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 →
8.8 KiB223 linesraw
1#include "models.h"
2
3void llama_model_grok::load_arch_hparams(llama_model_loader & ml) {
4    // defaults for old GGUFs
5    hparams.yarn_beta_fast = 8.0f;
6    hparams.f_logit_scale = 0.5773502691896257f;
7    hparams.f_embedding_scale = 78.38367176906169f;
8    hparams.f_attn_out_scale = 0.08838834764831845f;
9    hparams.f_attn_logit_softcapping = 30.0f;
10    hparams.f_router_logit_softcapping = 30.0f;
11    // no final_logit_softcapping in grok-1
12    hparams.f_final_logit_softcapping = 0.0f;
13
14    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
15    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
16    ml.get_key(LLM_KV_LOGIT_SCALE,                  hparams.f_logit_scale, false);
17    ml.get_key(LLM_KV_EMBEDDING_SCALE,              hparams.f_embedding_scale, false);
18    ml.get_key(LLM_KV_ATTENTION_OUTPUT_SCALE,       hparams.f_attn_out_scale, false);
19    ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,       hparams.f_attn_logit_softcapping, false);
20    ml.get_key(LLM_KV_ROUTER_LOGIT_SOFTCAPPING,     hparams.f_router_logit_softcapping, false);
21    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,      hparams.f_final_logit_softcapping, false);
22
23    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH,  hparams.attn_temp_length, false);
24    ml.get_key(LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,  hparams.yarn_ext_factor, false);
25    ml.get_key(LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, hparams.yarn_attn_factor, false);
26    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   hparams.yarn_beta_fast, false);
27    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   hparams.yarn_beta_slow, false);
28
29    switch (hparams.n_layer()) {
30        case 64: type = LLM_TYPE_314B; break;
31        default: type = LLM_TYPE_UNKNOWN;
32    }
33}
34
35void llama_model_grok::load_arch_tensors(llama_model_loader &) {
36    LLAMA_LOAD_LOCALS;
37
38    if (n_expert == 0) {
39        throw std::runtime_error(arch_name() + " model cannot have zero experts");
40    }
41
42    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
43
44    // output
45    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
46    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
47
48    // if output is NULL, init from the input tok embed
49    if (output == NULL) {
50        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
51    }
52
53    const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff/* / n_expert_used*/; // grok-1 n_ff_exp == n_ff
54    for (int i = 0; i < n_layer; ++i) {
55        auto & layer = layers[i];
56
57        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
58
59        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
60        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
61
62        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
63
64        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
65
66        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
67        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff,   n_embd}, TENSOR_NOT_REQUIRED);
68        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
69
70        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
71        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
72        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd,   n_expert}, 0);
73        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
74
75        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
76        if (!layer.ffn_post_norm) {
77            layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
78        }
79    }
80}
81
82std::unique_ptr<llm_graph_context> llama_model_grok::build_arch_graph(const llm_graph_params & params) const {
83    return std::make_unique<graph>(*this, params);
84}
85
86llama_model_grok::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
87    const int64_t n_embd_head = hparams.n_embd_head_v();
88
89    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
90    GGML_ASSERT(n_embd_head == n_rot);
91
92    ggml_tensor * cur;
93    ggml_tensor * inpL;
94
95    inpL = build_inp_embd(model.tok_embd);
96
97    // inp_pos - contains the positions
98    ggml_tensor * inp_pos = build_inp_pos();
99
100    auto * inp_attn = build_attn_inp_kv();
101
102    ggml_tensor * inp_out_ids = build_inp_out_ids();
103
104    for (int il = 0; il < n_layer; ++il) {
105        ggml_tensor * inpSA = inpL;
106
107        // norm
108        cur = build_norm(inpL,
109                model.layers[il].attn_norm, NULL,
110                LLM_NORM_RMS, il);
111        cb(cur, "attn_norm", il);
112
113        // self-attention
114        {
115            // compute Q and K and RoPE them
116            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
117                    n_embd_head, n_head, n_head_kv, il);
118
119            Qcur = ggml_rope_ext(
120                    ctx0, Qcur, inp_pos, nullptr,
121                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
122                    ext_factor, attn_factor, beta_fast, beta_slow
123                    );
124
125            Kcur = ggml_rope_ext(
126                    ctx0, Kcur, inp_pos, nullptr,
127                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
128                    ext_factor, attn_factor, beta_fast, beta_slow
129                    );
130
131            cb(Qcur, "Qcur", il);
132            cb(Kcur, "Kcur", il);
133            cb(Vcur, "Vcur", il);
134
135            cur = build_attn(inp_attn,
136                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
137                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
138        }
139        if (il == n_layer - 1 && inp_out_ids) {
140            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
141            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
142        }
143        cur = build_norm(cur,
144                model.layers[il].attn_out_norm, NULL,
145                LLM_NORM_RMS, il);
146        cb(cur, "attn_out_norm", il);
147
148        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
149        cb(ffn_inp, "ffn_inp", il);
150
151        // feed-forward network
152        cur = build_norm(ffn_inp,
153                model.layers[il].ffn_norm, NULL,
154                LLM_NORM_RMS, il);
155        cb(cur, "ffn_norm", il);
156
157        // MoE branch
158        ggml_tensor * moe_out = build_moe_ffn(cur,
159                model.layers[il].ffn_gate_inp,
160                model.layers[il].ffn_up_exps,
161                model.layers[il].ffn_gate_exps,
162                model.layers[il].ffn_down_exps,
163                nullptr,
164                n_expert, n_expert_used,
165                LLM_FFN_GELU, true,
166                hparams.expert_weights_scale,
167                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
168                il);
169        cb(moe_out, "ffn_moe_out", il);
170
171        if (model.layers[il].ffn_up) {
172            ggml_tensor * ffn_out = build_ffn(cur,
173                    model.layers[il].ffn_up,   NULL, NULL,
174                    model.layers[il].ffn_gate, NULL, NULL,
175                    model.layers[il].ffn_down, NULL, NULL,
176                    NULL,
177                    LLM_FFN_GELU, LLM_FFN_PAR, il);
178            cb(ffn_out, "ffn_out", il);
179
180            cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
181            cb(cur, "ffn_out", il);
182        } else {
183            cur = moe_out;
184        }
185        cur = build_norm(cur,
186                model.layers[il].ffn_post_norm, NULL,
187                LLM_NORM_RMS, il);
188        cb(cur, "ffn_post_norm", il);
189
190        cur = ggml_add(ctx0, cur, ffn_inp);
191        cb(cur, "ffn_out", il);
192
193        cur = build_cvec(cur, il);
194        cb(cur, "l_out", il);
195
196        // input for next layer
197        inpL = cur;
198    }
199    cur = inpL;
200
201    cur = build_norm(cur,
202            model.output_norm, NULL,
203            LLM_NORM_RMS, -1);
204
205    cb(cur, "result_norm", -1);
206    res->t_embd = cur;
207
208    // lm_head
209    cur = build_lora_mm(model.output, cur, model.output_s);
210
211    cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
212
213    // final logit soft-capping
214    if (hparams.f_final_logit_softcapping) {
215        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
216        cur = ggml_tanh(ctx0, cur);
217        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
218    }
219    cb(cur, "result_output", -1);
220    res->t_logits = cur;
221
222    ggml_build_forward_expand(gf, cur);
223}