Brunobkr/llama.cpp_AlgMor24_github

1#include "models.h"2 3void llama_model_deepseek2ocr::load_arch_hparams(llama_model_loader & ml) {4    // similar to deepseek2, but without MLA5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);6    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);7    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);8    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);9    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);10    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);11    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);12 13    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {14        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;15    }16 17    switch (hparams.n_layer()) {18        case 12: type = LLM_TYPE_3B; break;19        default: type = LLM_TYPE_UNKNOWN;20    }21}22 23void llama_model_deepseek2ocr::load_arch_tensors(llama_model_loader &) {24    LLAMA_LOAD_LOCALS;25    const int64_t n_expert_shared = hparams.n_expert_shared;26 27    // similar to deepseek2, but without MLA28    const int64_t n_ff_exp        = hparams.n_ff_exp;29 30    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);31 32    // output33    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);34    // try to load output.weight, if not found, use token_embd (tied embeddings)35    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);36    if (!output) {37        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);38    }39 40    for (int i = 0; i < n_layer; ++i) {41        auto & layer = layers[i];42 43        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);44        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd}, 0);45        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd}, 0);46        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);47 48        // norm49        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);50        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);51 52        if (i < (int) hparams.n_layer_dense_lead) {53            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);54            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);55            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);56        } else {57            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);58            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);59 60            if (n_expert == 0) {61                throw std::runtime_error("n_expert must be > 0");62            }63            if (n_expert_used == 0) {64                throw std::runtime_error("n_expert_used must be > 0");65            }66 67            // MoE branch68            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);69            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);70 71            // Shared expert branch72            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);73            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);74            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);75        }76    }77}78 79std::unique_ptr<llm_graph_context> llama_model_deepseek2ocr::build_arch_graph(const llm_graph_params & params) const {80    return std::make_unique<graph>(*this, params);81}82 83