151 152        cur = build_cvec(cur, il);153        cb(cur, "l_out", il);154 155        // input for next layer156        inpL = cur;157    }158    cur = inpL;159 160    cur = build_norm(cur,161            model.output_norm, NULL,162            LLM_NORM_RMS, -1);163 164    cb(cur, "result_norm", -1);165    res->t_embd = cur;166 167    // lm_head168    cur = build_lora_mm(model.output, cur, model.output_s);169 170    cb(cur, "result_output", -1);171    res->t_logits = cur;172 173    ggml_build_forward_expand(gf, cur);174}175