Merge remote-tracking branch 'upstream/compilade/fix-mpt-pretok' into tokenizer-fixes

jaime-m-p · jaime-m-p · commit fb46a1556406 · 2024-07-11T23:07:36.000+02:00
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
@@ -373,6 +373,29 @@ def from_model_architecture(cls, arch: str) -> type[Model]:
         except KeyError:
             raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
 
+    def does_token_look_special(self, token: str | bytes) -> bool:
+        if isinstance(token, (bytes, bytearray)):
+            token_text = token.decode(encoding="utf-8")
+        elif isinstance(token, memoryview):
+            token_text = token.tobytes().decode(encoding="utf-8")
+        else:
+            token_text = token
+
+        # Some models mark some added tokens which ought to be control tokens as not special.
+        # (e.g. command-r, command-r-plus, deepseek-coder, gemma{,-2})
+        seems_special = token_text in (
+            "<pad>",  # deepseek-coder
+            "<mask>", "<2mass>", "[@BOS@]",  # gemma{,-2}
+        )
+
+        seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>"))
+        seems_special = seems_special or (token_text.startswith("<｜") and token_text.endswith("｜>"))  # deepseek-coder
+
+        # TODO: should these be marked as UNUSED instead? (maybe not)
+        seems_special = seems_special or (token_text.startswith("<unused") and token_text.endswith(">"))  # gemma{,-2}
+
+        return seems_special
+
     # used for GPT-2 BPE and WordPiece vocabs
     def get_vocab_base(self) -> tuple[list[str], list[int], str]:
         tokens: list[str] = []
@@ -391,16 +414,18 @@ def get_vocab_base(self) -> tuple[list[str], list[int], str]:
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.USER_DEFINED)
-            elif reverse_vocab[i] in added_vocab:
-                tokens.append(reverse_vocab[i])
-                if tokenizer.added_tokens_decoder[i].special:
-                    toktypes.append(gguf.TokenType.CONTROL)
-                else:
-                    toktypes.append(gguf.TokenType.USER_DEFINED)
+                toktypes.append(gguf.TokenType.UNUSED)
             else:
-                tokens.append(reverse_vocab[i])
-                toktypes.append(gguf.TokenType.NORMAL)
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
 
         return tokens, toktypes, tokpre
 
@@ -559,7 +584,7 @@ def _set_vocab_qwen(self):
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.USER_DEFINED)
+                toktypes.append(gguf.TokenType.UNUSED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.CONTROL)
@@ -609,7 +634,7 @@ def _create_vocab_sentencepiece(self):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.IdToPiece(token_id)
@@ -644,6 +669,25 @@ def _create_vocab_sentencepiece(self):
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
+                for token_id, token_data in added_tokens_decoder.items():
+                    token_id = int(token_id)
+                    token: str = token_data["content"]
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
+                        assert tokens[token_id] == token.encode("utf-8")
+                    if token_data.get("special") or self.does_token_look_special(token):
+                        toktypes[token_id] = SentencePieceTokenTypes.CONTROL
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+                    scores[token_id] = -1000.0
+                    tokens[token_id] = token.encode("utf-8")
+
         if vocab_size > len(tokens):
             pad_count = vocab_size - len(tokens)
             logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
@@ -1267,7 +1311,7 @@ def set_vocab(self):
         if (self.dir_model / "tokenizer.json").is_file():
             self._set_vocab_gpt2()
         else:
-            # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab
+            # StableLM 2 1.6B used to have a vocab in a similar format to Qwen's vocab
             self._set_vocab_qwen()
 
     def set_gguf_parameters(self):
@@ -1579,7 +1623,6 @@ def set_gguf_parameters(self):
         self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
 
         self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
-        self.gguf_writer.add_file_type(self.ftype)
 
         self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
         self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
@@ -1873,7 +1916,7 @@ def set_vocab(self):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
 
@@ -1918,7 +1961,7 @@ def set_vocab(self):
                 for token_id, foken_data in added_tokens_decoder.items():
                     token_id = int(token_id)
                     token = foken_data["content"].encode("utf-8")
-                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         assert tokens[token_id] == token
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
@@ -1934,7 +1977,7 @@ def set_vocab(self):
                 for foken_data in added_tokens:
                     token_id = int(foken_data["id"])
                     token = foken_data["content"].encode("utf-8")
-                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         assert tokens[token_id] == token
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
@@ -2146,7 +2189,7 @@ def set_vocab(self):
                 toktype = SentencePieceTokenTypes.BYTE
             # take care of ununsed raw token
             if piece.startswith('[UNUSED'):
-                toktype = SentencePieceTokenTypes.UNKNOWN
+                toktype = SentencePieceTokenTypes.UNUSED
 
             tokens.append(text)
             scores.append(score)
@@ -2176,7 +2219,7 @@ def set_vocab(self):
                     if token == chat_eos_token:
                         chat_eos_token_id = token_id
                     token = token.encode("utf-8")
-                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         assert(tokens[token_id] == token)
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
@@ -2195,7 +2238,7 @@ def set_vocab(self):
                     if token == chat_eos_token:
                         chat_eos_token_id = token_id
                     token = token.encode("utf-8")
-                    if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
+                    if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                         assert(tokens[token_id] == token)
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
@@ -2435,19 +2478,7 @@ class Gemma2Model(Model):
     model_arch = gguf.MODEL_ARCH.GEMMA2
 
     def set_vocab(self):
-        tokens, scores, toktypes = self._create_vocab_sentencepiece()
-        # hack: This is required so that we can properly use start/end-of-turn for chat template
-        for i in range(108):
-            # including <unusedX>, <start_of_turn>, <end_of_turn>
-            toktypes[i] = SentencePieceTokenTypes.CONTROL
-        self.gguf_writer.add_tokenizer_model("llama")
-        self.gguf_writer.add_tokenizer_pre("default")
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-        special_vocab.add_to_gguf(self.gguf_writer)
+        self._set_vocab_sentencepiece()
 
         self.gguf_writer.add_add_space_prefix(False)
 
@@ -2771,7 +2802,7 @@ def set_vocab(self):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
 
@@ -3026,7 +3057,7 @@ def set_vocab(self):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.IdToPiece(token_id)
@@ -3244,15 +3275,14 @@ def set_vocab_chatglm3(self):
             if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():
                 score = tokenizer.tokenizer.sp_model.get_score(token_id)
 
-            if len(piece) == 0:
-                text = f"[PAD{token_id}]".encode("utf-8")
-
             if token_id >= tokenizer.tokenizer.sp_model.vocab_size():
                 if piece in special_tokens:
-                    # show special tokens in prompt
-                    toktype = SentencePieceTokenTypes.USER_DEFINED
+                    toktype = SentencePieceTokenTypes.CONTROL
+                elif len(piece) == 0:
+                    text = f"[PAD{token_id}]".encode("utf-8")
+                    toktype = SentencePieceTokenTypes.UNUSED
                 else:
-                    toktype = SentencePieceTokenTypes.UNKNOWN
+                    toktype = SentencePieceTokenTypes.USER_DEFINED
                 tokens.append(text)
                 scores.append(score)
                 toktypes.append(toktype)
@@ -3341,7 +3371,7 @@ def set_vocab(self):
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.USER_DEFINED)
+                toktypes.append(gguf.TokenType.UNUSED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
                 if tokenizer.added_tokens_decoder[i].special:
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -5419,6 +5419,7 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "command-r") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_COMMAND_R;
+                vocab.tokenizer_clean_spaces = false;
             } else if (
                 tokenizer_pre == "qwen2") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2;
@@ -5652,7 +5653,7 @@ static void llm_load_vocab(
     // build special tokens cache
     {
         for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) {
-            if (!(vocab.id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL)) {
+            if (vocab.id_to_token[id].attr & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_USER_DEFINED | LLAMA_TOKEN_ATTR_UNKNOWN)) {
                 vocab.cache_special_tokens.push_back(id);
             }
         }
@@ -15418,17 +15419,6 @@ struct llm_tokenizer_bpe {
                     "[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:
@@ -15438,6 +15428,7 @@ struct llm_tokenizer_bpe {
                 };
                 break;
             case LLAMA_VOCAB_PRE_TYPE_GPT2:
+            case LLAMA_VOCAB_PRE_TYPE_MPT:
             case LLAMA_VOCAB_PRE_TYPE_OLMO:
             case LLAMA_VOCAB_PRE_TYPE_JAIS:
                 regex_exprs = {
@@ -15464,8 +15455,8 @@ struct llm_tokenizer_bpe {
                 break;
             case LLAMA_VOCAB_PRE_TYPE_VIKING:
                 regex_exprs = {
-                    "\\p{N}",
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
+                    "\\p{N}",
                 };
                 break;
             default:
@@ -16185,12 +16176,20 @@ struct fragment_buffer_variant {
 
 // #define PRETOKENIZERDEBUG
 
-static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer) {
+static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer, bool parse_special) {
     // for each special token
     for (const llama_vocab::id special_id : vocab.cache_special_tokens) {
         const auto & data = vocab.id_to_token[special_id];
         const auto & special_token = data.text;
 
+        if (!parse_special && (data.attr & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_UNKNOWN))) {
+            // Ignore control and unknown tokens when parse_special == false
+            continue;
+            // User-defined tokens are still pre-tokenized before everything else
+            // ref: https://github.com/huggingface/tokenizers/blob/fdd26ba9a3f0c133427aab0423888cbde91362d7/tokenizers/src/tokenizer/mod.rs#L726
+            // This is mostly relevant for neox-style tokenizers (mpt, olmo, stablelm, etc.)
+        }
+
         // for each text fragment
         std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
         while (it != buffer.end()) {
@@ -16303,7 +16302,7 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
 
     if (!raw_text.empty()) {
         fragment_buffer.emplace_front(raw_text, 0, raw_text.length());
-        if (parse_special) tokenizer_st_partition(vocab, fragment_buffer);
+        tokenizer_st_partition(vocab, fragment_buffer, parse_special);
     }
 
     switch (vocab.type) {
diff --git a/tests/test-tokenizer-random.py b/tests/test-tokenizer-random.py
@@ -20,7 +20,7 @@
 from typing_extensions import Buffer
 
 import cffi
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, PreTrainedTokenizer
 
 
 logger = logging.getLogger("test-tokenizer-random")
@@ -145,7 +145,7 @@ def decode(self, ids: list[int]) -> str:
 class TokenizerGroundtruth (Tokenizer):
 
     def __init__(self, dir_tokenizer: str):
-        self.model = AutoTokenizer.from_pretrained(dir_tokenizer, trust_remote_code=False)
+        self.model: PreTrainedTokenizer = AutoTokenizer.from_pretrained(dir_tokenizer, trust_remote_code=False)
         # guess BOS and EOS
         ids = self.encode("a")
         assert 1 <= len(ids) <= 3
