From 6f4c300bfff3fcdf2df9d8a6bde09091d4baab61 Mon Sep 17 00:00:00 2001 From: jaime-m-p <> Date: Sat, 25 May 2024 02:20:55 +0200 Subject: [PATCH] Refactor llm_tokenizer_bpe: move code to constructor --- llama.cpp | 192 +++++++++++++++++++++++++++--------------------------- 1 file changed, 95 insertions(+), 97 deletions(-) diff --git a/llama.cpp b/llama.cpp index 15c660775..1ffe4d9af 100644 --- a/llama.cpp +++ b/llama.cpp @@ -12272,106 +12272,100 @@ struct llm_bigram_bpe { }; struct llm_tokenizer_bpe { - llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {} + llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) { + GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE); + switch (vocab.type_pre) { + case LLAMA_VOCAB_PRE_TYPE_LLAMA3: + ignore_merges = true; + regex_exprs = { + // original regex from tokenizer.json + //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + + // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DBRX: + regex_exprs = { + // same as llama3 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: + regex_exprs = { + "[\r\n]", + "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", + "\\s?[!-/:-~！-／：-～‘-‟　-。]+", + "\\s+$", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: + regex_exprs = { + "[\r\n]", + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_FALCON: + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "[0-9][0-9][0-9]", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_MPT: + // TODO: MPT pre-tokenization regexes are unknown + // the following are close, but not exact. run the following: + // ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf + GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed"); + regex_exprs = { + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STARCODER: + case LLAMA_VOCAB_PRE_TYPE_REFACT: + case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: + regex_exprs = { + "\\p{N}", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_GPT2: + case LLAMA_VOCAB_PRE_TYPE_OLMO: + regex_exprs = { + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STABLELM2: + case LLAMA_VOCAB_PRE_TYPE_QWEN2: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + default: + // default regex for BPE tokenization pre-processing + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "\\p{N}+", + "[0-9][0-9][0-9]", + }; + break; + } + } void tokenize(const std::string & text, std::vector & output) { int final_prev_index = -1; - bool ignore_merges = false; - std::vector word_collection; - switch (vocab.type) { - case LLAMA_VOCAB_TYPE_BPE: - switch (vocab.type_pre) { - case LLAMA_VOCAB_PRE_TYPE_LLAMA3: - ignore_merges = true; - word_collection = unicode_regex_split(text, { - // original regex from tokenizer.json - //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - - // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DBRX: - word_collection = unicode_regex_split(text, { - // same as llama3 - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: - word_collection = unicode_regex_split(text, { - "[\r\n]", - "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", - "\\s?[!-/:-~！-／：-～‘-‟　-。]+", - "\\s+$", - "[一-龥ࠀ-一가-퟿]+", - "\\p{N}+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: - word_collection = unicode_regex_split(text, { - "[\r\n]", - "\\s?\\p{L}+", - "\\s?\\p{P}+", - "[一-龥ࠀ-一가-퟿]+", - "\\p{N}", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_FALCON: - word_collection = unicode_regex_split(text, { - "[\\p{P}\\$\\+<=>\\^~\\|]+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - "[0-9][0-9][0-9]", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_MPT: - // TODO: MPT pre-tokenization regexes are unknown - // the following are close, but not exact. run the following: - // ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf - GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed"); - word_collection = unicode_regex_split(text, { - "\\s?\\p{L}+", - "\\s?\\p{P}+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_STARCODER: - case LLAMA_VOCAB_PRE_TYPE_REFACT: - case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: - word_collection = unicode_regex_split(text, { - "\\p{N}", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_GPT2: - case LLAMA_VOCAB_PRE_TYPE_OLMO: - word_collection = unicode_regex_split(text, { - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_STABLELM2: - case LLAMA_VOCAB_PRE_TYPE_QWEN2: - word_collection = unicode_regex_split(text, { - // original regex from tokenizer.json - // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - default: - // default regex for BPE tokenization pre-processing - word_collection = unicode_regex_split(text, { - "[\\p{P}\\$\\+<=>\\^~\\|]+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - "\\p{N}+", - "[0-9][0-9][0-9]", - }); - break; - } - break; - default: - GGML_ASSERT(false); - break; - } + const auto word_collection = unicode_regex_split(text, regex_exprs); symbols_final.clear(); @@ -12505,6 +12499,9 @@ private: const llama_vocab & vocab; + bool ignore_merges = false; + std::vector regex_exprs; + std::vector symbols; std::vector symbols_final; @@ -12852,6 +12849,8 @@ static std::vector llama_tokenize_internal(const llama_vocab & } break; case LLAMA_VOCAB_TYPE_BPE: { + llm_tokenizer_bpe tokenizer(vocab); + if (add_special && vocab.special_add_bos != 0) { GGML_ASSERT(vocab.special_bos_id != -1); output.push_back(vocab.special_bos_id); @@ -12864,7 +12863,6 @@ static std::vector llama_tokenize_internal(const llama_vocab & #ifdef PRETOKENIZERDEBUG LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); #endif - llm_tokenizer_bpe tokenizer(vocab); tokenizer.tokenize(raw_text, output); } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) output.push_back(fragment.token);