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 →
5.3 KiB157 linesraw
1#include "models.h"
2
3void llama_model_arcee::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    // Arcee uses the same structure as Llama
7    switch (hparams.n_layer()) {
8        case 36: type = LLM_TYPE_4B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_arcee::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}, TENSOR_NOT_REQUIRED);
21
22    // if output is NULL, init from the input tok embed
23    if (output == NULL) {
24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25    }
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
31
32        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
33        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
34
35        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
36
37        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
38
39        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
40        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
41    }
42}
43
44std::unique_ptr<llm_graph_context> llama_model_arcee::build_arch_graph(const llm_graph_params & params) const {
45    return std::make_unique<graph>(*this, params);
46}
47
48llama_model_arcee::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
49    const int64_t n_embd_head = hparams.n_embd_head_v();
50
51    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
52    GGML_ASSERT(n_embd_head == n_rot);
53
54    ggml_tensor * cur;
55    ggml_tensor * inpL;
56
57    inpL = build_inp_embd(model.tok_embd);
58
59    // inp_pos - contains the positions
60    ggml_tensor * inp_pos = build_inp_pos();
61
62    auto * inp_attn = build_attn_inp_kv();
63
64    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
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        // norm
72        cur = build_norm(inpL,
73                model.layers[il].attn_norm, NULL,
74                LLM_NORM_RMS, il);
75        cb(cur, "attn_norm", il);
76
77        // self-attention
78        {
79            // rope freq factors for llama3; may return nullptr for llama2 and other models
80            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
81
82            // compute Q and K and RoPE them
83            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
84                    n_embd_head, n_head, n_head_kv, il);
85
86            Qcur = ggml_rope_ext(
87                    ctx0, Qcur, inp_pos, rope_factors,
88                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
89                    ext_factor, attn_factor, beta_fast, beta_slow
90                    );
91
92            Kcur = ggml_rope_ext(
93                    ctx0, Kcur, inp_pos, rope_factors,
94                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
95                    ext_factor, attn_factor, beta_fast, beta_slow
96                    );
97
98            cb(Qcur, "Qcur", il);
99            cb(Kcur, "Kcur", il);
100            cb(Vcur, "Vcur", il);
101
102            cur = build_attn(inp_attn,
103                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
104                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
105            cb(cur, "attn_out", il);
106        }
107
108        if (il == n_layer - 1 && inp_out_ids) {
109            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
110            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
111        }
112
113        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
114        cb(ffn_inp, "ffn_inp", il);
115
116        // feed-forward network
117        // ARCEE uses relu^2 instead of silu
118        cur = build_norm(ffn_inp,
119                model.layers[il].ffn_norm, NULL,
120                LLM_NORM_RMS, il);
121        cb(cur, "ffn_norm", il);
122
123        cur = build_ffn(cur,
124                model.layers[il].ffn_up,   NULL, NULL,
125                NULL,                      NULL, NULL,
126                model.layers[il].ffn_down, NULL, NULL,
127                NULL,
128                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
129        cb(cur, "ffn_out", il);
130
131        cur = ggml_add(ctx0, cur, ffn_inp);
132        cb(cur, "ffn_out", il);
133
134        cur = build_cvec(cur, il);
135        cb(cur, "l_out", il);
136
137        // input for next layer
138        inpL = cur;
139    }
140
141    cur = inpL;
142
143    cur = build_norm(cur,
144            model.output_norm, NULL,
145            LLM_NORM_RMS, -1);
146
147    cb(cur, "result_norm", -1);
148    res->t_embd = cur;
149
150    // lm_head
151    cur = build_lora_mm(model.output, cur, model.output_s);
152
153    cb(cur, "result_output", -1);
154    res->t_logits = cur;
155
156    ggml_build_forward_expand(gf, cur);
157}