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 →
14.7 KiB370 linesraw
1#include "models.h"
2
3void llama_model_t5::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_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts);
6
7    uint32_t dec_start_token_id;
8    if (ml.get_key(LLM_KV_DECODER_START_TOKEN_ID, dec_start_token_id, false)) {
9        hparams.dec_start_token_id = dec_start_token_id;
10    }
11
12    hparams.dec_n_layer = hparams.n_layer();
13    ml.get_key(LLM_KV_DECODER_BLOCK_COUNT, hparams.dec_n_layer, false);
14
15    switch (hparams.n_layer()) {
16        case 6:  type = LLM_TYPE_60M;  break; // t5-small
17        case 8:  type = LLM_TYPE_80M;  break; // flan-t5-small
18        case 12:
19            switch (hparams.n_ff()) {
20                case 3072: type = LLM_TYPE_220M; break; // t5-base
21                case 2048: type = LLM_TYPE_250M; break; // flan-t5-base
22                default: type = LLM_TYPE_UNKNOWN;
23            } break;
24        case 24:
25            switch (hparams.n_ff()) {
26                case 4096:  type = LLM_TYPE_770M; break; // t5-large
27                case 2816:  type = LLM_TYPE_780M; break; // flan-t5-large
28                case 16384: type = LLM_TYPE_3B;   break; // t5-3b
29                case 5120:  type = LLM_TYPE_3B;   break; // flan-t5-xl
30                case 65536: type = LLM_TYPE_11B;  break; // t5-11b
31                case 10240: type = LLM_TYPE_11B;  break; // flan-t5-xxl
32                default: type = LLM_TYPE_UNKNOWN;
33            } break;
34        default: type = LLM_TYPE_UNKNOWN;
35   }
36}
37
38void llama_model_t5::load_arch_tensors(llama_model_loader &) {
39    LLAMA_LOAD_LOCALS;
40
41    const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts;
42
43    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
44
45    // output
46    output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}, 0);
47    output_norm     = create_tensor(tn(LLM_TENSOR_DEC_OUTPUT_NORM, "weight"), {n_embd}, 0);
48
49    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
50    // if output is NULL, init from the input tok embed
51    if (output == NULL) {
52        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
53    }
54
55    // n_layer:     number of encoder_layers
56    // dec_n_layer: number of decoder_layers
57    const int dec_n_layer = hparams.dec_n_layer;
58    if (dec_n_layer > n_layer) {
59        layers.resize(dec_n_layer);
60    }
61
62    // load encoder layers
63    for (int i = 0; i < n_layer; ++i) {
64        auto & layer = layers[i];
65
66        layer.attn_norm_enc  = create_tensor(tn(LLM_TENSOR_ENC_ATTN_NORM,  "weight", i), {n_embd}, 0);
67        layer.attn_rel_b_enc = create_tensor(tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, TENSOR_NOT_REQUIRED);
68
69        layer.wq_enc = create_tensor(tn(LLM_TENSOR_ENC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
70        layer.wk_enc = create_tensor(tn(LLM_TENSOR_ENC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
71        layer.wv_enc = create_tensor(tn(LLM_TENSOR_ENC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
72        layer.wo_enc = create_tensor(tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}, 0);
73
74        layer.ffn_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd}, 0);
75        layer.ffn_gate_enc = create_tensor(tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
76        layer.ffn_down_enc = create_tensor(tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
77        layer.ffn_up_enc   = create_tensor(tn(LLM_TENSOR_ENC_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
78    }
79
80    // load decoder layers
81    for (int i = 0; i < dec_n_layer; ++i) {
82        auto & layer = layers[i];
83
84        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_DEC_ATTN_NORM,  "weight", i), {n_embd}, 0);
85        layer.attn_rel_b = create_tensor(tn(LLM_TENSOR_DEC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, TENSOR_NOT_REQUIRED);
86
87        layer.wq = create_tensor(tn(LLM_TENSOR_DEC_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
88        layer.wk = create_tensor(tn(LLM_TENSOR_DEC_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
89        layer.wv = create_tensor(tn(LLM_TENSOR_DEC_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
90        layer.wo = create_tensor(tn(LLM_TENSOR_DEC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}, 0);
91
92        layer.attn_norm_cross  = create_tensor(tn(LLM_TENSOR_DEC_CROSS_ATTN_NORM,  "weight", i), {n_embd}, 0);
93        // this tensor seems to be unused in HF transformers implementation
94        layer.attn_rel_b_cross = create_tensor(
95            tn(LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
96
97        layer.wq_cross = create_tensor(tn(LLM_TENSOR_DEC_CROSS_ATTN_Q,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
98        layer.wk_cross = create_tensor(tn(LLM_TENSOR_DEC_CROSS_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
99        layer.wv_cross = create_tensor(tn(LLM_TENSOR_DEC_CROSS_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
100        layer.wo_cross = create_tensor(tn(LLM_TENSOR_DEC_CROSS_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}, 0);
101
102        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_DEC_FFN_NORM, "weight", i), {n_embd}, 0);
103        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_DEC_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED);
104        layer.ffn_down = create_tensor(tn(LLM_TENSOR_DEC_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
105        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_DEC_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
106    }
107}
108
109std::unique_ptr<llm_graph_context> llama_model_t5::build_arch_graph(const llm_graph_params & params) const {
110    switch (params.gtype) {
111        case LLM_GRAPH_TYPE_ENCODER:
112            return std::make_unique<graph<true>>(*this, params);
113        case LLM_GRAPH_TYPE_DEFAULT:
114        case LLM_GRAPH_TYPE_DECODER:
115            return std::make_unique<graph<false>>(*this, params);
116        default:
117            GGML_ABORT("invalid graph type");
118    };
119}
120
121template <>
122llama_model_t5::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
123    const int64_t n_embd_head = hparams.n_embd_head_v();
124    //const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
125
126    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
127
128    ggml_tensor * cur;
129    ggml_tensor * inpL;
130
131    inpL = build_inp_embd(model.tok_embd);
132
133    ggml_tensor * embd_enc       = build_inp_cross_embd();
134    ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();
135
136    const int64_t n_outputs_enc = embd_enc->ne[1];
137
138    auto * inp_attn_self  = build_attn_inp_kv();
139    auto * inp_attn_cross = build_attn_inp_cross();
140
141    ggml_tensor * inp_out_ids = build_inp_out_ids();
142
143    const int64_t dec_n_layer = hparams.dec_n_layer;
144
145    for (int il = 0; il < dec_n_layer; ++il) {
146        ggml_tensor * inpSA = inpL;
147
148        // norm
149        cur = build_norm(inpL,
150                model.layers[il].attn_norm, NULL,
151                LLM_NORM_RMS, il);
152        cb(cur, "attn_norm", il);
153
154        // self-attention
155        {
156            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il);
157
158            ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
159            ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
160
161            cur = build_attn(inp_attn_self,
162                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
163                    Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
164            cb(cur, "kqv_out", il);
165        }
166        cur = ggml_add(ctx0, cur, inpSA);
167        cb(cur, "cross_inp", il);
168
169        ggml_tensor * inpCA = cur;
170
171        // norm
172        cur = build_norm(cur,
173                model.layers[il].attn_norm_cross, NULL,
174                LLM_NORM_RMS, il);
175        cb(cur, "attn_norm_cross", il);
176
177        // cross-attention
178        {
179            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);
180            cb(Qcur, "Qcur", il);
181
182            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);
183            cb(Kcur, "Kcur", il);
184
185            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);
186            cb(Vcur, "Vcur", il);
187
188            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
189            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
190            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);
191
192            cur = build_attn(inp_attn_cross,
193                    model.layers[il].wo_cross, nullptr, nullptr,
194                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
195            cb(cur, "kqv_out", il);
196
197            //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
198            //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
199
200            //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
201            //cb(kq, "kq", il);
202
203            //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
204            //cb(kq, "kq_soft_max_ext", il);
205
206            //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
207            //cb(v, "v", il);
208
209            //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
210            //cb(kqv, "kqv", il);
211
212            //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
213            //cb(kqv_merged, "kqv_merged", il);
214
215            //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
216            //cb(cur, "kqv_merged_cont", il);
217
218            //ggml_build_forward_expand(gf, cur);
219
220            //cur = build_lora_mm(model.layers[il].wo_cross, cur);
221            //cb(cur, "kqv_out", il);
222        }
223        if (il == dec_n_layer - 1 && inp_out_ids) {
224            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
225            inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
226        }
227        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
228        cb(ffn_inp, "ffn_inp", il);
229
230        // feed-forward network
231        {
232            cur = build_norm(ffn_inp,
233                    model.layers[il].ffn_norm, NULL,
234                    LLM_NORM_RMS, il);
235            cb(cur, "ffn_norm", il);
236
237            // T5 uses relu, flan-T5 uses gelu-gated
238            cur = build_ffn(cur,
239                    model.layers[il].ffn_up,   NULL, NULL,
240                    model.layers[il].ffn_gate, NULL, NULL,
241                    model.layers[il].ffn_down, NULL, NULL,
242                    NULL,
243                    model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,
244                    model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,
245                    il);
246            cb(cur, "ffn_out", il);
247        }
248        cur = ggml_add(ctx0, cur, ffn_inp);
249        cb(cur, "ffn_out", il);
250
251        cur = build_cvec(cur, il);
252        cb(cur, "l_out", il);
253
254        // input for next layer
255        inpL = cur;
256    }
257    cur = inpL;
258    cb(cur, "result_embd", -1);
259
260    cur = build_norm(cur,
261            model.output_norm, NULL,
262            LLM_NORM_RMS, -1);
263
264    cb(cur, "result_norm", -1);
265    res->t_embd = cur;
266
267    // lm_head
268    cur = build_lora_mm(model.output, cur, model.output_s);
269
270    cb(cur, "result_output", -1);
271    res->t_logits = cur;
272
273    ggml_build_forward_expand(gf, cur);
274}
275
276template <>
277llama_model_t5::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
278    const int64_t n_embd_head = hparams.n_embd_head_v();
279
280    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
281
282    ggml_tensor * cur;
283    ggml_tensor * inpL;
284
285    inpL = build_inp_embd(model.tok_embd);
286
287    ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc();
288
289    auto * inp_attn = build_attn_inp_no_cache();
290
291    ggml_tensor * inp_out_ids = build_inp_out_ids();
292
293    for (int il = 0; il < n_layer; ++il) {
294        ggml_tensor * inpSA = inpL;
295
296        // norm
297        cur = build_norm(inpL,
298                model.layers[il].attn_norm_enc, NULL,
299                LLM_NORM_RMS, il);
300        cb(cur, "attn_norm", il);
301
302        // self-attention
303        {
304            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur);
305            cb(Qcur, "Qcur", il);
306
307            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur);
308            cb(Kcur, "Kcur", il);
309
310            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur);
311            cb(Vcur, "Vcur", il);
312
313            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
314            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
315            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
316
317            ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
318            ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b);
319
320            cur = build_attn(inp_attn,
321                    model.layers[il].wo_enc, nullptr, nullptr,
322                    Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
323            cb(cur, "kqv_out", il);
324        }
325        if (il == n_layer - 1 && inp_out_ids) {
326            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
327            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
328        }
329        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
330        cb(ffn_inp, "ffn_inp", il);
331
332        // feed-forward network
333        {
334            cur = build_norm(ffn_inp,
335                    model.layers[il].ffn_norm_enc, NULL,
336                    LLM_NORM_RMS, il);
337            cb(cur, "ffn_norm", il);
338
339            // T5 uses relu, flan-T5 uses gelu-gated
340            cur = build_ffn(cur,
341                    model.layers[il].ffn_up_enc,   NULL, NULL,
342                    model.layers[il].ffn_gate_enc, NULL, NULL,
343                    model.layers[il].ffn_down_enc, NULL, NULL,
344                    NULL,
345                    model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
346                    model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
347                    il);
348            cb(cur, "ffn_out", il);
349        }
350        cur = ggml_add(ctx0, cur, ffn_inp);
351        cb(cur, "ffn_out", il);
352
353        cur = build_cvec(cur, il);
354        cb(cur, "l_out", il);
355
356        // input for next layer
357        inpL = cur;
358    }
359    cur = inpL;
360    cb(cur, "result_embd", -1);
361
362    cur = build_norm(cur,
363            model.output_norm_enc, NULL,
364            LLM_NORM_RMS, -1);
365
366    cb(cur, "result_norm", -1);
367    res->t_embd = cur;
368
369    ggml_build_forward_expand(gf, cur);
370}