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 →
5.8 KiB163 linesraw
1#include "models.h"
2
3void llama_model_llada_moe::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all, false);
5    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
6
7    // diffusion language model uses non-causal attention
8    hparams.causal_attn = false;
9
10    switch (hparams.n_layer()) {
11        case 16: type = LLM_TYPE_A1_7B; break;
12        default: type = LLM_TYPE_UNKNOWN;
13    }
14}
15
16void llama_model_llada_moe::load_arch_tensors(llama_model_loader &) {
17    LLAMA_LOAD_LOCALS;
18
19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
20
21    // output
22    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
24
25    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for llada-moe");
26    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for llada-moe");
27
28    for (int i = 0; i < n_layer; ++i) {
29        auto & layer = layers[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, n_embd_gqa, n_embd_gqa, 0);
34        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
35        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
36        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
37
38        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
39
40        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
41
42        const int64_t n_ff_exp = hparams.n_ff_exp() ? hparams.n_ff_exp() : n_ff / n_expert_used;
43
44        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
45        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
46        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
47    }
48}
49
50std::unique_ptr<llm_graph_context> llama_model_llada_moe::build_arch_graph(const llm_graph_params & params) const {
51    return std::make_unique<graph>(*this, params);
52}
53
54llama_model_llada_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
55    const int64_t n_embd_head = hparams.n_embd_head_v();
56
57    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
58    GGML_ASSERT(n_embd_head == n_rot);
59
60    ggml_tensor * cur;
61    ggml_tensor * inpL;
62
63    inpL = build_inp_embd(model.tok_embd);
64
65    // inp_pos - contains the positions
66    ggml_tensor * inp_pos = build_inp_pos();
67
68    auto * inp_attn = build_attn_inp_no_cache();
69
70    ggml_tensor * inp_out_ids = build_inp_out_ids();
71
72    for (int il = 0; il < n_layer; ++il) {
73        ggml_tensor * inpSA = inpL;
74
75        // norm
76        cur = build_norm(inpL,
77                model.layers[il].attn_norm, NULL,
78                LLM_NORM_RMS, il);
79        cb(cur, "attn_norm", il);
80
81        // self_attention
82        {
83            // compute Q and K and RoPE them
84            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
85                    n_embd_head, n_head, n_head_kv, il);
86
87            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
88            cb(Qcur, "Qcur_normed", il);
89
90            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
91            cb(Kcur, "Kcur_normed", il);
92
93            Qcur = ggml_rope_ext(
94                    ctx0, Qcur, inp_pos, nullptr,
95                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
96                    ext_factor, attn_factor, beta_fast, beta_slow
97                    );
98
99            Kcur = ggml_rope_ext(
100                    ctx0, Kcur, inp_pos, nullptr,
101                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
102                    ext_factor, attn_factor, beta_fast, beta_slow
103                    );
104
105            cb(Qcur, "Qcur", il);
106            cb(Kcur, "Kcur", il);
107            cb(Vcur, "Vcur", il);
108
109            cur = build_attn(inp_attn,
110                    model.layers[il].wo, NULL, model.layers[il].wo_s,
111                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
112        }
113        if (il == n_layer - 1 && inp_out_ids) {
114            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
115            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
116        }
117        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
118        cb(ffn_inp, "ffn_inp", il);
119
120        // MoE branch
121        cur = build_norm(ffn_inp,
122                model.layers[il].ffn_norm, NULL,
123                LLM_NORM_RMS, il);
124        cb(cur, "ffn_norm", il);
125
126        cur = build_moe_ffn(cur,
127                model.layers[il].ffn_gate_inp,
128                model.layers[il].ffn_up_exps,
129                model.layers[il].ffn_gate_exps,
130                model.layers[il].ffn_down_exps,
131                nullptr,
132                n_expert, n_expert_used,
133                LLM_FFN_SILU, false,
134                hparams.expert_weights_scale,
135                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
136                il);
137        cb(cur, "ffn_moe_out", il);
138
139        cur = ggml_add(ctx0, cur, ffn_inp);
140
141        cur = build_cvec(cur, il);
142        cb(cur, "l_out", il);
143
144        // input for next layer
145        inpL = cur;
146    }
147    cur = inpL;
148
149    cur = build_norm(cur,
150            model.output_norm, NULL,
151            LLM_NORM_RMS, -1);
152
153    cb(cur, "result_norm", -1);
154    res->t_embd = cur;
155
156    // lm_head
157    cur = build_lora_mm(model.output, cur, model.output_s);
158
159    cb(cur, "result_output", -1);
160    res->t_logits = cur;
161
162    ggml_build_forward_expand(gf, cur);
163}