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 →
8.6 KiB225 linesraw
1#include "models.h"
2
3void llama_model_gemma3::load_arch_hparams(llama_model_loader & ml) {
4    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
5    if (found_swa && hparams.n_swa > 0) {
6        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
7        uint32_t swa_period = 6;
8        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
9        hparams.set_swa_pattern(swa_period);
10
11        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
12    } else {
13        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
14    }
15
16    hparams.f_final_logit_softcapping = 0.0f;
17    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
18    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
19
20    switch (hparams.n_layer()) {
21        case 18: type = LLM_TYPE_270M; break;
22        case 26: type = LLM_TYPE_1B; break;
23        case 32: type = LLM_TYPE_8B; break; // Rnj-1
24        case 34: type = LLM_TYPE_4B; break;
25        case 48: type = LLM_TYPE_12B; break;
26        case 62: type = LLM_TYPE_27B; break;
27        default: type = LLM_TYPE_UNKNOWN;
28    }
29
30    // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
31    hparams.f_attention_scale = type == LLM_TYPE_27B
32        ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
33        : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
34}
35
36void llama_model_gemma3::load_arch_tensors(llama_model_loader &) {
37    LLAMA_LOAD_LOCALS;
38
39    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
40
41    // output
42    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
43    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
44
45    // if output is NULL, init from the input tok embed
46    if (output == NULL) {
47        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
48    }
49
50    // Dense linear weights
51    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);
52    dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);
53
54
55    for (int i = 0; i < n_layer; ++i) {
56        auto & layer = layers[i];
57
58        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
59
60        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
61        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
62
63        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
64        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
65        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
66
67        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
68        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
69        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
70        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
71        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
72    }
73}
74
75std::unique_ptr<llm_graph_context> llama_model_gemma3::build_arch_graph(const llm_graph_params & params) const {
76    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
77        return std::make_unique<graph<true>>(*this, params);
78    } else {
79        return std::make_unique<graph<false>>(*this, params);
80    }
81}
82
83template <bool iswa>
84llama_model_gemma3::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
85    const int64_t n_embd_head = hparams.n_embd_head_k();
86
87    ggml_tensor * cur;
88    ggml_tensor * inpL;
89
90    inpL = build_inp_embd(model.tok_embd);
91
92    // important: do not normalize weights for raw embeddings input (i.e. encoded image embeddings)
93    inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
94    cb(inpL, "inp_scaled", -1);
95
96    // inp_pos - contains the positions
97    ggml_tensor * inp_pos = build_inp_pos();
98
99    // TODO: is causal == true correct? might need some changes
100    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
101    inp_attn_type * inp_attn = nullptr;
102
103    if constexpr (iswa) {
104        inp_attn = build_attn_inp_kv_iswa();
105    } else {
106        inp_attn = build_attn_inp_kv();
107    }
108
109    ggml_tensor * inp_out_ids = build_inp_out_ids();
110
111    for (int il = 0; il < n_layer; ++il) {
112        float freq_base_l  = 0.0f;
113        float freq_scale_l = 0.0f;
114
115        if constexpr (iswa) {
116            freq_base_l  = model.get_rope_freq_base (cparams, il);
117            freq_scale_l = model.get_rope_freq_scale(cparams, il);
118        } else {
119            freq_base_l  = freq_base;
120            freq_scale_l = freq_scale;
121        }
122
123        // norm
124        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
125        cb(cur, "attn_norm", il);
126
127        // self-attention
128        {
129            // compute Q and K and RoPE them
130            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
131                    n_embd_head, n_head, n_head_kv, il);
132
133            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
134            cb(Qcur, "Qcur_normed", il);
135
136            Qcur = ggml_rope_ext(
137                    ctx0, Qcur, inp_pos, nullptr,
138                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
139                    ext_factor, attn_factor, beta_fast, beta_slow);
140
141            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
142            cb(Kcur, "Kcur_normed", il);
143
144            Kcur = ggml_rope_ext(
145                    ctx0, Kcur, inp_pos, nullptr,
146                    n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
147                    ext_factor, attn_factor, beta_fast, beta_slow);
148
149            cb(Qcur, "Qcur", il);
150            cb(Kcur, "Kcur", il);
151            cb(Vcur, "Vcur", il);
152
153            // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
154            Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
155
156            cur = build_attn(inp_attn,
157                    model.layers[il].wo, NULL, model.layers[il].wo_s,
158                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
159        }
160        if (il == n_layer - 1 && inp_out_ids) {
161            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
162            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
163        }
164        cur = build_norm(cur,
165                model.layers[il].attn_post_norm, NULL,
166                LLM_NORM_RMS, il);
167        cb(cur, "attn_post_norm", il);
168
169        ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
170        cb(sa_out, "sa_out", il);
171
172        cur = build_norm(sa_out,
173                model.layers[il].ffn_norm, NULL,
174                LLM_NORM_RMS, il);
175        cb(cur, "ffn_norm", il);
176
177        // feed-forward network
178        {
179            cur = build_ffn(cur,
180                    model.layers[il].ffn_up,   NULL, NULL,
181                    model.layers[il].ffn_gate, NULL, NULL,
182                    model.layers[il].ffn_down, NULL, NULL,
183                    NULL,
184                    LLM_FFN_GELU, LLM_FFN_PAR, il);
185            cb(cur, "ffn_out", il);
186        }
187        cur = build_norm(cur,
188                model.layers[il].ffn_post_norm, NULL,
189                LLM_NORM_RMS, -1);
190        cb(cur, "ffn_post_norm", il);
191
192        cur = ggml_add(ctx0, cur, sa_out);
193
194        cur = build_cvec(cur, il);
195        cb(cur, "l_out", il);
196
197        // input for next layer
198        inpL = cur;
199    }
200    cur = inpL;
201
202    cur = build_norm(cur,
203            model.output_norm, NULL,
204            LLM_NORM_RMS, -1);
205
206    cb(cur, "result_norm", -1);
207    res->t_embd = cur;
208
209    // lm_head
210    cur = build_lora_mm(model.output, cur, model.output_s);
211
212    if (hparams.f_final_logit_softcapping) {
213        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
214        cur = ggml_tanh(ctx0, cur);
215        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
216    }
217
218    cb(cur, "result_output", -1);
219    res->t_logits = cur;
220
221    ggml_build_forward_expand(gf, cur);
222}
223
224template struct llama_model_gemma3::graph<false>;
225template struct llama_model_gemma3::graph<true>;