char-fork/llama.cpp

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

PikaPikachuconvert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion (#22780)465e49b9c

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.2 KiB168 linesraw
1#include "models.h"
2
3void llama_model_minimax_m2::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_GATING_FUNC,           hparams.expert_gating_func, false);
7
8    switch (hparams.n_layer()) {
9        case 62: type = LLM_TYPE_230B_A10B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11    }
12}
13
14void llama_model_minimax_m2::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}, 0);
22
23    for (int i = 0; i < n_layer; ++i) {
24        auto & layer = layers[i];
25
26        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
27        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
28
29        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
30        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k * n_head}, 0);
31        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_k_gqa}, 0);
32
33        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
34
35        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
36        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
37        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
38        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
39        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
40    }
41}
42
43std::unique_ptr<llm_graph_context> llama_model_minimax_m2::build_arch_graph(const llm_graph_params & params) const {
44    return std::make_unique<graph>(*this, params);
45}
46
47llama_model_minimax_m2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
48    const int64_t n_embd_head = hparams.n_embd_head_v();
49
50    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
51    // GGML_ASSERT(n_embd_head == n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64
52
53    ggml_tensor * cur;
54    ggml_tensor * inpL;
55
56    inpL = build_inp_embd(model.tok_embd);
57
58    ggml_tensor * inp_pos = build_inp_pos();
59    auto inp_attn = build_attn_inp_kv();
60    ggml_tensor * inp_out_ids = build_inp_out_ids();
61
62    for (int il = 0; il < n_layer; ++il) {
63        res->t_layer_inp[il] = inpL;
64
65        ggml_tensor * inpSA = inpL;
66
67        cur = inpL;
68
69        // self_attention
70        {
71            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
72            cb(cur, "attn_norm", il);
73
74            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
75                    n_embd_head, n_head,
76                    n_embd_head, n_head_kv,
77                    n_embd_head, n_head_kv,
78                    il, false);
79            cb(Qcur, "Qcur", il);
80            cb(Kcur, "Kcur", il);
81            cb(Vcur, "Vcur", il);
82
83            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL,
84                    LLM_NORM_RMS, il);
85            cb(Qcur, "Qcur_normed", il);
86
87            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL,
88                    LLM_NORM_RMS, il);
89            cb(Kcur, "Kcur_normed", il);
90
91            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
92            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
93            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
94
95            Qcur = ggml_rope_ext(
96                ctx0, Qcur, inp_pos, nullptr,
97                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
98                ext_factor, attn_factor, beta_fast, beta_slow
99                );
100
101            Kcur = ggml_rope_ext(
102                ctx0, Kcur, inp_pos, nullptr,
103                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
104                ext_factor, attn_factor, beta_fast, beta_slow
105                );
106
107            cb(Qcur, "Qcur", il);
108            cb(Kcur, "Kcur", il);
109            cb(Vcur, "Vcur", il);
110
111            cur = build_attn(inp_attn,
112                    model.layers[il].wo, NULL, model.layers[il].wo_s,
113                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
114        }
115
116        if (il == n_layer - 1 && inp_out_ids) {
117            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
118            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
119        }
120
121        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
122        cb(ffn_inp, "ffn_inp", il);
123
124        // MoE branch
125        cur = build_norm(ffn_inp,
126                model.layers[il].ffn_norm, NULL,
127                LLM_NORM_RMS, il);
128        cb(cur, "ffn_norm", il);
129
130        cur = build_moe_ffn(cur,
131                model.layers[il].ffn_gate_inp,
132                model.layers[il].ffn_up_exps,
133                model.layers[il].ffn_gate_exps,
134                model.layers[il].ffn_down_exps,
135                model.layers[il].ffn_exp_probs_b,
136                n_expert, n_expert_used,
137                LLM_FFN_SILU, true,
138                hparams.expert_weights_scale,
139                (llama_expert_gating_func_type) hparams.expert_gating_func,
140                il);
141        cb(cur, "ffn_moe_out", il);
142
143        cur = ggml_add(ctx0, cur, ffn_inp);
144
145        cur = build_cvec(cur, il);
146        cb(cur, "l_out", il);
147
148        // input for next layer
149        inpL = cur;
150    }
151
152    cur = inpL;
153
154    cur = build_norm(cur,
155            model.output_norm, NULL,
156            LLM_NORM_RMS, -1);
157
158    cb(cur, "result_norm", -1);
159    res->t_embd = cur;
160
161    // lm_head
162    cur = build_lora_mm(model.output, cur, model.output_s);
163
164    cb(cur, "result_output", -1);
165    res->t_logits = cur;
166
167    ggml_build_forward_expand(gf, cur);
168}