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 →
7.1 KiB187 linesraw
1#include "models.h"
2
3void llama_model_hunyuan_moe::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_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
7
8    switch (hparams.n_layer()) {
9        case 32: type = LLM_TYPE_A13B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_hunyuan_moe::load_arch_tensors(llama_model_loader &) {
15    LLAMA_LOAD_LOCALS;
16
17    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
18
19    // output
20    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
21    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
22    // if output is NULL, init from the input tok embed
23    if (output == NULL) {
24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25    }
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29        const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(i);
30
31        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
32
33        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
34        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
35
36        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
37        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
38
39        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
40
41        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
42        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, 0);
43        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
44        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
45
46        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0);
47        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
48        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
49    }
50}
51
52std::unique_ptr<llm_graph_context> llama_model_hunyuan_moe::build_arch_graph(const llm_graph_params & params) const {
53    return std::make_unique<graph>(*this, params);
54}
55
56llama_model_hunyuan_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
57    const int64_t n_embd_head = hparams.n_embd_head_v();
58
59    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
60    GGML_ASSERT(n_embd_head == n_rot);
61
62    ggml_tensor * cur;
63    ggml_tensor * inpL;
64
65    inpL = build_inp_embd(model.tok_embd);
66
67    // inp_pos - contains the positions
68    ggml_tensor * inp_pos = build_inp_pos();
69
70    auto * inp_attn = build_attn_inp_kv();
71
72    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
73
74    ggml_tensor * inp_out_ids = build_inp_out_ids();
75
76    for (int il = 0; il < n_layer; ++il) {
77        ggml_tensor * inpSA = inpL;
78
79        // norm
80        cur = build_norm(inpL,
81                model.layers[il].attn_norm, NULL,
82                LLM_NORM_RMS, il);
83        cb(cur, "attn_norm", il);
84
85        // self-attention
86        {
87            // rope freq factors for llama3; may return nullptr for llama2 and other models
88            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
89
90            // compute Q and K and RoPE them
91            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
92                    n_embd_head, n_head, n_head_kv, il);
93
94            Qcur = ggml_rope_ext(
95                    ctx0, Qcur, inp_pos, rope_factors,
96                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
97                    ext_factor, attn_factor, beta_fast, beta_slow
98                    );
99
100            cb(Qcur, "Qcur", il);
101            cb(Kcur, "Kcur", il);
102            cb(Vcur, "Vcur", il);
103
104            Kcur = ggml_rope_ext(
105                    ctx0, Kcur, inp_pos, rope_factors,
106                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
107                    ext_factor, attn_factor, beta_fast, beta_slow
108                    );
109
110            Kcur = build_norm(Kcur,
111                    model.layers[il].attn_k_norm, nullptr,
112                    LLM_NORM_RMS, il);
113            cb(Kcur, "Kcur_norm", il);
114
115            Qcur = build_norm(Qcur,
116                    model.layers[il].attn_q_norm, nullptr,
117                    LLM_NORM_RMS, il);
118            cb(Qcur, "Qcur_norm", il);
119
120            cur = build_attn(inp_attn,
121                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
122                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
123            cb(cur, "attn_out", il);
124        }
125        if (il == n_layer - 1 && inp_out_ids) {
126            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
127            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
128        }
129        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
130        cb(ffn_inp, "ffn_inp", il);
131
132        cur = build_norm(ffn_inp,
133            model.layers[il].ffn_norm, NULL,
134            LLM_NORM_RMS, il);
135        cb(cur, "ffn_norm", il);
136
137        // feed-forward network (non-MoE)
138        ggml_tensor * cur_mlp = build_ffn(cur,
139                model.layers[il].ffn_up_shexp,   NULL, NULL,
140                model.layers[il].ffn_gate_shexp, NULL, NULL,
141                model.layers[il].ffn_down_shexp, NULL, NULL,
142                NULL,
143                LLM_FFN_SILU, LLM_FFN_PAR, il);
144        cb(cur_mlp, "ffn_mlp", il);
145
146        // MoE branch
147        ggml_tensor * cur_moe = build_moe_ffn(cur,
148                model.layers[il].ffn_gate_inp,
149                model.layers[il].ffn_up_exps,
150                model.layers[il].ffn_gate_exps,
151                model.layers[il].ffn_down_exps,
152                nullptr,
153                n_expert, n_expert_used,
154                LLM_FFN_SILU,
155                true, // norm_topk_prob
156                hparams.expert_weights_scale,
157                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
158                il);
159        cb(cur_moe, "ffn_moe_out", il);
160
161        ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp);
162        cb(ffn_out, "ffn_out", il);
163
164        cur = ggml_add(ctx0, ffn_out, ffn_inp);
165
166        cur = build_cvec(cur, il);
167        cb(cur, "l_out", il);
168
169        // input for next layer
170        inpL = cur;
171    }
172    cur = inpL;
173
174    cur = build_norm(cur,
175            model.output_norm, NULL,
176            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    cb(cur, "result_output", -1);
184    res->t_logits = cur;
185
186    ggml_build_forward_expand(gf, cur);
187}