char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

PikaPikachuconvert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)465e49b9c

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
3.9 KiB80 linesraw
1#include "models.h"
2
3void llama_model_deepseek2ocr::load_arch_hparams(llama_model_loader & ml) {
4    // similar to deepseek2, but without MLA
5    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_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
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 MLA
28    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    // output
33    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        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd, n_embd, 0);
44        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
45
46        // norm
47        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
48        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
49
50        if (i < (int) hparams.n_layer_dense_lead) {
51            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
52            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
53            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
54        } else {
55            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
56            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
57
58            if (n_expert == 0) {
59                throw std::runtime_error("n_expert must be > 0");
60            }
61            if (n_expert_used == 0) {
62                throw std::runtime_error("n_expert_used must be > 0");
63            }
64
65            // MoE branch
66            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
67            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
68
69            // Shared expert branch
70            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
71            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
72            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
73        }
74    }
75}
76
77std::unique_ptr<llm_graph_context> llama_model_deepseek2ocr::build_arch_graph(const llm_graph_params & params) const {
78    return std::make_unique<graph>(*this, params);
79}
80