vocab : refactor normalizer flags into options struct, add strip_accents (#24371)

* vocab : refactor normalizer flags into options struct, add strip_accents

* Update src/llama-vocab.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update src/llama-vocab.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

gguf-py/gguf/constants.py

@@ -273,6 +273,7 @@ class Keys:
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # Normalizer constants
         NORMALIZER_LOWERCASE     = "tokenizer.ggml.normalizer.lowercase"
+        NORMALIZER_STRIP_ACCENTS = "tokenizer.ggml.normalizer.strip_accents"
         # FIM/Infill special tokens constants
         FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
         FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"

gguf-py/gguf/gguf_writer.py

@@ -1124,6 +1124,9 @@ class GGUFWriter:
     def add_normalizer_lowercase(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.NORMALIZER_LOWERCASE, value)
 
+    def add_normalizer_strip_accents(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.NORMALIZER_STRIP_ACCENTS, value)
+
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 

gguf-py/gguf/vocab.py

@@ -53,6 +53,7 @@ class SpecialVocab:
     special_token_ids: dict[str, int]
     chat_template: str | Sequence[Mapping[str, str]] | None
     normalizer_lowercase: bool | None
+    normalizer_strip_accents: bool | None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
@@ -66,6 +67,7 @@ class SpecialVocab:
         self.merges = []
         self.chat_template = None
         self.normalizer_lowercase = None
+        self.normalizer_strip_accents = None
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
@@ -108,6 +110,10 @@ class SpecialVocab:
             if not quiet:
                 logger.info(f'Setting normalizer_lowercase to {self.normalizer_lowercase}')
             gw.add_normalizer_lowercase(self.normalizer_lowercase)
+        if self.normalizer_strip_accents is not None:
+            if not quiet:
+                logger.info(f'Setting normalizer_strip_accents to {self.normalizer_strip_accents}')
+            gw.add_normalizer_strip_accents(self.normalizer_strip_accents)
 
     def _load(self, path: Path) -> None:
         self._try_load_from_tokenizer_json(path)
@@ -155,17 +161,21 @@ class SpecialVocab:
     def _parse_normalizer(self, normalizer: dict) -> None:
         # ref: https://huggingface.co/docs/tokenizers/api/normalizers
         #
-        # Detects lowercase normalization in three possible formats:
+        # Extracts normalizer flags from three possible formats:
         # 1. Standalone:           {"type": "Lowercase"}
-        # 2. BertNormalizer attribute: {"type": "BertNormalizer", "lowercase": true, ...}
+        # 2. BertNormalizer attrs: {"type": "BertNormalizer", ...}
         # 3. Nested in Sequence:   {"type": "Sequence", "normalizers": [...]}
 
         normalizer_type = normalizer.get('type')
         if normalizer_type == 'Lowercase':
             self.normalizer_lowercase = True
+        elif normalizer_type == 'StripAccents':
+            self.normalizer_strip_accents = True
         elif normalizer_type == 'BertNormalizer':
             if 'lowercase' in normalizer:
                 self.normalizer_lowercase = normalizer['lowercase']
+            if 'strip_accents' in normalizer:
+                self.normalizer_strip_accents = normalizer['strip_accents']
         elif normalizer_type == 'Sequence':
             for norm in normalizer.get('normalizers', []):
                 self._parse_normalizer(norm)

src/llama-arch.cpp

@@ -325,6 +325,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_RWKV,                     "tokenizer.rwkv.world"                    },
     { LLM_KV_TOKENIZER_CHAT_TEMPLATE,            "tokenizer.chat_template"                 },
     { LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     "tokenizer.ggml.normalizer.lowercase"     },
+    { LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, "tokenizer.ggml.normalizer.strip_accents" },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,               "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,               "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,               "tokenizer.ggml.fim_mid_token_id"         },

src/llama-arch.h

@@ -314,6 +314,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
     LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,
+    LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,

src/llama-vocab.cpp

@@ -764,7 +764,7 @@ struct llm_tokenizer_wpm_session {
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) {
         // normalize and split by whitespace
-        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_lowercase());
+        std::vector<std::string> words = preprocess(text, vocab.get_normalizer_opts());
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -809,11 +809,14 @@ struct llm_tokenizer_wpm_session {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    static std::vector<std::string> preprocess(const std::string & text, bool lowercase)  {
+    static std::vector<std::string> preprocess(const std::string & text, const llama_vocab::normalizer_options & normalizer_opts)  {
-        const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
+        std::vector<uint32_t> cpts = unicode_cpts_from_utf8(text);
+        if (normalizer_opts.strip_accents) {
+            cpts = unicode_cpts_normalize_nfd(cpts);
+        }
         std::vector<std::string> words(1, "");
 
-        for (const uint32_t cpt : cpts_nfd) {
+        for (const uint32_t cpt : cpts) {
             const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
@@ -828,7 +831,11 @@ struct llm_tokenizer_wpm_session {
                 continue;
             }
 
-            const std::string s = unicode_cpt_to_utf8(lowercase ? unicode_tolower(cpt) : cpt);
+            if (normalizer_opts.strip_accents && flags.is_accent_mark) {
+                continue;
+            }
+
+            const std::string s = unicode_cpt_to_utf8(normalizer_opts.lowercase ? unicode_tolower(cpt) : cpt);
             if (flags.is_punctuation || ( cpt < 0x7F && flags.is_symbol ) || is_chinese_char(cpt)) {
                 if (words.back().size()) {  // finish previous word if any
                     words.emplace_back();
@@ -1692,7 +1699,7 @@ struct llm_tokenizer_whitespace_session : llm_tokenizer_bpe_session {
     llm_tokenizer_whitespace_session(const llama_vocab & vocab, const llm_tokenizer_bpe & tokenizer) : llm_tokenizer_bpe_session{vocab, tokenizer}, vocab{vocab} {}
 
     void tokenize(const std::string & text, std::vector<llama_token> & output) override {
-        const bool lowercase = vocab.get_normalizer_lowercase();
+        const bool lowercase = vocab.get_normalizer_opts().lowercase;
 
         std::string segment;
         auto flush = [&]() {
@@ -1797,7 +1804,9 @@ struct llama_vocab::impl {
     bool remove_extra_whitespaces   = false;
     bool escape_whitespaces         = true;
     bool treat_whitespace_as_suffix = false;
-    bool normalizer_lowercase       = true; // Lowercase normalizer (tokenizer.json)
+
+    // BertNormalizer options
+    llama_vocab::normalizer_options normalizer_opts;
 
     std::unordered_map<std::string, llama_token> token_to_id;
     std::vector<token_data>                      id_to_token;
@@ -2172,7 +2181,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                     tokenizer_pre == "whitespace") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_WHITESPACE;
-                normalizer_lowercase = false;
+                normalizer_opts.lowercase = false;
             } else if (
                     tokenizer_pre == "refact") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT;
@@ -2532,8 +2541,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
-        // Lowercase normalizer flag (consulted by WPM / whitespace BPE)
+        // BertNormalizer options
-        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE, normalizer_lowercase, false);
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_LOWERCASE,     normalizer_opts.lowercase,     false);
+        normalizer_opts.strip_accents = normalizer_opts.lowercase;
+        ml.get_key(LLM_KV_TOKENIZER_NORMALIZER_STRIP_ACCENTS, normalizer_opts.strip_accents, false);
 
         // suppress tokens
         {
@@ -3969,8 +3980,8 @@ bool llama_vocab::get_treat_whitespace_as_suffix() const {
     return pimpl->treat_whitespace_as_suffix;
 }
 
-bool llama_vocab::get_normalizer_lowercase() const {
+const llama_vocab::normalizer_options & llama_vocab::get_normalizer_opts() const {
-    return pimpl->normalizer_lowercase;
+    return pimpl->normalizer_opts;
 }
 
 const std::vector<llama_token> & llama_vocab::get_suppress_tokens() const {

src/llama-vocab.h

@@ -76,6 +76,12 @@ struct llama_vocab {
         llama_token_attr attr;
     };
 
+    struct normalizer_options {
+        bool lowercase     = true;
+        bool strip_accents = true;
+        // TODO: clean_text, handle_chinese_chars
+    };
+
     llama_vocab();
     ~llama_vocab();
 
@@ -141,7 +147,7 @@ struct llama_vocab {
     bool get_remove_extra_whitespaces  () const;
     bool get_escape_whitespaces        () const;
     bool get_treat_whitespace_as_suffix() const;
-    bool get_normalizer_lowercase      () const;
+    const normalizer_options & get_normalizer_opts() const;
 
     const std::vector<llama_token> & get_suppress_tokens() const;