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 →
7.8 KiB219 linesraw
1#include "models.h"
2
3void llama_model_gptneox::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
5    ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL,   hparams.use_par_res);
6
7    switch (hparams.n_layer()) {
8        case 6:
9            switch (hparams.n_ff()) {
10                case 512:  type = LLM_TYPE_14M; break;
11                case 2048: type = LLM_TYPE_70M; break;
12                default:   type = LLM_TYPE_UNKNOWN;
13            } break;
14        case 12:
15            switch (hparams.n_ff()) {
16                case 3072: type = LLM_TYPE_160M; break;
17                default: type = LLM_TYPE_UNKNOWN;
18            } break;
19        case 16:
20            switch (hparams.n_ff()) {
21                case 8192: type = LLM_TYPE_1B; break;
22                default: type = LLM_TYPE_UNKNOWN;
23            } break;
24        case 24:
25            switch (hparams.n_ff()) {
26                case 4096: type = LLM_TYPE_410M; break;
27                case 8192: type = LLM_TYPE_1_4B; break;
28                default: type = LLM_TYPE_UNKNOWN;
29            } break;
30        case 32:
31            switch (hparams.n_ff()) {
32                case 10240: type = LLM_TYPE_2_8B; break;
33                case 16384: type = LLM_TYPE_6_9B; break;
34                default: type = LLM_TYPE_UNKNOWN;
35            } break;
36        case 36:
37            switch (hparams.n_ff()) {
38                case 20480: type = LLM_TYPE_12B; break;
39                default: type = LLM_TYPE_UNKNOWN;
40            } break;
41        case 44:
42            switch (hparams.n_ff()) {
43                case 24576: type = LLM_TYPE_20B; break;
44                default: type = LLM_TYPE_UNKNOWN;
45            } break;
46        default: type = LLM_TYPE_UNKNOWN;
47    }
48}
49
50void llama_model_gptneox::load_arch_tensors(llama_model_loader &) {
51    LLAMA_LOAD_LOCALS;
52
53    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
54
55    // output
56    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
57    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
58    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
59
60    for (int i = 0; i < n_layer; ++i) {
61        auto & layer = layers[i];
62
63        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
64        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
65
66        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
67        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
68
69        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
70        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
71
72        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
73        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
74
75        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
76        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
77
78        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
79        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
80    }
81}
82
83std::unique_ptr<llm_graph_context> llama_model_gptneox::build_arch_graph(const llm_graph_params & params) const {
84    return std::make_unique<graph>(*this, params);
85}
86
87llama_model_gptneox::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
88    const int64_t n_embd_head = hparams.n_embd_head_v();
89
90    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
91
92    ggml_tensor * cur;
93    ggml_tensor * inpL;
94
95    inpL = build_inp_embd(model.tok_embd);
96
97    // inp_pos - contains the positions
98    ggml_tensor * inp_pos = build_inp_pos();
99
100    auto * inp_attn = build_attn_inp_kv();
101
102    ggml_tensor * inp_out_ids = build_inp_out_ids();
103
104    for (int il = 0; il < n_layer; ++il) {
105        cur = build_norm(inpL,
106                model.layers[il].attn_norm,
107                model.layers[il].attn_norm_b,
108                LLM_NORM, il);
109        cb(cur, "attn_norm", il);
110
111        // self-attention
112        {
113            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
114                    n_embd_head, n_head, n_head_kv, il);
115
116            Qcur = ggml_rope_ext(
117                    ctx0, Qcur, inp_pos, nullptr,
118                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
119                    ext_factor, attn_factor, beta_fast, beta_slow
120                    );
121
122            Kcur = ggml_rope_ext(
123                    ctx0, Kcur, inp_pos, nullptr,
124                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
125                    ext_factor, attn_factor, beta_fast, beta_slow
126                    );
127
128            cb(Qcur, "Qcur", il);
129            cb(Kcur, "Kcur", il);
130            cb(Vcur, "Vcur", il);
131
132            cur = build_attn(inp_attn,
133                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
134                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
135        }
136
137        if (il == n_layer - 1 && inp_out_ids) {
138            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
139            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
140        }
141
142        // ffn
143        if (hparams.use_par_res) {
144            // attention and ffn are computed in parallel
145            // x = x + attn(ln1(x)) + ffn(ln2(x))
146
147            ggml_tensor * attn_out = cur;
148
149            cur = build_norm(inpL,
150                    model.layers[il].ffn_norm,
151                    model.layers[il].ffn_norm_b,
152                    LLM_NORM, il);
153            cb(cur, "ffn_norm", il);
154
155            cur = build_ffn(cur,
156                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
157                    NULL,                      NULL,                        NULL,
158                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
159                    NULL,
160                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
161            cb(cur, "ffn_out", il);
162
163            cur = ggml_add(ctx0, cur, inpL);
164            cb(cur, "ffn_out", il);
165
166            cur = ggml_add(ctx0, cur, attn_out);
167
168            cur = build_cvec(cur, il);
169            cb(cur, "l_out", il);
170
171            // input for next layer
172            inpL = cur;
173        } else {
174            // attention and ffn are computed sequentially
175            // x = x + attn(ln1(x))
176            // x = x + ffn(ln2(x))
177
178            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
179            cb(ffn_inp, "ffn_inp", il);
180
181            cur = build_norm(ffn_inp,
182                    model.layers[il].ffn_norm,
183                    model.layers[il].ffn_norm_b,
184                    LLM_NORM, il);
185            cb(cur, "ffn_norm", il);
186
187            cur = build_ffn(cur,
188                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
189                    NULL,                      NULL,                        NULL,
190                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
191                    NULL,
192                    LLM_FFN_GELU, LLM_FFN_SEQ, il);
193            cb(cur, "ffn_out", il);
194
195            cur = ggml_add(ctx0, cur, ffn_inp);
196
197            cur = build_cvec(cur, il);
198            cb(cur, "l_out", il);
199
200            // input for next layer
201            inpL = cur;
202        }
203    }
204
205    cur = build_norm(inpL,
206            model.output_norm,
207            model.output_norm_b,
208            LLM_NORM, -1);
209
210    cb(cur, "result_norm", -1);
211    res->t_embd = cur;
212
213    cur = build_lora_mm(model.output, cur, model.output_s);
214
215    cb(cur, "result_output", -1);
216    res->t_logits = cur;
217
218    ggml_build_forward_expand(gf, cur);
219}