Coverage Report

Created: 2026-06-22 06:47

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/src/llama.cpp/src/models/granite-moe.cpp
Line
Count
Source
1
#include "models.h"
2
3
0
void llama_model_granite_moe::load_arch_hparams(llama_model_loader & ml) {
4
0
    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
0
    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
6
0
    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
7
0
    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
8
0
    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);
9
10
    // Granite uses rope_finetuned as a switch for rope, so default to true
11
0
    bool rope_finetuned = true;
12
0
    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
13
0
    hparams.rope_finetuned = rope_finetuned;
14
15
0
    switch (hparams.n_layer()) {
16
0
        case 32: type = LLM_TYPE_3B; break;
17
0
        case 40: type = LLM_TYPE_3B; break;
18
        // Add additional layer/vocab/etc checks here for other model sizes
19
0
        default: type = LLM_TYPE_UNKNOWN;
20
0
    }
21
22
    // For Granite MoE Shared
23
0
    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
24
0
}
25
26
0
void llama_model_granite_moe::load_arch_tensors(llama_model_loader &) {
27
0
    LLAMA_LOAD_LOCALS;
28
29
0
    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
30
31
    // output
32
0
    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
33
0
    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
34
35
    // if output is NULL, init from the input tok embed
36
0
    if (output == NULL) {
37
0
        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
38
0
    }
39
40
0
    for (int i = 0; i < n_layer; ++i) {
41
0
        auto & layer = layers[i];
42
43
0
        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
44
45
0
        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
46
0
        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
47
48
        // optional bias tensors
49
0
        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
50
51
0
        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
52
53
0
        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
54
0
            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
55
0
            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
56
0
        }
57
0
        else {
58
0
            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
59
0
        }
60
61
0
        if (n_expert == 0) {
62
0
            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
63
0
            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
64
0
            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
65
66
            // optional MLP bias
67
0
            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
68
0
            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
69
0
            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
70
0
        } else {
71
0
            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
72
0
            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
73
0
            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
74
0
            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
75
76
            // For Granite MoE Shared
77
0
            if (hparams.n_ff_shexp > 0) {
78
0
                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
79
0
                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
80
0
                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
81
0
            }
82
0
        }
83
0
    }
84
0
}
85
86
0
std::unique_ptr<llm_graph_context> llama_model_granite_moe::build_arch_graph(const llm_graph_params & params) const {
87
0
    return std::make_unique<graph>(*this, params);
88
0
}
89