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 →
8.3 KiB225 linesraw
1#include "models.h"
2
3void llama_model_mellum::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
6    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
7
8    if (hparams.n_swa > 0) {
9        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
10
11        uint32_t swa_period = 4;
12        const auto res = ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
13        if (res) {
14            hparams.set_swa_pattern(swa_period);
15        } else {
16            ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
17        }
18
19        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
20        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
21
22        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
23    } else {
24        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
25    }
26
27    switch (hparams.n_layer()) {
28        case 28: type = LLM_TYPE_12B_A2_5B; break;
29        default: type = LLM_TYPE_UNKNOWN;
30    }
31}
32
33void llama_model_mellum::load_arch_tensors(llama_model_loader &) {
34    LLAMA_LOAD_LOCALS;
35
36    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
37
38    // output
39    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
40    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
41
42    for (int i = 0; i < n_layer; ++i) {
43        auto & layer = layers[i];
44
45        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
46
47        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
48        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
49
50        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
51        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
52
53        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
54
55        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
56
57        if (n_expert == 0) {
58            throw std::runtime_error("n_expert must be > 0 for Mellum");
59        }
60        if (n_expert_used == 0) {
61            throw std::runtime_error("n_expert_used must be > 0 for Mellum");
62        }
63
64        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
65
66        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
67        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
68        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
69    }
70}
71
72std::unique_ptr<llm_graph_context> llama_model_mellum::build_arch_graph(const llm_graph_params & params) const {
73    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
74        return std::make_unique<graph<true>>(*this, params);
75    }
76    return std::make_unique<graph<false>>(*this, params);
77}
78
79template <bool iswa>
80llama_model_mellum::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
81    const int64_t n_embd_head = hparams.n_embd_head_v();
82
83    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
84    GGML_ASSERT(n_embd_head == n_rot);
85
86    ggml_tensor * cur;
87    ggml_tensor * inpL;
88
89    inpL = build_inp_embd(model.tok_embd);
90
91    // inp_pos - contains the positions
92    ggml_tensor * inp_pos = build_inp_pos();
93
94    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
95    inp_attn_type * inp_attn = nullptr;
96
97    if constexpr (iswa) {
98        inp_attn = build_attn_inp_kv_iswa();
99    } else {
100        inp_attn = build_attn_inp_kv();
101    }
102
103    ggml_tensor * inp_out_ids = build_inp_out_ids();
104
105    for (int il = 0; il < n_layer; ++il) {
106        ggml_tensor * inpSA = inpL;
107
108        // norm
109        cur = build_norm(inpL,
110                model.layers[il].attn_norm, nullptr,
111                LLM_NORM_RMS, il);
112        cb(cur, "attn_norm", il);
113
114        // self_attention
115        {
116            // compute Q and K and RoPE them
117            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
118                    n_embd_head, n_head, n_head_kv, il);
119
120            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
121            cb(Qcur, "Qcur_normed", il);
122
123            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
124            cb(Kcur, "Kcur_normed", il);
125
126            const bool is_swa = hparams.is_swa(il);
127
128            if (is_swa) {
129                // For sliding window layers, use regular rope with no yarn rope scaling.
130                // This is achieved here by setting freq_scale and attn_factor to 1.
131                // We also set ext_factor to 0 to avoid a few unnecessary computations.
132                Qcur = ggml_rope_ext(
133                    ctx0, Qcur, inp_pos, nullptr,
134                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
135                    0.0, 1.0, beta_fast, beta_slow
136                    );
137
138                Kcur = ggml_rope_ext(
139                    ctx0, Kcur, inp_pos, nullptr,
140                    n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
141                    0.0, 1.0, beta_fast, beta_slow
142                    );
143            } else {
144                Qcur = ggml_rope_ext(
145                    ctx0, Qcur, inp_pos, nullptr,
146                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
147                    ext_factor, attn_factor, beta_fast, beta_slow
148                    );
149
150                Kcur = ggml_rope_ext(
151                    ctx0, Kcur, inp_pos, nullptr,
152                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
153                    ext_factor, attn_factor, beta_fast, beta_slow
154                    );
155            }
156
157            cb(Qcur, "Qcur", il);
158            cb(Kcur, "Kcur", il);
159            cb(Vcur, "Vcur", il);
160
161            cur = build_attn(inp_attn,
162                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
163                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
164        }
165        if (il == n_layer - 1 && inp_out_ids) {
166            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
167            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
168        }
169        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
170        cb(ffn_inp, "ffn_inp", il);
171
172        // MoE
173        cur = build_norm(ffn_inp,
174                model.layers[il].ffn_norm, nullptr,
175                LLM_NORM_RMS, il);
176        cb(cur, "ffn_norm", il);
177
178        ggml_tensor * moe_out =
179            build_moe_ffn(cur,
180                    model.layers[il].ffn_gate_inp,
181                    model.layers[il].ffn_up_exps,
182                    model.layers[il].ffn_gate_exps,
183                    model.layers[il].ffn_down_exps,
184                    nullptr,
185                    n_expert, n_expert_used,
186                    LLM_FFN_SILU, true,
187                    hparams.expert_weights_scale,
188                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
189                    il,
190                    nullptr, nullptr,
191                    model.layers[il].ffn_up_exps_s,
192                    model.layers[il].ffn_gate_exps_s,
193                    model.layers[il].ffn_down_exps_s);
194        cb(moe_out, "ffn_moe_out", il);
195        cur = moe_out;
196
197        cur = ggml_add(ctx0, cur, ffn_inp);
198        cb(cur, "ffn_out", il);
199
200        cur = build_cvec(cur, il);
201        cb(cur, "l_out", il);
202
203        // input for next layer
204        inpL = cur;
205    }
206    cur = inpL;
207
208    cur = build_norm(cur,
209            model.output_norm, nullptr,
210            LLM_NORM_RMS, -1);
211
212    cb(cur, "result_norm", -1);
213    res->t_embd = cur;
214
215    // lm_head
216    cur = build_lora_mm(model.output, cur, model.output_s);
217
218    cb(cur, "result_output", -1);
219    res->t_logits = cur;
220
221    ggml_build_forward_expand(gf, cur);
222}
223
224template struct llama_model_mellum::graph<false>;
225template struct llama_model_mellum::graph<true>;