Fix up a few things

llmfoundry/__init__.py

@@ -23,6 +23,7 @@
     from llmfoundry.models.mpt import (ComposerMPTCausalLM, MPTConfig,
                                        MPTForCausalLM, MPTModel,
                                        MPTPreTrainedModel)
+    from llmfoundry.tokenizers import TiktokenTokenizerWrapper
 
 except ImportError as e:
     try:
@@ -64,6 +65,7 @@
     'build_alibi_bias',
     'optim',
     'utils',
+    'TiktokenTokenizerWrapper'
 ]
 
 __version__ = '0.2.0'

llmfoundry/tokenizers/__init__.py

@@ -0,0 +1,8 @@
+# Copyright 2022 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from llmfoundry.tokenizers.tiktoken import TiktokenTokenizerWrapper
+
+__all__ = [
+    'TiktokenTokenizerWrapper',
+]

llmfoundry/tokenizers/tiktoken.py

@@ -0,0 +1,292 @@
+# Copyright 2022 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import PreTrainedTokenizer
+
+
+class TiktokenTokenizerWrapper(PreTrainedTokenizer):
+    """A thin wrapper around tiktoken to make it compatible with Hugging Face.
+
+    tokenizers.
+
+    See HuggingFace for further documentation on general tokenizer methods.
+    """
+
+    model_input_names = ['input_ids', 'attention_mask']
+
+    def __init__(self,
+                 model_name: Optional[str] = None,
+                 encoding_name: Optional[str] = None,
+                 add_bos_token: bool = False,
+                 unk_token: Optional[str] = '<|endoftext|>',
+                 eos_token: Optional[str] = '<|endoftext|>',
+                 bos_token: Optional[str] = '<|endoftext|>',
+                 mask_token: Optional[str] = "[MASK]",
+                 pad_token: Optional[str] = None,
+                 **kwargs: Dict[str, Any]):
+        """Constructor creates a tiktoken tokenizer to use as the underlying.
+
+        tokenizer.
+
+        Args:
+            model_name (Optional[str], optional): The name of the model to load from tiktoken. Defaults to None.
+                Either model_name or encoding_name must be set, but not both.
+            encoding_name (Optional[str], optional): The name of the encoding to load from tiktoken. Defaults to None.
+                Either model_name or encoding_name must be set, but not both.
+            add_bos_token (bool, optional): Whether to add bos tokens. Defaults to False.
+            unk_token (Optional[str], optional): The unk token. Defaults to '<|endoftext|>'.
+            eos_token (Optional[str], optional): The eos token. Defaults to '<|endoftext|>'.
+            bos_token (Optional[str], optional): The bos token. Defaults to '<|endoftext|>'.
+            pad_token (Optional[str], optional): The pad token. Defaults to None.
+        """
+        try:
+            import tiktoken
+        except:
+            raise ImportError(
+                'You need to install tiktoken to use TiktokenTokenizerWrapper.')
+
+        if model_name is not None and encoding_name is not None:
+            raise ValueError(
+                'You need to specify either model_name or encoding_name, not both.'
+            )
+
+        self.model_name = model_name
+        self.encoding_name = encoding_name
+
+        if self.model_name is not None:
+            self.encoding = tiktoken.encoding_for_model(  # type: ignore (thirdParty)
+                self.model_name)
+        elif self.encoding_name is not None:
+            self.encoding = tiktoken.get_encoding(  # type: ignore (thirdParty)
+                self.encoding_name)
+        else:
+            raise ValueError(
+                'You need to specify either model_name or encoding_name.')
+
+        self.add_bos_token = add_bos_token
+
+        super().__init__(model_name=model_name,
+                         encoding_name=encoding_name,
+                         add_bos_token=add_bos_token,
+                         unk_token=unk_token,
+                         eos_token=eos_token,
+                         bos_token=bos_token,
+                         pad_token=pad_token,
+                         mask_token=mask_token,
+                         **kwargs)
+
+    @property
+    def vocab_size(self) -> int:
+        """Returns vocab size."""
+        return self.encoding.n_vocab
+
+    @property
+    def is_fast(self) -> bool:
+        return False
+
+    def get_vocab(self) -> Dict[str, int]:
+        """Returns vocab as a dict."""
+        vocab = {}
+        for i in range(self.vocab_size):
+            try:
+                # need to try this first, so that we get a proper KeyError,
+                # otherwise it crashes in the rust code
+                _ = self.encoding.decode_single_token_bytes(i)
+                vocab[self.encoding.decode([i])] = i
+            except KeyError:
+                pass
+
+        return vocab
+
+    def _tokenize(self, text: str) -> List[int]:
+        """Returns a tokenized string.
+
+        Note: We have slightly redefined the expected contract between this method and
+        the _convert_token_to_id method. Normally, this method turns a string, into a list of strings,
+        and then the _convert_token_to_id method turns that list of strings into a list of integers.
+        However, not all vocab indices can be decoded into a string, so instead we just return the integers
+        from this function, and have adjusted the _convert_token_to_id method to handle integers as well as strings.
+        The only use of _tokenize that I could find was in this way, so this _should_ be safe.
+        """
+        if not isinstance(text, str):
+            raise ValueError(
+                f'Expected a string input to _tokenize but got {type(text)}.')
+
+        tokens = [t for t in self.encoding.encode(text, allowed_special='all')]
+
+        return tokens
+
+    def _convert_token_to_id(self, token: Union[int, str]) -> int:
+        """Converts a token (str) into an id using the vocab."""
+        if isinstance(token, int):
+            return token
+
+        return self.encoding.encode(token, allowed_special='all')[0]
+
+    def _convert_id_to_token(self, index: int) -> str:
+        """Converts an index (integer) into a token (str) using the vocab."""
+        return self.encoding.decode([index])
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """Converts a sequence of tokens (string) in a single string."""
+        return ''.join(tokens)
+
+    def convert_ids_to_tokens(
+            self,
+            ids: Union[int, List[int]],
+            skip_special_tokens: bool = False) -> Union[str, List[str]]:
+        """Converts a single index or a sequence of indices into a token or a.
+
+        sequence of tokens, using the vocabulary and added tokens.
+
+        Args:
+            ids (`int` or `List[int]`):
+                The token id (or token ids) to convert to tokens.
+            skip_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not to remove special tokens in the decoding.
+
+        Returns:
+            `str` or `List[str]`: The decoded token(s).
+        """
+        if isinstance(ids, int):
+            if ids in self.added_tokens_decoder:
+                return self.added_tokens_decoder[ids]
+
+            return self._convert_id_to_token(ids)
+
+        # current_stream will collect multiple tokens, and then separately add items
+        # for each added token. This is done so that decode works properly with token ids
+        # that cannot be represented naively in utf-8.
+        tokens = []
+        current_stream = []
+        for index in ids:
+            if skip_special_tokens and index in self.all_special_ids:
+                continue
+
+            if index in self.added_tokens_decoder:
+                tokens.append(self.encoding.decode(current_stream))
+                current_stream = []
+                tokens.append(self.added_tokens_decoder[index])
+            else:
+                current_stream.append(index)
+
+        if len(current_stream) > 0:
+            tokens.append(self.encoding.decode(current_stream))
+        return tokens
+
+    def build_inputs_with_special_tokens(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None) -> List[int]:
+        if self.add_bos_token:
+            bos_token_ids = [self.bos_token_id]
+        else:
+            bos_token_ids = []
+
+        output = bos_token_ids + token_ids_0
+
+        if token_ids_1 is None:
+            return output
+
+        return output + bos_token_ids + token_ids_1
+
+    def get_special_tokens_mask(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None,
+            already_has_special_tokens: bool = False) -> List[int]:
+        """Retrieves sequence ids from a token list that has no special tokens.
+
+        Function copied from
+        https://github.com/huggingface/transformers/blob/e3a4bd2bee212a2d0fd9f03b27fe7bfc1debe42d/src/transformers/models/gpt2/tokenization_gpt2.py#L265-L295
+
+        added. This method is called when adding special tokens using the
+        tokenizer `prepare_for_model` or `encode_plus` methods.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0,
+                token_ids_1=token_ids_1,
+                already_has_special_tokens=True)
+
+        if not self.add_bos_token:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0,
+                token_ids_1=token_ids_1,
+                already_has_special_tokens=False)
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0))
+        return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1))
+
+    def create_token_type_ids_from_sequences(
+            self,
+            token_ids_0: List[int],
+            token_ids_1: Optional[List[int]] = None) -> List[int]:
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self,
+                        save_directory: str,
+                        filename_prefix: Optional[str] = None) -> Tuple[str]:
+
+        # ignore the below type to keep the original signature
+        # we are knowingly breaking the signature here, although not 100% certain
+        # it doesn't have side effects
+        # There is some code in huggingface that calls this function to get the vocab files,
+        # but it doesn't seem to access them (or at least checks for their existence
+        # before accessing them)
+        return (None, None)  # type: ignore
+
+    def sanitize_special_tokens(self) -> int:
+        """Make sure that all the special tokens attributes of the tokenizer.
+
+        (`tokenizer.mask_token`, `tokenizer.cls_token`, etc.) are in the
+        vocabulary.
+
+        Add the missing ones to the vocabulary if needed.
+
+        Return:
+            `int`: The number of tokens added in the vocabulary during the operation.
+        """
+        actual_new_tokens = []
+        for token in self.all_special_tokens_extended:
+            encoded = self.encoding.encode(token, allowed_special='all')
+            if len(encoded) > 1:
+                actual_new_tokens.append(token)
+
+        return self.add_tokens(actual_new_tokens, special_tokens=True)
+
+    def construct_logit_tensor(self, logprobs: Dict[str,
+                                                    float]) -> torch.Tensor:
+        """Construct tensor of shape (vocab_size,) mapping words to logprobs.
+
+        Args:
+            logprobs (Dict[str, float]): Dictionary mapping tokens to log probabilities assigned to them by the model.
+        """
+        tensor = torch.tensor([min(logprobs.values()) - 1] * (self.vocab_size))
+        for k in logprobs:
+            encoding = self(k)['input_ids']
+            idx = encoding[0]
+            tensor[idx] = logprobs[k]
+        return tensor
+
+
+TiktokenTokenizerWrapper.register_for_auto_class()

llmfoundry/utils/builders.py

@@ -29,6 +29,7 @@
                                   ScheduledGarbageCollector)
 from llmfoundry.optim import (DecoupledAdaLRLion, DecoupledClipLion,
                               DecoupledLionW, DecoupledLionW_8bit)
+from llmfoundry.tokenizers.tiktoken import TiktokenTokenizerWrapper
 
 
 def build_icl_data_and_gauntlet(
@@ -153,25 +154,25 @@ def build_scheduler(name: str, scheduler_config: Dict[str, Any]):
         raise ValueError(f'Not sure how to build scheduler: {name}')
 
 
-def build_tokenizer(om_tokenizer_config: DictConfig) -> PreTrainedTokenizerBase:
+def build_tokenizer(
+        tokenizer_name: str,
+        tokenizer_kwargs: Dict[str, Any]) -> PreTrainedTokenizerBase:
     os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = '1'
     os.environ['TOKENIZERS_PARALLELISM'] = 'false'
 
-    resolved_om_tokenizer_config = om.to_container(om_tokenizer_config,
-                                                   resolve=True)
-    tokenizer_kwargs = resolved_om_tokenizer_config.get(  # type: ignore
-        'kwargs', {})
-    tokenizer_name = resolved_om_tokenizer_config['name']  # type: ignore
-    tokenizer = AutoTokenizer.from_pretrained(tokenizer_name,
-                                              **tokenizer_kwargs)
+    if tokenizer_name.startswith('tiktoken'):
+        tokenizer = TiktokenTokenizerWrapper(**tokenizer_kwargs)
+    else:
+        tokenizer = AutoTokenizer.from_pretrained(tokenizer_name,
+                                                  **tokenizer_kwargs)
 
-    # HuggingFace does not respect the model_max_length kwarg, and overrides it with
-    # min(kwargs['model_max_length'], original_config['model_max_length']), so we
-    # explicitly set it here
-    tokenizer.model_max_length = tokenizer_kwargs.get(
-        'model_max_length',
-        int(1e30),
-    )
+        # HuggingFace does not respect the model_max_length kwarg, and overrides it with
+        # min(kwargs['model_max_length'], original_config['model_max_length']), so we
+        # explicitly set it here
+        tokenizer.model_max_length = tokenizer_kwargs.get(
+            'model_max_length',
+            int(1e30),
+        )
 
     return tokenizer
 

scripts/train/train.py

@@ -416,7 +416,9 @@ def main(cfg: DictConfig) -> Trainer:
     logged_cfg.update({'fsdp_config': fsdp_config}, merge=True)
 
     # Build tokenizer
-    tokenizer = build_tokenizer(tokenizer_config)
+    tokenizer_name = tokenizer_config['name']
+    tokenizer_kwargs = tokenizer_config.get('kwargs', {})
+    tokenizer = build_tokenizer(tokenizer_name, tokenizer_kwargs)
 
     # Scheduler
     scheduler_name: str = scheduler_config.pop('name')

setup.py

@@ -63,6 +63,7 @@
     'cmake>=3.25.0,<=3.26.3',  # required for triton-pre-mlir below
     # PyPI does not support direct dependencies, so we remove this line before uploading from PyPI
     'triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir_sm90#subdirectory=python',
+    'tiktoken==0.4.0'
 ]
 
 extra_deps = {}