char-fork/llama.cpp

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

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.8 KiB164 linesraw
1#include "models.h"
2
3void llama_model_ernie4_5::load_arch_hparams(llama_model_loader & ml) {
4    // paddleocr need mrope_section
5    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
6
7    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
8    if (arch == LLM_ARCH_ERNIE4_5_MOE) {
9        ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
10        ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
11        ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         hparams.n_moe_layer_step);
12        ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
13    }
14
15    switch (hparams.n_layer()) {
16        case 18: type = LLM_TYPE_0_3B; break;
17        case 28: type = LLM_TYPE_21B_A3B; break;
18        case 54: type = LLM_TYPE_300B_A47B; break;
19        default: type = LLM_TYPE_UNKNOWN;
20    }
21}
22
23void llama_model_ernie4_5::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    // output
29    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
30    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
31    // if output is NULL, init from the input tok embed
32    if (output == NULL) {
33        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
34    }
35
36    for (int i = 0; i < n_layer; ++i) {
37        auto & layer = layers[i];
38
39        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
40
41        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
42        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
43
44        // optional bias tensors
45        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
46
47        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
48
49        if (arch == LLM_ARCH_ERNIE4_5_MOE && static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
50            int n_ff_exp = hparams.n_ff_exp();
51
52            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
53            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
54            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
55            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);
56            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
57
58            // Shared expert (if present)
59            if (hparams.n_ff_shexp > 0) {
60                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, hparams.n_ff_shexp}, 0);
61                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd    }, 0);
62                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, hparams.n_ff_shexp}, 0);
63            }
64        } else { // Dense layers
65            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
66            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
67            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
68        }
69    }
70}
71
72std::unique_ptr<llm_graph_context> llama_model_ernie4_5::build_arch_graph(const llm_graph_params & params) const {
73    return std::make_unique<graph>(*this, params);
74}
75
76llama_model_ernie4_5::graph::graph(const llama_model & model, const llm_graph_params & params) :
77    llm_graph_context(params) {
78    const int64_t n_embd_head = hparams.n_embd_head_v();
79
80    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
81    GGML_ASSERT(n_embd_head == n_rot);
82
83    ggml_tensor * cur;
84    ggml_tensor * inpL;
85
86    inpL = build_inp_embd(model.tok_embd);
87
88    // inp_pos - contains the positions
89    ggml_tensor * inp_pos = build_inp_pos();
90
91    auto * inp_attn = build_attn_inp_kv();
92
93    ggml_tensor * inp_out_ids = build_inp_out_ids();
94
95    for (int il = 0; il < n_layer; ++il) {
96        ggml_tensor * inpSA = inpL;
97
98        // norm
99        {
100            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
101            cb(cur, "attn_norm", il);
102        }
103        // self-attention
104        {
105            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
106                    n_embd_head, n_head, n_head_kv, il);
107
108            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
109                                 ext_factor, attn_factor, beta_fast, beta_slow);
110
111            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
112                                 ext_factor, attn_factor, beta_fast, beta_slow);
113
114            cb(Qcur, "Qcur", il);
115            cb(Kcur, "Kcur", il);
116            cb(Vcur, "Vcur", il);
117
118            cur = build_attn(inp_attn,
119                    model.layers[il].wo, NULL, model.layers[il].wo_s,
120                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
121        }
122        if (il == n_layer - 1) {
123            // skip computing output for unused tokens
124            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
125            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
126        }
127        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
128        cb(ffn_inp, "ffn_inp", il);
129
130        // feed-forward network
131        {
132            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
133            cb(cur, "ffn_norm", il);
134
135            cur = build_ffn(cur,
136                    model.layers[il].ffn_up, NULL, NULL,
137                    model.layers[il].ffn_gate, NULL, NULL,
138                    model.layers[il].ffn_down, NULL, NULL,
139                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
140            cb(cur, "ffn_out", il);
141        }
142        cur = ggml_add(ctx0, cur, ffn_inp);
143
144        cur = build_cvec(cur, il);
145        cb(cur, "l_out", il);
146
147        // input for next layer
148        inpL = cur;
149    }
150    cur = inpL;
151
152    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
153
154    cb(cur, "result_norm", -1);
155    res->t_embd = cur;
156
157    // lm_head
158    cur = build_lora_mm(model.output, cur, model.output_s);
159
160    cb(cur, "result_output", -1);
161    res->t_logits = cur;
162
163    ggml_build_forward_expand(gf, cur);
164}