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 KiB177 linesraw
1#include "models.h"
2
3void llama_model_openai_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_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
7
8    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
9    uint32_t swa_period = 2;
10    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
11    hparams.set_swa_pattern(swa_period);
12
13    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
14    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
15    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
16
17    switch (hparams.n_layer()) {
18        case 24: type = LLM_TYPE_20B; break;
19        case 36: type = LLM_TYPE_120B; break;
20        default: type = LLM_TYPE_UNKNOWN;
21    }
22}
23
24void llama_model_openai_moe::load_arch_tensors(llama_model_loader &) {
25    LLAMA_LOAD_LOCALS;
26
27    const int64_t n_ff_exp = hparams.n_ff_exp();
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}, 0);
34
35    for (int i = 0; i < n_layer; ++i) {
36        auto & layer = layers[i];
37
38        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
39        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
40
41        create_tensor_qkv(layer, i, n_embd, n_head * n_rot, n_head_kv * n_rot, n_head_kv * n_rot, 0);
42        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);
43
44        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, 0);
45
46        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {  n_embd, n_expert}, 0);
47        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
48        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
49        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
50
51        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
52
53        layer.ffn_gate_inp_b  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "bias", i), {n_expert}, 0);
54        layer.ffn_gate_exps_b = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0);
55        layer.ffn_down_exps_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "bias", i), {  n_embd, n_expert}, 0);
56        layer.ffn_up_exps_b   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "bias", i), {n_ff_exp, n_expert}, 0);
57    }
58}
59
60std::unique_ptr<llm_graph_context> llama_model_openai_moe::build_arch_graph(const llm_graph_params & params) const {
61    return std::make_unique<graph>(*this, params);
62}
63
64llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
65    ggml_tensor * cur;
66    ggml_tensor * inpL;
67
68    inpL = build_inp_embd(model.tok_embd);
69
70    // inp_pos - contains the positions
71    ggml_tensor * inp_pos = build_inp_pos();
72
73    auto * inp_attn = build_attn_inp_kv_iswa();
74
75    ggml_tensor * inp_out_ids = build_inp_out_ids();
76
77    for (int il = 0; il < n_layer; ++il) {
78        res->t_layer_inp[il] = inpL;
79
80        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
81        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
82
83        ggml_tensor * inpSA = inpL;
84
85        // norm
86        cur = build_norm(inpL,
87                model.layers[il].attn_norm, nullptr,
88                LLM_NORM_RMS, il);
89        cb(cur, "attn_norm", il);
90
91        // self-attention
92        {
93            // compute Q and K and RoPE them
94            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
95                    n_rot, n_head, n_head_kv, il);
96
97            Qcur = ggml_rope_ext(
98                    ctx0, Qcur, inp_pos, nullptr,
99                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
100                    ext_factor, attn_factor, beta_fast, beta_slow
101                    );
102
103            Kcur = ggml_rope_ext(
104                    ctx0, Kcur, inp_pos, nullptr,
105                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
106                    ext_factor, attn_factor, beta_fast, beta_slow
107                    );
108
109            cb(Qcur, "Qcur", il);
110            cb(Kcur, "Kcur", il);
111            cb(Vcur, "Vcur", il);
112
113            cur = build_attn(inp_attn,
114                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
115                    Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
116
117            cb(cur, "attn_out", il);
118        }
119        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
120            // skip computing output for unused tokens
121            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
122            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
123        }
124        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
125        cb(ffn_inp, "ffn_inp", il);
126
127        cur = ffn_inp;
128        cur = build_norm(cur,
129                model.layers[il].attn_post_norm, nullptr,
130                LLM_NORM_RMS, il);
131        cb(cur, "attn_post_norm", il);
132
133        // MoE branch
134        cur = build_moe_ffn(cur,
135                model.layers[il].ffn_gate_inp,  model.layers[il].ffn_gate_inp_b,
136                model.layers[il].ffn_up_exps,   model.layers[il].ffn_up_exps_b,
137                model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b,
138                model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b,
139                nullptr,
140                n_expert, n_expert_used,
141                LLM_FFN_SWIGLU_OAI_MOE, false,
142                hparams.expert_weights_scale,
143                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT,
144                il);
145        cb(cur, "ffn_moe_out", il);
146
147        cur = ggml_add(ctx0, cur, ffn_inp);
148
149        cur = build_cvec(cur, il);
150        cb(cur, "l_out", il);
151
152        // input for next layer
153        inpL = cur;
154    }
155    cur = inpL;
156
157    res->t_h_nextn = cur;
158
159    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
160        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
161    }
162
163    cur = build_norm(cur,
164            model.output_norm, NULL,
165            LLM_NORM_RMS, -1);
166
167    cb(cur, "result_norm", -1);
168    res->t_embd = cur;
169
170    // lm_head
171    cur = build_lora_mm(model.output, cur, model.output_s);
172
173    cb(cur, "result_output", -1);
174    res->t_logits = cur;
175
176    ggml_build_forward_expand(gf, cur);
177}