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 →
5.3 KiB152 linesraw
1#include "models.h"
2
3void llama_model_smollm3::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    hparams.n_no_rope_layer_step = 4;
6
7    switch (hparams.n_layer()) {
8        case 36: type = LLM_TYPE_3B; break;
9        default: type = LLM_TYPE_UNKNOWN;
10    }
11}
12
13void llama_model_smollm3::load_arch_tensors(llama_model_loader &) {
14    LLAMA_LOAD_LOCALS;
15
16    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
17
18    // output
19    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
20    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
21
22    // if output is NULL, init from the input tok embed
23    if (output == NULL) {
24        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
25    }
26
27    for (int i = 0; i < n_layer; ++i) {
28        auto & layer = layers[i];
29
30        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
31
32        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
33        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
34
35        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
36        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
37        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
38        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
39    }
40}
41
42std::unique_ptr<llm_graph_context> llama_model_smollm3::build_arch_graph(const llm_graph_params & params) const {
43    return std::make_unique<graph>(*this, params);
44}
45
46llama_model_smollm3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
47    const int64_t n_embd_head = hparams.n_embd_head_v();
48
49    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
50    GGML_ASSERT(n_embd_head == n_rot);
51
52    ggml_tensor * cur;
53    ggml_tensor * inpL;
54
55    inpL = build_inp_embd(model.tok_embd);
56
57    // inp_pos - contains the positions
58    ggml_tensor * inp_pos = build_inp_pos();
59
60    auto * inp_attn = build_attn_inp_kv();
61
62    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
63
64    ggml_tensor * inp_out_ids = build_inp_out_ids();
65
66    for (int il = 0; il < n_layer; ++il) {
67        ggml_tensor * inpSA = inpL;
68
69        const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
70
71        // norm
72        cur = build_norm(inpL,
73                model.layers[il].attn_norm, NULL,
74                LLM_NORM_RMS, il);
75        cb(cur, "attn_norm", il);
76
77        // self-attention
78        {
79            // compute Q and K and RoPE them
80            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
81                    n_embd_head, n_head, n_head_kv, il);
82
83            if (use_rope) {
84                Qcur = ggml_rope_ext(
85                        ctx0, Qcur, inp_pos, nullptr,
86                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
87                        ext_factor, attn_factor, beta_fast, beta_slow
88                        );
89
90                Kcur = ggml_rope_ext(
91                        ctx0, Kcur, inp_pos, nullptr,
92                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
93                        ext_factor, attn_factor, beta_fast, beta_slow
94                        );
95            }
96            cb(Qcur, "Qcur", il);
97            cb(Kcur, "Kcur", il);
98            cb(Vcur, "Vcur", il);
99
100            cur = build_attn(inp_attn,
101                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
102                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
103            cb(cur, "attn_out", il);
104        }
105        if (il == n_layer - 1 && inp_out_ids) {
106            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
107            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
108        }
109        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
110        cb(ffn_inp, "ffn_inp", il);
111
112        // feed-forward network
113        {
114            cur = build_norm(ffn_inp,
115                    model.layers[il].ffn_norm, NULL,
116                    LLM_NORM_RMS, il);
117            cb(cur, "ffn_norm", il);
118
119            cur = build_ffn(cur,
120                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
121                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
122                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
123                    NULL,
124                    LLM_FFN_SILU, LLM_FFN_PAR, il);
125            cb(cur, "ffn_out", il);
126        }
127        cur = ggml_add(ctx0, cur, ffn_inp);
128        cb(cur, "ffn_out", il);
129
130        cur = build_cvec(cur, il);
131        cb(cur, "l_out", il);
132
133        // input for next layer
134        inpL = cur;
135    }
136    cur = inpL;
137
138    cur = build_norm(cur,
139            model.output_norm, NULL,
140            LLM_NORM_RMS, -1);
141
142    cb(cur, "result_norm", -1);
143    res->t_embd = cur;
144
145    // lm_head
146    cur = build_lora_mm(model.output, cur, model.output_s);
147
148    cb(cur, "result_output", -1);
149    res->t_logits = cur;
150
151    ggml_build_forward_expand(gf, cur);
152}