Encode a single pre-tokenized piece using BPE merges.
| 362 | |
| 363 | // Encode a single pre-tokenized piece using BPE merges. |
| 364 | std::vector<int32_t> Tokenizer::bpe_encode_piece(const std::string & piece) const { |
| 365 | if (piece.empty()) return {}; |
| 366 | |
| 367 | std::vector<std::string> symbols; |
| 368 | |
| 369 | if (is_sentencepiece_) { |
| 370 | // SentencePiece: replace leading space with ▁, tokens are raw UTF-8. |
| 371 | std::string sp_piece; |
| 372 | sp_piece.reserve(piece.size() + 2); |
| 373 | size_t start = 0; |
| 374 | if (!piece.empty() && piece[0] == ' ') { |
| 375 | sp_piece += "\xe2\x96\x81"; // ▁ (U+2581) |
| 376 | start = 1; |
| 377 | } |
| 378 | sp_piece += piece.substr(start); |
| 379 | // Replace any remaining spaces with ▁ |
| 380 | std::string encoded; |
| 381 | encoded.reserve(sp_piece.size()); |
| 382 | for (char c : sp_piece) { |
| 383 | if (c == ' ') { |
| 384 | encoded += "\xe2\x96\x81"; |
| 385 | } else { |
| 386 | encoded += c; |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | // Try whole piece as single token. |
| 391 | auto it = token_to_id_.find(encoded); |
| 392 | if (it != token_to_id_.end()) { |
| 393 | return { it->second }; |
| 394 | } |
| 395 | |
| 396 | // Split into individual UTF-8 characters as initial BPE symbols. |
| 397 | const char * p = encoded.c_str(); |
| 398 | const char * end = p + encoded.size(); |
| 399 | while (p < end) { |
| 400 | int cplen; |
| 401 | utf8_decode(p, (size_t)(end - p), &cplen); |
| 402 | if (cplen <= 0) cplen = 1; |
| 403 | std::string sym(p, cplen); |
| 404 | auto sit = token_to_id_.find(sym); |
| 405 | if (sit != token_to_id_.end()) { |
| 406 | symbols.push_back(sym); |
| 407 | } else { |
| 408 | // Byte-fallback: <0xNN> |
| 409 | char buf[8]; |
| 410 | std::snprintf(buf, sizeof(buf), "<0x%02X>", (unsigned)(uint8_t)*p); |
| 411 | symbols.push_back(buf); |
| 412 | } |
| 413 | p += cplen; |
| 414 | } |
| 415 | } else { |
| 416 | // GPT-2 BPE: convert raw text to GPT-2 byte encoding for vocab lookup. |
| 417 | std::string encoded = encode_gpt2_bpe(piece); |
| 418 | |
| 419 | // Try to find the encoded piece as a single token first. |
| 420 | auto it = token_to_id_.find(encoded); |
| 421 | if (it != token_to_id_.end()) { |
nothing calls this directly
no test coverage detected