char-fork/llama.cpp

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

Sigbjørn Skjæretmodel : load hparams.n_layer_nextn before n_layer() calls (#28159)9d817213a

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
7.5 KiB192 linesraw
1#include "models.h"
2
3void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
4    if (hparams.n_layer() == 64) {    // 32B
5        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
6        hparams.n_swa = 4096;
7        uint32_t swa_period = 4;
8        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
9        hparams.set_swa_pattern(swa_period);
10
11        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
12        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
13        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
14    }
15
16    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
17    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
18
19    switch (hparams.n_layer()) {
20        case 30: type = LLM_TYPE_1_2B; break;
21        case 64: type = LLM_TYPE_32B; break;
22        default: type = LLM_TYPE_UNKNOWN;
23    }
24}
25
26void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {
27    LLAMA_LOAD_LOCALS;
28
29    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
30
31    // output
32    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
33    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    if (output == NULL) {
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_all; ++i) {
41        const bool is_nextn = i >= n_layer;
42        int flags = 0;
43        if (is_nextn) {
44            // NextN/MTP layers are preserved in GGUF but are not executed yet.
45            flags |= TENSOR_SKIP;
46        }
47
48        auto & layer = layers[i];
49
50        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
51        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, flags);
52
53        if (!is_nextn) {
54            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
55        }
56
57        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
58        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
59        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
60
61        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
62        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
63        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
64        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
65
66        if (is_nextn) {
67            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
68            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);
69            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);
70            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
71        }
72    }
73}
74
75std::unique_ptr<llm_graph_context> llama_model_exaone4::build_arch_graph(const llm_graph_params & params) const {
76    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
77        return std::make_unique<graph<true>>(*this, params);
78    } else {
79        return std::make_unique<graph<false>>(*this, params);
80    }
81}
82
83template <bool iswa>
84llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :
85    llm_graph_context(params) {
86    const int64_t n_embd_head = hparams.n_embd_head_k();
87
88    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
89    GGML_ASSERT(n_embd_head == n_rot);
90
91    ggml_tensor * cur;
92    ggml_tensor * inpL;
93
94    inpL = build_inp_embd(model.tok_embd);
95
96    // inp_pos - contains the positions
97    ggml_tensor * inp_pos = build_inp_pos();
98
99    using inp_attn_type      = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
100    inp_attn_type * inp_attn = nullptr;
101
102    if constexpr (iswa) {
103        inp_attn = build_attn_inp_kv_iswa();
104    } else {
105        inp_attn = build_attn_inp_kv();
106    }
107    ggml_tensor * inp_out_ids = build_inp_out_ids();
108
109    for (int il = 0; il < n_layer; ++il) {
110        ggml_tensor * inpSA = inpL;
111
112        // use RoPE for SWA layers or non-SWA models
113        const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE;
114
115        cur = inpL;
116
117        // self-attention
118        {
119            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
120
121            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
122                    n_embd_head, n_head, n_head_kv, il);
123
124            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
125            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
126            cb(Qcur, "Qcur_normed", il);
127            cb(Kcur, "Kcur_normed", il);
128
129            if (use_rope) {
130                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
131                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
132
133                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
134                                     freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
135            }
136            cb(Qcur, "Qcur", il);
137            cb(Kcur, "Kcur", il);
138            cb(Vcur, "Vcur", il);
139
140            cur = build_attn(inp_attn,
141                    model.layers[il].wo, NULL, model.layers[il].wo_s,
142                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
143            cb(cur, "attn_out", il);
144        }
145        if (il == n_layer - 1 && inp_out_ids) {
146            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
147            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
148        }
149        cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
150        cb(cur, "attn_post_norm", il);
151
152        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
153        cb(ffn_inp, "ffn_inp", il);
154
155        // feed-forward network
156        cur = build_ffn(ffn_inp,
157                model.layers[il].ffn_up, NULL, NULL,
158                model.layers[il].ffn_gate, NULL, NULL,
159                model.layers[il].ffn_down, NULL, NULL, NULL,
160                LLM_FFN_SILU, LLM_FFN_PAR, il);
161        cb(cur, "ffn_out", il);
162
163        cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1);
164        cb(cur, "ffn_post_norm", -1);
165
166        cur = ggml_add(ctx0, cur, ffn_inp);
167
168        cur = build_cvec(cur, il);
169        cb(cur, "l_out", il);
170
171        // input for next layer
172        inpL = cur;
173    }
174    cur = inpL;
175
176    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
177
178    cb(cur, "result_norm", -1);
179    res->t_embd = cur;
180
181    // lm_head
182    cur = build_lora_mm(model.output, cur, model.output_s);
183
184    cb(cur, "result_output", -1);
185    res->t_logits = cur;
186
187    ggml_build_forward_expand(gf, cur);
188}
189
190// Explicit template instantiations
191template struct llama_model_exaone4::graph<false>;
192template struct llama_model_exaone4::graph<true>;