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 KiB150 linesraw
1#include "models.h"
2
3void llama_model_nemotron::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5
6    switch (hparams.n_layer()) {
7        case 32: type = LLM_TYPE_4B; break;
8        default: type = LLM_TYPE_UNKNOWN;
9    }
10}
11
12void llama_model_nemotron::load_arch_tensors(llama_model_loader &) {
13    LLAMA_LOAD_LOCALS;
14
15    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
16
17    // output
18    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
19    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0);
20    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
21
22    for (int i = 0; i < n_layer; ++i) {
23        auto & layer = layers[i];
24
25        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
26        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
27
28        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
29        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
30
31        // optional bias tensors
32        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
33
34        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
35        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
36
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        // optional MLP bias
41        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
42        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
43    }
44}
45
46std::unique_ptr<llm_graph_context> llama_model_nemotron::build_arch_graph(const llm_graph_params & params) const {
47    return std::make_unique<graph>(*this, params);
48}
49
50llama_model_nemotron::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
51    const int64_t n_embd_head = hparams.n_embd_head_v();
52
53    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
54    //GGML_ASSERT(n_embd_head == n_rot);
55
56    ggml_tensor * cur;
57    ggml_tensor * inpL;
58
59    inpL = build_inp_embd(model.tok_embd);
60
61    // inp_pos - contains the positions
62    ggml_tensor * inp_pos = build_inp_pos();
63
64    auto * inp_attn = build_attn_inp_kv();
65
66    ggml_tensor * inp_out_ids = build_inp_out_ids();
67
68    for (int il = 0; il < n_layer; ++il) {
69        ggml_tensor * inpSA = inpL;
70
71        // norm
72        cur = build_norm(inpL,
73                model.layers[il].attn_norm,
74                model.layers[il].attn_norm_b,
75                LLM_NORM, il);
76        cb(cur, "attn_norm", il);
77
78        // self-attention
79        {
80            // compute Q and K and RoPE them
81            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
82                    n_embd_head, n_head, n_head_kv, il);
83
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, 1.0f/sqrtf(float(n_embd_head)), il);
103        }
104        if (il == n_layer - 1 && inp_out_ids) {
105            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
106            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
107        }
108        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
109        cb(ffn_inp, "ffn_inp", il);
110
111        // feed-forward network
112        cur = build_norm(ffn_inp,
113                model.layers[il].ffn_norm,
114                model.layers[il].ffn_norm_b,
115                LLM_NORM, il);
116        cb(cur, "ffn_norm", il);
117
118        cur = build_ffn(cur,
119                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
120                NULL,                      NULL,                        NULL,
121                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
122                NULL,
123                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
124
125        cur = ggml_add(ctx0, cur, ffn_inp);
126        cb(cur, "ffn_out", il);
127
128        cur = build_cvec(cur, il);
129        cb(cur, "l_out", il);
130
131        // input for next layer
132        inpL = cur;
133    }
134    cur = inpL;
135
136    cur = build_norm(cur,
137            model.output_norm, model.output_norm_b,
138            LLM_NORM, -1);
139
140    cb(cur, "result_norm", -1);
141    res->t_embd = cur;
142
143    // lm_head
144    cur = build_lora_mm(model.output, cur, model.output_s);
145
146    cb(cur, "result_output", -1);
147    res->t_logits = cur;
148
149    ggml_build_forward_expand(gf, cur);
150}