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 →
9.6 KiB235 linesraw
1#include "models.h"
2
3void llama_model_mistral3::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
6
7    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast,    false);
8    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
9    ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, false);
10
11    hparams.f_attn_temp_offset = 0.0f;
12
13    // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
14    if (hparams.f_attn_temp_scale != 0.0f) {
15        hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
16        if (hparams.n_attn_temp_floor_scale == 0) {
17            throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
18        }
19    }
20
21    switch (hparams.n_layer()) {
22        case 26: type = LLM_TYPE_3B; break;
23        case 34: type = LLM_TYPE_8B; break;
24        case 40: type = LLM_TYPE_14B; break;
25        default: type = LLM_TYPE_UNKNOWN;
26    }
27}
28
29void llama_model_mistral3::load_arch_tensors(llama_model_loader &) {
30    LLAMA_LOAD_LOCALS;
31
32    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
33
34    // output
35    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
36    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
37
38    // if output is NULL, init from the input tok embed
39    if (output == NULL) {
40        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
41    }
42
43    for (int i = 0; i < n_layer; ++i) {
44        auto & layer = layers[i];
45
46        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
47
48        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
49        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
50
51        // optional bias tensors
52        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
53
54        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
55
56        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
57            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
58            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
59        }
60        else {
61            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
62        }
63
64        if (n_expert == 0) {
65            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
66            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
67            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
68
69            // optional MLP bias
70            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
71            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
72            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
73        } else {
74            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
75            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
76            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
77            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
78
79            // For Granite MoE Shared
80            if (hparams.n_ff_shexp > 0) {
81                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
82                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
83                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
84            }
85        }
86    }
87}
88
89std::unique_ptr<llm_graph_context> llama_model_mistral3::build_arch_graph(const llm_graph_params & params) const {
90    return std::make_unique<graph>(*this, params);
91}
92
93llama_model_mistral3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
94    const int64_t n_embd_head = hparams.n_embd_head_v();
95
96    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
97    GGML_ASSERT(n_embd_head == n_rot);
98
99    ggml_tensor * cur;
100    ggml_tensor * inpL;
101
102    inpL = build_inp_embd(model.tok_embd);
103
104    // inp_pos - contains the positions
105    ggml_tensor * inp_pos = build_inp_pos();
106
107    // (optional) temperature tuning
108    ggml_tensor * inp_attn_scale = nullptr;
109    if (hparams.f_attn_temp_scale != 0.0f) {
110        inp_attn_scale = build_inp_attn_scale();
111    }
112
113    auto * inp_attn = build_attn_inp_kv();
114
115    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
116
117    ggml_tensor * inp_out_ids = build_inp_out_ids();
118
119    for (int il = 0; il < n_layer; ++il) {
120        ggml_tensor * inpSA = inpL;
121
122        // norm
123        cur = build_norm(inpL,
124                model.layers[il].attn_norm, NULL,
125                LLM_NORM_RMS, il);
126        cb(cur, "attn_norm", il);
127
128        // self-attention
129        {
130            // rope freq factors for llama3; may return nullptr for llama2 and other models
131            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
132
133            // compute Q and K and RoPE them
134            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
135                    n_embd_head, n_head, n_head_kv, il);
136
137            Qcur = ggml_rope_ext(
138                    ctx0, Qcur, inp_pos, rope_factors,
139                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
140                    ext_factor, attn_factor, beta_fast, beta_slow
141                    );
142
143            Kcur = ggml_rope_ext(
144                    ctx0, Kcur, inp_pos, rope_factors,
145                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
146                    ext_factor, attn_factor, beta_fast, beta_slow
147                    );
148
149            cb(Qcur, "Qcur", il);
150            cb(Kcur, "Kcur", il);
151            cb(Vcur, "Vcur", il);
152
153            if (inp_attn_scale) {
154                // apply llama 4 temperature scaling
155                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
156                cb(Qcur, "Qcur_attn_temp_scaled", il);
157            }
158
159            cur = build_attn(inp_attn,
160                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
161                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
162            cb(cur, "attn_out", il);
163        }
164        if (il == n_layer - 1 && inp_out_ids) {
165            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
166            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
167        }
168        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
169        cb(ffn_inp, "ffn_inp", il);
170
171        // feed-forward network (non-MoE)
172        if (model.layers[il].ffn_gate_inp == nullptr) {
173
174            cur = build_norm(ffn_inp,
175                    model.layers[il].ffn_norm, NULL,
176                    LLM_NORM_RMS, il);
177            cb(cur, "ffn_norm", il);
178
179            cur = build_ffn(cur,
180                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,
181                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
182                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
183                    NULL,
184                    LLM_FFN_SILU, LLM_FFN_PAR, il);
185            cb(cur, "ffn_out", il);
186        } else {
187            // MoE branch
188            cur = build_norm(ffn_inp,
189                    model.layers[il].ffn_norm, NULL,
190                    LLM_NORM_RMS, il);
191            cb(cur, "ffn_norm", il);
192
193            cur = build_moe_ffn(cur,
194                    model.layers[il].ffn_gate_inp,
195                    model.layers[il].ffn_up_exps,
196                    model.layers[il].ffn_gate_exps,
197                    model.layers[il].ffn_down_exps,
198                    nullptr,
199                    n_expert, n_expert_used,
200                    LLM_FFN_SILU, true,
201                    hparams.expert_weights_scale,
202                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
203                    il,
204                    nullptr, nullptr,
205                    model.layers[il].ffn_up_exps_s,
206                    model.layers[il].ffn_gate_exps_s,
207                    model.layers[il].ffn_down_exps_s);
208            cb(cur, "ffn_moe_out", il);
209        }
210        cur = ggml_add(ctx0, cur, ffn_inp);
211        cb(cur, "ffn_out", il);
212
213        cur = build_cvec(cur, il);
214        cb(cur, "l_out", il);
215
216        // input for next layer
217        inpL = cur;
218    }
219    cur = inpL;
220
221    cur = build_norm(cur,
222            model.output_norm, NULL,
223            LLM_NORM_RMS, -1);
224
225    cb(cur, "result_norm", -1);
226    res->t_embd = cur;
227
228    // lm_head
229    cur = build_lora_mm(model.output, cur, model.output_s);
230
231    cb(cur, "result_output", -1);
232    res->t_logits = cur;
233
234    ggml_build_forward_expand(gf, cur);
235}