Felipe97/llama-cpp-compiled
01.1k
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-small17 case 8: type = LLM_TYPE_80M; break; // flan-t5-small18 case 12:19 switch (hparams.n_ff()) {20 case 3072: type = LLM_TYPE_220M; break; // t5-base21 case 2048: type = LLM_TYPE_250M; break; // flan-t5-base22 default: type = LLM_TYPE_UNKNOWN;23 } break;24 case 24:25 switch (hparams.n_ff()) {26 case 4096: type = LLM_TYPE_770M; break; // t5-large27 case 2816: type = LLM_TYPE_780M; break; // flan-t5-large28 case 16384: type = LLM_TYPE_3B; break; // t5-3b29 case 5120: type = LLM_TYPE_3B; break; // flan-t5-xl30 case 65536: type = LLM_TYPE_11B; break; // t5-11b31 case 10240: type = LLM_TYPE_11B; break; // flan-t5-xxl32 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 // output46 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 embed51 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_layers56 // dec_n_layer: number of decoder_layers57 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 layers63 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 layers81 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 implementation94 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 109template <>110llama_model_t5::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {111 const int64_t n_embd_head = hparams.n_embd_head_v();112 //const int64_t n_embd_gqa = hparams.n_embd_v_gqa();113 114 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());115 116 ggml_tensor * cur;117 ggml_tensor * inpL;118 119 inpL = build_inp_embd(model.tok_embd);120 121 ggml_tensor * embd_enc = build_inp_cross_embd();122 ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();123 124 const int64_t n_outputs_enc = embd_enc->ne[1];125 126 auto * inp_attn_self = build_attn_inp_kv();127 auto * inp_attn_cross = build_attn_inp_cross();128 129 ggml_tensor * inp_out_ids = build_inp_out_ids();130 131 const int64_t dec_n_layer = hparams.dec_n_layer;132 133 for (int il = 0; il < dec_n_layer; ++il) {134 ggml_tensor * inpSA = inpL;135 136 // norm137 cur = build_norm(inpL,138 model.layers[il].attn_norm, NULL,139 LLM_NORM_RMS, il);140 cb(cur, "attn_norm", il);141 142 // self-attention143 {144 auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur, n_embd_head, n_head, n_head_kv, il);145 146 ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;147 ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);148 149 cur = build_attn(inp_attn_self,150 model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,151 Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);152 cb(cur, "kqv_out", il);153 }154 cur = ggml_add(ctx0, cur, inpSA);155 cb(cur, "cross_inp", il);156 157 ggml_tensor * inpCA = cur;158 159 // norm160 cur = build_norm(cur,161 model.layers[il].attn_norm_cross, NULL,162 LLM_NORM_RMS, il);163 cb(cur, "attn_norm_cross", il);164 165 // cross-attention166 {167 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);168 cb(Qcur, "Qcur", il);169 170 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);171 cb(Kcur, "Kcur", il);172 173 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);174 cb(Vcur, "Vcur", il);175 176 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);177 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);178 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);179 180 cur = build_attn(inp_attn_cross,181 model.layers[il].wo_cross, nullptr, nullptr,182 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);183 cb(cur, "kqv_out", il);184 185 //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);186 //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));187 188 //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);189 //cb(kq, "kq", il);190 191 //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);192 //cb(kq, "kq_soft_max_ext", il);193 194 //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));195 //cb(v, "v", il);196 197 //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);198 //cb(kqv, "kqv", il);199 200 //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);201 //cb(kqv_merged, "kqv_merged", il);202 203 //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);204 //cb(cur, "kqv_merged_cont", il);205 206 //ggml_build_forward_expand(gf, cur);207 208 //cur = build_lora_mm(model.layers[il].wo_cross, cur);209 //cb(cur, "kqv_out", il);210 }211 if (il == dec_n_layer - 1 && inp_out_ids) {212 cur = ggml_get_rows(ctx0, cur, inp_out_ids);213 inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);214 }215 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);216 cb(ffn_inp, "ffn_inp", il);217 218 // feed-forward network219 {220 cur = build_norm(ffn_inp,221 model.layers[il].ffn_norm, NULL,222 LLM_NORM_RMS, il);223 cb(cur, "ffn_norm", il);224 225 // T5 uses relu, flan-T5 uses gelu-gated226 cur = build_ffn(cur,227 model.layers[il].ffn_up, NULL, NULL,228 model.layers[il].ffn_gate, NULL, NULL,229 model.layers[il].ffn_down, NULL, NULL,230 NULL,231 model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,232 model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,233 il);234 cb(cur, "ffn_out", il);235 }236 cur = ggml_add(ctx0, cur, ffn_inp);237 cb(cur, "ffn_out", il);238 239 cur = build_cvec(cur, il);240 cb(cur, "l_out", il);241 242 // input for next layer243 inpL = cur;244 }245 cur = inpL;246 cb(cur, "result_embd", -1);247 248 cur = build_norm(cur,249 model.output_norm, NULL,250 LLM_NORM_RMS, -1);251 252 cb(cur, "result_norm", -1);253 res->t_embd = cur;254 255 // lm_head256 cur = build_lora_mm(model.output, cur, model.output_s);257 258 cb(cur, "result_output", -1);259 res->t_logits = cur;260 261 ggml_build_forward_expand(gf, cur);262}263 264template <>265llama_model_t5::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {266 const int64_t n_embd_head = hparams.n_embd_head_v();267 268 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());269 270 ggml_tensor * cur;271 ggml_tensor * inpL;272 273 inpL = build_inp_embd(model.tok_embd);274 275 ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc();276 277 auto * inp_attn = build_attn_inp_no_cache();278 279 ggml_tensor * inp_out_ids = build_inp_out_ids();280 281 for (int il = 0; il < n_layer; ++il) {282 ggml_tensor * inpSA = inpL;283 284 // norm285 cur = build_norm(inpL,286 model.layers[il].attn_norm_enc, NULL,287 LLM_NORM_RMS, il);288 cb(cur, "attn_norm", il);289 290 // self-attention291 {292 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur);293 cb(Qcur, "Qcur", il);294 295 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur);296 cb(Kcur, "Kcur", il);297 298 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur);299 cb(Vcur, "Vcur", il);300 301 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);302 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);303 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);304 305 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;306 ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b);307 308 cur = build_attn(inp_attn,309 model.layers[il].wo_enc, nullptr, nullptr,310 Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);311 cb(cur, "kqv_out", il);312 }313 if (il == n_layer - 1 && inp_out_ids) {314 cur = ggml_get_rows(ctx0, cur, inp_out_ids);315 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);316 }317 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);318 cb(ffn_inp, "ffn_inp", il);319 320 // feed-forward network321 {322 cur = build_norm(ffn_inp,323 model.layers[il].ffn_norm_enc, NULL,324 LLM_NORM_RMS, il);325 cb(cur, "ffn_norm", il);326 327 // T5 uses relu, flan-T5 uses gelu-gated328 cur = build_ffn(cur,329 model.layers[il].ffn_up_enc, NULL, NULL,330 model.layers[il].ffn_gate_enc, NULL, NULL,331 model.layers[il].ffn_down_enc, NULL, NULL,332 NULL,333 model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,334 model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,335 il);336 cb(cur, "ffn_out", il);337 }338 cur = ggml_add(ctx0, cur, ffn_inp);339 cb(cur, "ffn_out", il);340 341 cur = build_cvec(cur, il);342 cb(cur, "l_out", il);343 344 // input for next layer345 inpL = cur;346 }347 cur = inpL;348 cb(cur, "result_embd", -1);349 350 cur = build_norm(cur,351 model.output_norm_enc, NULL,352 LLM_NORM_RMS, -1);353 354 cb(cur, "result_norm", -1);355 res->t_embd = cur;356 357 ggml_build_forward_expand(gf, cur);358}359 360std::unique_ptr<llm_graph_context> llama_model_t5::build_arch_graph(const llm_graph_params & params) const {361 switch (params.gtype) {362 case LLM_GRAPH_TYPE_ENCODER:363 return std::make_unique<graph<true>>(*this, params);364 case LLM_GRAPH_TYPE_DEFAULT:365 case LLM_GRAPH_TYPE_DECODER:366 return std::make_unique<graph<false>>(*this, params);367 default:368 GGML_ABORT("invalid graph type");369 };370}371