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 →
4.7 KiB140 linesraw
1#include "models.h"
2
3void llama_model_qwen::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 32: type = LLM_TYPE_7B; break;
8        case 40: type = LLM_TYPE_13B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_qwen::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
17
18    // output
19    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
20    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
21
22    for (int i = 0; i < n_layer; ++i) {
23        auto & layer = layers[i];
24
25        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
26
27        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd*3}, 0);
28        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd*3}, 0);
29        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
30
31        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
32
33        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff/2}, 0);
34        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff/2, n_embd}, 0);
35        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff/2}, 0);
36    }
37}
38
39std::unique_ptr<llm_graph_context> llama_model_qwen::build_arch_graph(const llm_graph_params & params) const {
40    return std::make_unique<graph>(*this, params);
41}
42
43llama_model_qwen::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
44    const int64_t n_embd_head = hparams.n_embd_head_v();
45
46    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
47
48    ggml_tensor * cur;
49    ggml_tensor * inpL;
50
51    inpL = build_inp_embd(model.tok_embd);
52
53    // inp_pos - contains the positions
54    ggml_tensor * inp_pos = build_inp_pos();
55
56    auto * inp_attn = build_attn_inp_kv();
57
58    ggml_tensor * inp_out_ids = build_inp_out_ids();
59
60    for (int il = 0; il < n_layer; ++il) {
61        ggml_tensor * inpSA = inpL;
62
63        cur = build_norm(inpL,
64                model.layers[il].attn_norm, NULL,
65                LLM_NORM_RMS, il);
66        cb(cur, "attn_norm", il);
67
68        // self-attention
69        {
70            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
71                    n_embd_head, n_head, n_head_kv, il);
72
73            // using mode = 2 for neox mode
74            Qcur = ggml_rope_ext(
75                    ctx0, Qcur, inp_pos, nullptr,
76                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
77                    ext_factor, attn_factor, beta_fast, beta_slow
78                    );
79
80            Kcur = ggml_rope_ext(
81                    ctx0, Kcur, inp_pos, nullptr,
82                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
83                    ext_factor, attn_factor, beta_fast, beta_slow
84                    );
85
86            cb(Qcur, "Qcur", il);
87            cb(Kcur, "Kcur", il);
88            cb(Vcur, "Vcur", il);
89
90            cur = build_attn(inp_attn,
91                    model.layers[il].wo, NULL, model.layers[il].wo_s,
92                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
93        }
94        if (il == n_layer - 1 && inp_out_ids) {
95            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
96            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
97        }
98        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
99        cb(ffn_inp, "ffn_inp", il);
100
101        // feed-forward forward
102        {
103            cur = build_norm(ffn_inp,
104                    model.layers[il].ffn_norm, NULL,
105                    LLM_NORM_RMS, il);
106            cb(cur, "ffn_norm", il);
107
108            cur = build_ffn(cur,
109                    model.layers[il].ffn_up,   NULL, NULL,
110                    model.layers[il].ffn_gate, NULL, NULL,
111                    model.layers[il].ffn_down, NULL, NULL,
112                    NULL,
113                    LLM_FFN_SILU, LLM_FFN_PAR, il);
114            cb(cur, "ffn_out", il);
115        }
116        cur = ggml_add(ctx0, cur, ffn_inp);
117
118        cur = build_cvec(cur, il);
119        cb(cur, "l_out", il);
120
121        // input for next layer
122        inpL = cur;
123    }
124    cur = inpL;
125
126    cur = build_norm(cur,
127            model.output_norm, NULL,
128            LLM_NORM_RMS, -1);
129
130    cb(cur, "result_norm", -1);
131    res->t_embd = cur;
132
133    // lm_head
134    cur = build_lora_mm(model.output, cur, model.output_s);
135
136    cb(cur, "result_output", -1);
137    res->t_logits = cur;
138
139    ggml_build_forward_expand(gf, cur);
140}