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 →
6.2 KiB172 linesraw
1#include "models.h"
2
3void llama_model_stablelm::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5
6    switch (hparams.n_layer()) {
7        case 24: type = LLM_TYPE_1B; break;
8        case 32: type = LLM_TYPE_3B; break;
9        case 40: type = LLM_TYPE_12B; break;
10        default: type = LLM_TYPE_UNKNOWN;
11   }
12}
13
14void llama_model_stablelm::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_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
21    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
22    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
23
24    for (int i = 0; i < n_layer; ++i) {
25        auto & layer = layers[i];
26
27        layer.attn_norm =   create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
28        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
29
30        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
31        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
32
33        // optional q and k layernorms, present in StableLM 2 12B
34        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head},    TENSOR_NOT_REQUIRED);
35        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, TENSOR_NOT_REQUIRED);
36
37        // optional FFN norm, not present in StableLM 2 12B which uses parallel residual
38        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
39        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
40
41        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
42        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
43        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
44    }
45}
46
47std::unique_ptr<llm_graph_context> llama_model_stablelm::build_arch_graph(const llm_graph_params & params) const {
48    return std::make_unique<graph>(*this, params);
49}
50
51llama_model_stablelm::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        // norm
70        cur = build_norm(inpL,
71                model.layers[il].attn_norm,
72                model.layers[il].attn_norm_b,
73                LLM_NORM, il);
74        cb(cur, "attn_norm", il);
75
76        ggml_tensor * inpSA = cur;
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            if (model.layers[il].attn_q_norm) {
85                Qcur = build_norm(Qcur,
86                        model.layers[il].attn_q_norm,
87                        NULL,
88                        LLM_NORM, il);
89                cb(Qcur, "Qcur", il);
90            }
91            if (model.layers[il].attn_k_norm) {
92                Kcur = build_norm(Kcur,
93                        model.layers[il].attn_k_norm,
94                        NULL,
95                        LLM_NORM, il);
96                cb(Kcur, "Kcur", il);
97            }
98
99            Qcur = ggml_rope_ext(
100                    ctx0, Qcur, inp_pos, nullptr,
101                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
102                    ext_factor, attn_factor, beta_fast, beta_slow
103                    );
104
105            Kcur = ggml_rope_ext(
106                    ctx0, Kcur, inp_pos, nullptr,
107                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
108                    ext_factor, attn_factor, beta_fast, beta_slow
109                    );
110
111            cb(Qcur, "Qcur", il);
112            cb(Kcur, "Kcur", il);
113            cb(Vcur, "Vcur", il);
114
115            cur = build_attn(inp_attn,
116                    model.layers[il].wo, NULL, model.layers[il].wo_s,
117                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
118        }
119        if (il == n_layer - 1 && inp_out_ids) {
120            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
121            inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
122            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
123        }
124        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
125        cb(ffn_inp, "ffn_inp", il);
126
127        // feed-forward network
128        {
129            if (model.layers[il].ffn_norm) {
130                cur = build_norm(ffn_inp,
131                        model.layers[il].ffn_norm,
132                        model.layers[il].ffn_norm_b,
133                        LLM_NORM, il);
134                cb(cur, "ffn_norm", il);
135            } else {
136                // parallel residual
137                cur = inpSA;
138            }
139            cur = build_ffn(cur,
140                    model.layers[il].ffn_up,   NULL, NULL,
141                    model.layers[il].ffn_gate, NULL, NULL,
142                    model.layers[il].ffn_down, NULL, NULL,
143                    NULL,
144                    LLM_FFN_SILU, LLM_FFN_PAR, il);
145            cb(cur, "ffn_out", il);
146        }
147        cur = ggml_add(ctx0, cur, ffn_inp);
148
149        cur = build_cvec(cur, il);
150        cb(cur, "l_out", il);
151
152        // input for next layer
153        inpL = cur;
154    }
155    cur = inpL;
156
157    cur = build_norm(cur,
158            model.output_norm,
159            model.output_norm_b,
160            LLM_NORM, -1);
161
162    cb(cur, "result_norm", -1);
163    res->t_embd = cur;
164
165    // lm_head
166    cur = build_lora_mm(model.output, cur, model.output_s);
167
168    cb(cur, "result_output", -1);
169    res->t_logits = cur;
170
171    ggml_build_forward_expand(gf, cur);
172}