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 →
6.8 KiB170 linesraw
1#include "models.h"
2
3void llama_model_apertus::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5
6    ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_N, hparams.xielu_alpha_n, hparams.n_layer());
7    ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_P, hparams.xielu_alpha_p, hparams.n_layer());
8    ml.get_key_or_arr(LLM_KV_XIELU_BETA,    hparams.xielu_beta,    hparams.n_layer());
9    ml.get_key_or_arr(LLM_KV_XIELU_EPS,     hparams.xielu_eps,     hparams.n_layer());
10
11    switch (hparams.n_layer()) {
12        case 32: type = LLM_TYPE_8B; break;
13        default: type = LLM_TYPE_UNKNOWN;
14    }
15}
16
17void llama_model_apertus::load_arch_tensors(llama_model_loader &) {
18    LLAMA_LOAD_LOCALS;
19
20    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
21
22    // output
23    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
24    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), { n_embd, n_vocab }, 0);
25
26    for (int i = 0; i < n_layer; ++i) {
27        auto & layer = layers[i];
28
29        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
30
31        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
32            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
33            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
34        } else {
35            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
36        }
37
38        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
39        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
40
41        // optional bias tensors
42        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
43
44        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
45        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
46        layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
47
48        // Q and K layernorms for Apertus
49        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
50        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
51        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
52        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
53    }
54}
55
56std::unique_ptr<llm_graph_context> llama_model_apertus::build_arch_graph(const llm_graph_params & params) const {
57    return std::make_unique<graph>(*this, params);
58}
59
60llama_model_apertus::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
61    const int64_t n_embd_head = hparams.n_embd_head_v();
62
63    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
64    GGML_ASSERT(n_embd_head == n_rot);
65
66    ggml_tensor * cur;
67    ggml_tensor * inpL;
68
69    inpL = build_inp_embd(model.tok_embd);
70
71    ggml_tensor * inp_pos  = build_inp_pos();
72    auto *        inp_attn = build_attn_inp_kv();
73
74    const float kq_scale =
75        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
76
77    ggml_tensor * inp_out_ids = build_inp_out_ids();
78
79    for (int il = 0; il < n_layer; ++il) {
80        ggml_tensor * inpSA = inpL;
81
82        cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
83        cb(cur, "attn_norm", il);
84
85        // self-attention
86        {
87            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
88
89            // compute Q and K and RoPE them
90            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
91                    n_embd_head, n_head, n_head_kv, il);
92
93            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
94            cb(Qcur, "Qcur_normed", il);
95
96            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
97            cb(Kcur, "Kcur_normed", il);
98
99            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
100                                 ext_factor, attn_factor, beta_fast, beta_slow);
101
102            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
103                                 ext_factor, attn_factor, beta_fast, beta_slow);
104
105            cb(Qcur, "Qcur_pos", il);
106            cb(Kcur, "Kcur_pos", il);
107            cb(Vcur, "Vcur_pos", il);
108
109            cur = build_attn(inp_attn,
110                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
111                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
112            cb(cur, "attn_out", il);
113        }
114
115        if (il == n_layer - 1 && inp_out_ids) {
116            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
117            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
118        }
119
120        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
121        cb(ffn_inp, "ffn_inp", il);
122
123        // feed-forward network with xIELU activation
124        {
125            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
126            cb(cur, "ffn_norm", il);
127
128            // Up projection
129            ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur);
130            cb(up, "ffn_up", il);
131
132            float alpha_n_val = hparams.xielu_alpha_n[il];
133            float alpha_p_val = hparams.xielu_alpha_p[il];
134            float beta_val    = hparams.xielu_beta[il];
135            float eps_val     = hparams.xielu_eps[il];
136
137            // Apply xIELU activation
138            ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val);
139            cb(activated, "ffn_xielu", il);
140
141            // Down projection
142            cur = build_lora_mm(model.layers[il].ffn_down, activated);
143            cb(cur, "ffn_down", il);
144        }
145
146        cur = ggml_add(ctx0, cur, ffn_inp);
147        cb(cur, "ffn_out", il);
148
149        cur = build_cvec(cur, il);
150        cb(cur, "l_out", il);
151
152        // input for next layer
153        inpL = cur;
154    }
155
156    cur = inpL;
157
158    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
159
160    cb(cur, "result_norm", -1);
161    res->t_embd = cur;
162
163    // lm_head
164    cur = build_lora_mm(model.output, cur, model.output_s);
165
166    cb(cur, "result_output", -1);
167    res->t_logits = cur;
168
169    ggml_build_forward_expand(gf, cur);
170}