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 →
6.7 KiB177 linesraw
1#include "models.h"
2
3void llama_model_gemma2::load_arch_hparams(llama_model_loader & ml) {
4    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
5    hparams.n_swa = 4096; // default value of gemma 2
6    uint32_t swa_period = 2;
7    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
8    hparams.set_swa_pattern(swa_period);
9    hparams.attn_soft_cap = true;
10    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
11    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
12
13    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
14    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
15    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
16    ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,      hparams.f_attn_logit_softcapping, false);
17    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
18
19    switch (hparams.n_layer()) {
20        case 26: type = LLM_TYPE_2B; break;
21        case 42: type = LLM_TYPE_9B; break;
22        case 46: type = LLM_TYPE_27B; break;
23        default: type = LLM_TYPE_UNKNOWN;
24   }
25
26    // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
27    hparams.f_attention_scale = type == LLM_TYPE_27B
28        ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
29        : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
30}
31
32void llama_model_gemma2::load_arch_tensors(llama_model_loader &) {
33    LLAMA_LOAD_LOCALS;
34
35    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
36
37    // output
38    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
39    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
40
41    for (int i = 0; i < n_layer; ++i) {
42        auto & layer = layers[i];
43
44        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
45
46        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
47        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
48        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
49
50        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
51        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
52        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
53        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
54        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
55    }
56}
57
58std::unique_ptr<llm_graph_context> llama_model_gemma2::build_arch_graph(const llm_graph_params & params) const {
59    return std::make_unique<graph>(*this, params);
60}
61
62llama_model_gemma2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
63    const int64_t n_embd_head = hparams.n_embd_head_k();
64
65    ggml_tensor * cur;
66    ggml_tensor * inpL;
67
68    inpL = build_inp_embd(model.tok_embd);
69
70    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
71    cb(inpL, "inp_scaled", -1);
72
73    // inp_pos - contains the positions
74    ggml_tensor * inp_pos = build_inp_pos();
75
76    auto * inp_attn = build_attn_inp_kv_iswa();
77
78    ggml_tensor * inp_out_ids = build_inp_out_ids();
79
80    for (int il = 0; il < n_layer; ++il) {
81        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
82        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
83
84        // norm
85        cur = build_norm(inpL,
86                model.layers[il].attn_norm, NULL,
87                LLM_NORM_RMS, il);
88        cb(cur, "attn_norm", il);
89
90        // self-attention
91        {
92            // compute Q and K and RoPE them
93            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
94                    n_embd_head, n_head, n_head_kv, il);
95
96            Qcur = ggml_rope_ext(
97                    ctx0, Qcur, inp_pos, nullptr,
98                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
99                    ext_factor, attn_factor, beta_fast, beta_slow);
100
101            Kcur = ggml_rope_ext(
102                    ctx0, Kcur, inp_pos, nullptr,
103                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
104                    ext_factor, attn_factor, beta_fast, beta_slow);
105
106            cb(Qcur, "Qcur", il);
107            cb(Kcur, "Kcur", il);
108            cb(Vcur, "Vcur", il);
109
110            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
111
112            cur = build_attn(inp_attn,
113                    model.layers[il].wo, NULL, model.layers[il].wo_s,
114                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
115        }
116        if (il == n_layer - 1 && inp_out_ids) {
117            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
118            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
119        }
120        cur = build_norm(cur,
121                model.layers[il].attn_post_norm, NULL,
122                LLM_NORM_RMS, il);
123        cb(cur, "attn_post_norm", il);
124
125        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
126        cb(sa_out, "sa_out", il);
127
128        cur = build_norm(sa_out,
129                model.layers[il].ffn_norm, NULL,
130                LLM_NORM_RMS, il);
131        cb(cur, "ffn_norm", il);
132
133        // feed-forward network
134        {
135            cur = build_ffn(cur,
136                    model.layers[il].ffn_up,   NULL, NULL,
137                    model.layers[il].ffn_gate, NULL, NULL,
138                    model.layers[il].ffn_down, NULL, NULL,
139                    NULL,
140                    LLM_FFN_GELU, LLM_FFN_PAR, il);
141            cb(cur, "ffn_out", il);
142        }
143        cur = build_norm(cur,
144                model.layers[il].ffn_post_norm, NULL,
145                LLM_NORM_RMS, -1);
146        cb(cur, "ffn_post_norm", -1);
147
148        cur = ggml_add(ctx0, cur, sa_out);
149
150        cur = build_cvec(cur, il);
151        cb(cur, "l_out", il);
152
153        // input for next layer
154        inpL = cur;
155    }
156    cur = inpL;
157
158    cur = build_norm(cur,
159            model.output_norm, NULL,
160            LLM_NORM_RMS, -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    // final logit soft-capping
169    cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
170    cur = ggml_tanh(ctx0, cur);
171    cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
172
173    cb(cur, "result_output", -1);
174    res->t_logits = cur;
175
176    ggml_build_forward_expand(gf, cur);
177}