char-fork/llama.cpp

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

Georgi Gerganovhparams : refactor `hparams.n_layer` (#24060)7acb4e8cd

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
6.3 KiB170 linesraw
1#include "models.h"
2
3void llama_model_bitnet::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    switch (hparams.n_layer()) {
7        case 26: type = LLM_TYPE_3B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_bitnet::load_arch_tensors(llama_model_loader &) {
13    LLAMA_LOAD_LOCALS;
14
15    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
16
17    // output
18    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
19
20    for (int i = 0; i < n_layer; ++i) {
21        auto & layer = layers[i];
22
23        layer.attn_norm     = create_tensor(tn(LLM_TENSOR_ATTN_NORM,     "weight", i), {n_embd}, 0);
24        layer.attn_sub_norm = create_tensor(tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd}, 0);
25
26        layer.wq       = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
27        layer.wq_s     = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
28        layer.wk       = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
29        layer.wk_s     = create_tensor(tn(LLM_TENSOR_ATTN_K,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
30        layer.wv       = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
31        layer.wv_s     = create_tensor(tn(LLM_TENSOR_ATTN_V,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
32        layer.wo       = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
33        layer.wo_s     = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
34
35        layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,     "weight", i), {n_embd}, 0);
36        layer.ffn_sub_norm = create_tensor(tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff}, 0);
37
38        layer.ffn_gate       = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
39        layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
40        layer.ffn_down       = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
41        layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
42        layer.ffn_up         = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
43        layer.ffn_up_s   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
44    }
45}
46
47std::unique_ptr<llm_graph_context> llama_model_bitnet::build_arch_graph(const llm_graph_params & params) const {
48    return std::make_unique<graph>(*this, params);
49}
50
51llama_model_bitnet::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
52    const int64_t n_embd_head = hparams.n_embd_head_v();
53
54    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
55
56    ggml_tensor * cur;
57    ggml_tensor * inpL;
58
59    inpL = build_inp_embd(model.tok_embd);
60
61    // inp_pos - contains the positions
62    ggml_tensor * inp_pos = build_inp_pos();
63
64    auto * inp_attn = build_attn_inp_kv();
65
66    ggml_tensor * inp_out_ids = build_inp_out_ids();
67
68    for (int il = 0; il < n_layer; ++il) {
69        ggml_tensor * inpSA = inpL;
70
71        cur = build_norm(inpL,
72                model.layers[il].attn_norm, NULL,
73                LLM_NORM_RMS, il);
74        cb(cur, "attn_norm", il);
75
76        // self-attention
77        {
78            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
79                    n_embd_head, n_head, n_head_kv, il);
80
81            Qcur = ggml_rope_ext(
82                    ctx0, Qcur, inp_pos, nullptr,
83                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
84                    ext_factor, attn_factor, beta_fast, beta_slow
85                    );
86
87            Kcur = ggml_rope_ext(
88                    ctx0, Kcur, inp_pos, nullptr,
89                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
90                    ext_factor, attn_factor, beta_fast, beta_slow
91                    );
92
93            cb(Qcur, "Qcur", il);
94            cb(Kcur, "Kcur", il);
95            cb(Vcur, "Vcur", il);
96
97            cur = build_attn(inp_attn,
98                    NULL, NULL, NULL,
99                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
100
101            cur = build_norm(cur,
102                    model.layers[il].attn_sub_norm, NULL,
103                    LLM_NORM_RMS, il);
104            cb(cur, "attn_sub_norm", il);
105
106            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
107            if (model.layers[il].wo_b) {
108                cur = ggml_add(ctx0, cur, model.layers[il].wo_b);
109            }
110            cb(cur, "attn_out", il);
111        }
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
118        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
119        cb(ffn_inp, "ffn_inp", il);
120
121        // feed-forward forward
122        cur = build_norm(ffn_inp,
123                model.layers[il].ffn_norm, NULL,
124                LLM_NORM_RMS, il);
125        cb(cur, "ffn_norm", il);
126
127        cur = build_ffn(cur,
128                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,
129                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
130                NULL,                      NULL, NULL,
131                NULL,
132                LLM_FFN_SILU, LLM_FFN_PAR, il);
133        cb(cur, "ffn_sub_out", il);
134
135        cur = build_norm(cur,
136                model.layers[il].ffn_sub_norm, NULL,
137                LLM_NORM_RMS, il);
138        cb(cur, "ffn_sub_norm", il);
139
140        cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
141        cb(cur, "ffn_down", il);
142
143        cur = ggml_add(ctx0, cur, ffn_inp);
144        cb(cur, "l_out", il);
145
146        cur = build_cvec(cur, il);
147        cb(cur, "l_out", il);
148
149        // input for next layer
150        inpL = cur;
151    }
152
153    cur = inpL;
154
155    cur = build_norm(cur,
156            model.output_norm, NULL,
157            LLM_NORM_RMS, -1);
158
159    cb(cur, "result_norm", -1);
160    res->t_embd = cur;
161
162    // lm_head
163    // FIXME: do not use model.tok_embd directly, duplicate as model.output
164    cur = build_lora_mm(model.tok_embd, cur);
165
166    cb(cur, "result_output", -1);
167    res->t_logits = cur;
168
169    ggml_build_forward_expand(gf, cur);
170}