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 →
5.2 KiB151 linesraw
1#include "models.h"
2
3void llama_model_seed_oss::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 64: type = LLM_TYPE_36B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_seed_oss::load_arch_tensors(llama_model_loader &) {
13    LLAMA_LOAD_LOCALS;
14
15    const uint32_t head_dim             = hparams.n_embd_head_k();
16    const int64_t n_qo_dim              = n_head * head_dim;
17    const int64_t n_kv_dim              = n_head_kv * head_dim;
18
19    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
20
21    // output
22    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
23    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
24    // if output is NULL, init from the input tok embed
25    if (output == NULL) {
26        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
27    }
28
29    for (int i = 0; i < n_layer; ++i) {
30        auto & layer = layers[i];
31
32        create_tensor_qkv(layer, i, n_embd, n_qo_dim, n_kv_dim, n_kv_dim, 0);
33        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, 0);
34
35
36        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
37        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
38
39        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
40        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
41        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
42    }
43}
44
45std::unique_ptr<llm_graph_context> llama_model_seed_oss::build_arch_graph(const llm_graph_params & params) const {
46    return std::make_unique<graph>(*this, params);
47}
48
49llama_model_seed_oss::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
50    const int64_t n_embd_head = hparams.n_embd_head_v();
51
52    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
53    GGML_ASSERT(n_embd_head == n_rot);
54
55    ggml_tensor * cur;
56    ggml_tensor * inpL;
57
58    inpL = build_inp_embd(model.tok_embd);
59
60    // inp_pos - contains the positions
61    ggml_tensor * inp_pos = build_inp_pos();
62
63    auto * inp_attn = build_attn_inp_kv();
64
65    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
66
67    ggml_tensor * inp_out_ids = build_inp_out_ids();
68
69    for (int il = 0; il < n_layer; ++il) {
70        ggml_tensor * inpSA = inpL;
71
72        // norm
73        cur = build_norm(inpL,
74                model.layers[il].attn_norm, NULL,
75                LLM_NORM_RMS, il);
76        cb(cur, "attn_norm", il);
77
78        // self-attention
79        {
80            // compute Q and K and RoPE them
81            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
82                    n_embd_head, n_head, n_head_kv, il);
83
84            Qcur = ggml_rope_ext(
85                    ctx0, Qcur, inp_pos, nullptr,
86                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
87                    ext_factor, attn_factor, beta_fast, beta_slow
88                    );
89
90            Kcur = ggml_rope_ext(
91                    ctx0, Kcur, inp_pos, nullptr,
92                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
93                    ext_factor, attn_factor, beta_fast, beta_slow
94                    );
95
96            cb(Qcur, "Qcur", il);
97            cb(Kcur, "Kcur", il);
98            cb(Vcur, "Vcur", il);
99
100            cur = build_attn(inp_attn,
101                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
102                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
103            cb(cur, "attn_out", il);
104        }
105        if (il == n_layer - 1 && inp_out_ids) {
106            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
107            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
108        }
109        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
110        cb(ffn_inp, "ffn_inp", il);
111
112        // feed-forward network
113        cur = build_norm(ffn_inp,
114                model.layers[il].attn_post_norm, NULL,
115                LLM_NORM_RMS, il);
116        cb(cur, "attn_post_norm", il);
117
118        cur = build_ffn(cur,
119                model.layers[il].ffn_up,   NULL, NULL,
120                model.layers[il].ffn_gate, NULL, NULL,
121                model.layers[il].ffn_down, NULL, NULL,
122                NULL,
123                LLM_FFN_SILU, LLM_FFN_PAR, il);
124        cb(cur, "ffn_out", il);
125
126        cur = ggml_add(ctx0, cur, ffn_inp);
127        cb(cur, "ffn_out", il);
128
129        cur = build_cvec(cur, il);
130        cb(cur, "l_out", il);
131
132        // input for next layer
133        inpL = cur;
134    }
135    cur = inpL;
136
137    cur = build_norm(cur,
138            model.output_norm, NULL,
139            LLM_NORM_RMS, -1);
140
141    cb(cur, "result_norm", -1);
142    res->t_embd = cur;
143
144    // lm_head
145    cur = build_lora_mm(model.output, cur, model.output_s);
146
147    cb(cur, "result_output", -1);
148    res->t_logits = cur;
149
150    ggml_build_forward_expand(gf, cur);
151}