char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Alexllama: add Maple 20B-A1B ternary MoE architecture (CPU) (#27000)3d10bcd19

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.2 KiB150 linesraw
1#include "models.h"
2
3void llama_model_maple::load_arch_hparams(llama_model_loader & ml) {
4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
5
6    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
7    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
8    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
9
10    ml.get_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl);
11
12    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
13    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
14    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
15
16    ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer_all);
17
18    switch (hparams.n_layer()) {
19        case 24: type = LLM_TYPE_20B; break;
20        default: type = LLM_TYPE_UNKNOWN;
21    }
22}
23
24void llama_model_maple::load_arch_tensors(llama_model_loader &) {
25    LLAMA_LOAD_LOCALS;
26
27    const int64_t n_ff_exp = hparams.n_ff_exp();
28    const int64_t head_dim = hparams.n_embd_head_k();
29
30    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
31
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    if (n_expert == 0) {
36        throw std::runtime_error("n_expert must be > 0 for Maple");
37    }
38    if (n_expert_used == 0) {
39        throw std::runtime_error("n_expert_used must be > 0 for Maple");
40    }
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_head * head_dim, n_head_kv * head_dim, n_head_kv * head_dim, 0);
48        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * head_dim, n_embd}, 0);
49
50        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {head_dim}, 0);
51        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {head_dim}, 0);
52        layer.ffn_norm    = create_tensor(tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd}, 0);
53
54        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
55        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
56        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
57        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
58    }
59}
60
61std::unique_ptr<llm_graph_context> llama_model_maple::build_arch_graph(const llm_graph_params & params) const {
62    return std::make_unique<graph>(*this, params);
63}
64
65llama_model_maple::graph::graph(const llama_model & model, const llm_graph_params & params) :
66    llm_graph_context(params) {
67    const int64_t n_embd_head = hparams.n_embd_head_k();
68
69    GGML_ASSERT(n_embd_head == hparams.n_embd_head_v());
70
71    ggml_tensor * inpL = build_inp_embd(model.tok_embd);
72    ggml_tensor * inp_pos = build_inp_pos();
73    auto * inp_attn = build_attn_inp_kv_iswa();
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        ggml_tensor * cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
80        cb(cur, "attn_norm", il);
81
82        {
83            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il);
84
85            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
86            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
87            cb(Qcur, "Qcur_normed", il);
88            cb(Kcur, "Kcur_normed", il);
89
90            if (hparams.is_swa(il)) {
91                const int64_t n_rot_l = hparams.n_rot(il);
92                const float freq_base_l = model.get_rope_freq_base(cparams, il);
93                const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
94
95                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l,
96                                     freq_scale_l, ext_factor, attn_factor, beta_fast, beta_slow);
97                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot_l, rope_type, n_ctx_orig, freq_base_l,
98                                     freq_scale_l, ext_factor, attn_factor, beta_fast, beta_slow);
99            }
100            cb(Qcur, "Qcur", il);
101            cb(Kcur, "Kcur", il);
102            cb(Vcur, "Vcur", il);
103
104            cur = build_attn(inp_attn,
105                    model.layers[il].wo, nullptr, model.layers[il].wo_s,
106                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
107            cb(cur, "attn_out", il);
108        }
109
110        if (il == n_layer - 1 && inp_out_ids) {
111            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
112            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
113        }
114
115        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
116        cb(ffn_inp, "ffn_inp", il);
117
118        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
119        cb(cur, "ffn_norm", il);
120
121        cur = build_moe_ffn(cur,
122                model.layers[il].ffn_gate_inp,
123                model.layers[il].ffn_up_exps,
124                model.layers[il].ffn_gate_exps,
125                model.layers[il].ffn_down_exps,
126                nullptr,
127                n_expert, n_expert_used,
128                LLM_FFN_SILU, true,
129                1.0f,
130                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
131                il);
132        cb(cur, "ffn_moe_out", il);
133
134        cur = ggml_add(ctx0, cur, ffn_inp);
135        cur = build_cvec(cur, il);
136        cb(cur, "l_out", il);
137
138        inpL = cur;
139    }
140
141    ggml_tensor * cur = build_norm(inpL, model.output_norm, nullptr, LLM_NORM_RMS, -1);
142    cb(cur, "result_norm", -1);
143    res->t_embd = cur;
144
145    cur = build_lora_mm(model.output, cur, model.output_s);
146    cb(cur, "result_output", -1);
147    res->t_logits = cur;
148
149    ggml_build_forward_expand(gf, cur);
150}