Add qwen2 (#28436)

* add config, modeling, and tokenization

* add auto and init

* update readme

* update readme

* update team name

* fixup

* fixup

* update config

* update code style

* update for fixup

* update for fixup

* update for fixup

* update for testing

* update for testing

* fix bug for config and tokenization

* fix bug for bos token

* not doctest

* debug tokenizer

* not doctest

* debug tokenization

* debug init for tokenizer

* fix style

* update init

* delete if in token auto

* add tokenizer doc

* add tokenizer in init

* Update dummy_tokenizers_objects.py

* update

* update

* debug

* Update tokenization_qwen2.py

* debug

* Update convert_slow_tokenizer.py

* add copies

* add copied from and make style

* update files map

* update test

* fix style

* fix merge reading and update tests

* fix tests

* fix tests

* fix style

* debug a variable in readme

* Update src/transformers/models/qwen2/configuration_qwen2.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* update test and copied from

* fix style

* update qwen2 tokenization  and tests

* Update tokenization_qwen2.py

* delete the copied from after property

* fix style

* update tests

* update tests

* add copied from

* fix bugs

* update doc

* add warning for sliding window attention

* update qwen2 tokenization

* fix style

* Update src/transformers/models/qwen2/modeling_qwen2.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix tokenizer fast

---------
Co-authored-by: Ren Xuancheng <jklj077@users.noreply.github.com>
Co-authored-by: renxuancheng.rxc <renxuancheng.rxc@alibaba-inc.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/qwen2/configuration_qwen2.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Qwen2 model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "Qwen/Qwen2-7B-beta": "https://huggingface.co/Qwen/Qwen2-7B-beta/resolve/main/config.json",
+}
+
+
+class Qwen2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
+    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen2Model, Qwen2Config
+
+    >>> # Initializing a Qwen2 style configuration
+    >>> configuration = Qwen2Config()
+
+    >>> # Initializing a model from the Qwen2-7B style configuration
+    >>> model = Qwen2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

src/transformers/models/qwen2/tokenization_qwen2.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for Qwen2."""
+
+import json
+import os
+import unicodedata
+from functools import lru_cache
+from typing import Optional, Tuple
+
+import regex as re
+
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/vocab.json"},
+    "merges_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/merges.txt"},
+}
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+PRETOKENIZE_REGEX = r"""(?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+"""
+
+
+@lru_cache()
+# Copied from transformers.models.gpt2.tokenization_gpt2.bytes_to_unicode
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+# Copied from transformers.models.gpt2.tokenization_gpt2.get_pairs
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class Qwen2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a Qwen2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Same with GPT2Tokenzier, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import Qwen2Tokenizer
+
+    >>> tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-tokenizer")
+    >>> tokenizer("Hello world")["input_ids"]
+    [9707, 1879]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [21927, 1879]
+    ```
+    This is expected.
+
+    You should not use GPT2Tokenizer instead, because of the different pretokenization rules.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        errors (`str`, *optional*, defaults to `"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See
+            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*):
+            The beginning of sequence token. Not applicable for this tokenizer.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should cleanup the spaces that were added when splitting the input text during the
+            tokenization process. Not applicable to this tokenizer, since tokenization does not add spaces.
+        split_special_tokens (`bool`, *optional*, defaults to `False`):
+            Whether or not the special tokens should be split during the tokenization process. The default behavior is
+            to not split special tokens. This means that if `<|endoftext|>` is the `eos_token`, then `tokenizer.tokenize("<|endoftext|>") =
+            ['<|endoftext|>`]. Otherwise, if `split_special_tokens=True`, then `tokenizer.tokenize("<|endoftext|>")` will be give `['<',
+            '|', 'endo', 'ft', 'ext', '|', '>']`. This argument is only supported for `slow` tokenizers for the moment.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = MAX_MODEL_INPUT_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token=None,
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",
+        clean_up_tokenization_spaces=False,
+        split_special_tokens=False,
+        **kwargs,
+    ):
+        # Qwen vocab does not contain control tokens; added tokens need to be special
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        bpe_merges = []
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            for line in merges_handle:
+                line = line.strip()
+                if not line or line.startswith("#"):
+                    continue
+                bpe_merges.append(tuple(line.split()))
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        # NOTE: the cache can grow without bound and will get really large for long running processes
+        # (esp. for texts of language that do not use space between word, e.g. Chinese); technically
+        # not a memory leak but appears as one.
+        # GPT2Tokenizer has the same problem, so let's be consistent.
+        self.cache = {}
+
+        self.pat = re.compile(PRETOKENIZE_REGEX)
+
+        if kwargs.get("add_prefix_space", False):
+            logger.warning_once(
+                f"{self.__class__.__name} does not support `add_prefix_space`, setting it to True has no effect."
+            )
+
+        super().__init__(
+            errors=errors,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            unk_token=unk_token,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            split_special_tokens=split_special_tokens,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.get_vocab
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.bpe
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._tokenize
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.convert_tokens_to_string
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def decode(
+        self,
+        token_ids,
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = False,
+        spaces_between_special_tokens: bool = False,
+        **kwargs,
+    ) -> str:
+        # `spaces_between_special_tokens` defaults to True for _decode in slow tokenizers
+        # and cannot be configured elsewhere, but it should default to False for Qwen2Tokenizer
+        return super().decode(
+            token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, **kwargs):
+        text = unicodedata.normalize("NFC", text)
+        return (text, kwargs)

src/transformers/models/qwen2/tokenization_qwen2_fast.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for Qwen2."""
+
+from typing import Optional, Tuple
+
+from ...tokenization_utils import AddedToken
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_qwen2 import Qwen2Tokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_file": "tokenizer.json",
+}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/vocab.json"},
+    "merges_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/merges.txt"},
+    "tokenizer_file": {
+        "qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/tokenizer.json"
+    },
+}
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+
+class Qwen2TokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" Qwen2 tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    Same with GPT2Tokenzier, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import Qwen2TokenizerFast
+
+    >>> tokenizer = Qwen2TokenizerFast.from_pretrained("Qwen/Qwen-tokenizer")
+    >>> tokenizer("Hello world")["input_ids"]
+    [9707, 1879]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [21927, 1879]
+    ```
+    This is expected.
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`, *optional*):
+            Path to the vocabulary file.
+        merges_file (`str`, *optional*):
+            Path to the merges file.
+        tokenizer_file (`str`, *optional*):
+            Path to [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
+            contains everything needed to load the tokenizer.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead. Not applicable to this tokenizer.
+        bos_token (`str`, *optional*):
+            The beginning of sequence token. Not applicable for this tokenizer.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = MAX_MODEL_INPUT_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = Qwen2Tokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        unk_token="<|endoftext|>",
+        bos_token=None,
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",
+        **kwargs,
+    ):
+        # We need to at least pass vocab_file and merges_file to base class
+        # in case a slow tokenizer needs to be initialized; other can be
+        # configured through files.
+        # following GPT2TokenizerFast, also adding unk_token, bos_token, and eos_token
+
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+
+        super().__init__(
+            vocab_file,
+            merges_file,
+            tokenizer_file=tokenizer_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2_fast.GPT2TokenizerFast.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)

src/transformers/utils/dummy_pt_objects.py

@@ -6715,6 +6715,34 @@ def load_tf_weights_in_qdqbert(*args, **kwargs):
     requires_backends(load_tf_weights_in_qdqbert, ["torch"])
 
 
+class Qwen2ForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2ForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2Model(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2PreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class RagModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/transformers/utils/dummy_tokenizers_objects.py

@@ -331,6 +331,13 @@ class PegasusTokenizerFast(metaclass=DummyObject):
         requires_backends(self, ["tokenizers"])
 
 
+class Qwen2TokenizerFast(metaclass=DummyObject):
+    _backends = ["tokenizers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tokenizers"])
+
+
 class RealmTokenizerFast(metaclass=DummyObject):
     _backends = ["tokenizers"]
 

tests/models/qwen2/test_tokenization_qwen2.py

+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import json
+import os
+import unittest
+
+from transformers import AddedToken, Qwen2Tokenizer, Qwen2TokenizerFast
+from transformers.models.qwen2.tokenization_qwen2 import VOCAB_FILES_NAMES, bytes_to_unicode
+from transformers.testing_utils import require_tokenizers, slow
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+@require_tokenizers
+class Qwen2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    tokenizer_class = Qwen2Tokenizer
+    rust_tokenizer_class = Qwen2TokenizerFast
+    test_slow_tokenizer = True
+    test_rust_tokenizer = True
+    space_between_special_tokens = False
+    from_pretrained_kwargs = None
+    test_seq2seq = False
+
+    def setUp(self):
+        super().setUp()
+
+        # this make sure the vocabuary is complete at the byte level.
+        vocab = list(bytes_to_unicode().values())
+        # the vocabulary, note:
+        # - `"\u0120n"`, `"\u0120lowest"`, `"\u0120newer"`, and `"\u0120wider"` are ineffective, because there are
+        #   not in the merges.
+        # - `"01"` is ineffective, because the merge is ineffective due to pretokenization.
+        vocab.extend(
+            [
+                "\u0120l",
+                "\u0120n",
+                "\u0120lo",
+                "\u0120low",
+                "er",
+                "\u0120lowest",
+                "\u0120newer",
+                "\u0120wider",
+                "01",
+                ";}",
+                ";}\u010a",
+                "\u00cf\u0135",
+            ]
+        )
+
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+
+        # note: `"0 1"` is in the merges, but the pretokenization rules render it ineffective
+        merges = [
+            "#version: 0.2",
+            "\u0120 l",
+            "\u0120l o",
+            "\u0120lo w",
+            "e r",
+            "0 1",
+            "; }",
+            ";} \u010a",
+            "\u00cf \u0135",
+        ]
+
+        self.special_tokens_map = {"eos_token": "<|endoftext|>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+    def get_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return Qwen2Tokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return Qwen2TokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self, tokenizer):
+        # this case should cover
+        # - NFC normalization (code point U+03D3 has different normalization forms under NFC, NFD, NFKC, and NFKD)
+        # - the pretokenization rules (spliting digits and merging symbols with \n\r)
+        input_text = "lower lower newer 010;}\n<|endoftext|>\u03d2\u0301"
+        output_text = "lower lower newer 010;}\n<|endoftext|>\u03d3"
+        return input_text, output_text
+
+    def test_python_full_tokenizer(self):
+        tokenizer = self.get_tokenizer()
+        sequence, _ = self.get_input_output_texts(tokenizer)
+        bpe_tokens = [
+            "l",
+            "o",
+            "w",
+            "er",
+            "\u0120low",
+            "er",
+            "\u0120",
+            "n",
+            "e",
+            "w",
+            "er",
+            "\u0120",
+            "0",
+            "1",
+            "0",
+            ";}\u010a",
+            "<|endoftext|>",
+            "\u00cf\u0135",
+        ]
+        tokens = tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens
+        input_bpe_tokens = [75, 78, 86, 260, 259, 260, 220, 77, 68, 86, 260, 220, 15, 16, 15, 266, 268, 267]
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    @unittest.skip("We disable the test of pretokenization as it is not reversible.")
+    def test_pretokenized_inputs(self):
+        # the test case in parent class uses str.split to "pretokenize",
+        # which eats the whitespaces, which, in turn, is not reversible.
+        # the results, by nature, should be different.
+        pass
+
+    def test_nfc_normalization(self):
+        # per https://unicode.org/faq/normalization.html, there are three characters whose normalization forms
+        # under NFC, NFD, NFKC, and NFKD are all different
+        # using these, we can make sure only NFC is applied
+        input_string = "\u03d2\u0301\u03d2\u0308\u017f\u0307"  # the NFD form
+        output_string = "\u03d3\u03d4\u1e9b"  # the NFC form
+
+        if self.test_slow_tokenizer:
+            tokenizer = self.get_tokenizer()
+            tokenizer_output_string, _ = tokenizer.prepare_for_tokenization(input_string)
+            self.assertEqual(tokenizer_output_string, output_string)
+
+        if self.test_rust_tokenizer:
+            tokenizer = self.get_rust_tokenizer()
+            # we can check the class of the normalizer, but it would be okay if Sequence([NFD, NFC]) is used
+            # let's check the output instead
+            tokenizer_output_string = tokenizer.backend_tokenizer.normalizer.normalize_str(input_string)
+            self.assertEqual(tokenizer_output_string, output_string)
+
+    def test_slow_tokenizer_decode_spaces_between_special_tokens_default(self):
+        # Qwen2Tokenzier changes the default `spaces_between_special_tokens` in `decode` to False
+        if not self.test_slow_tokenizer:
+            return
+
+        # tokenizer has a special token: `"<|endfotext|>"` as eos, but it is not `legacy_added_tokens`
+        # special tokens in `spaces_between_special_tokens` means spaces between `legacy_added_tokens`
+        # that would be `"<|im_start|>"` and `"<|im_end|>"` in Qwen/Qwen2 Models
+        token_ids = [259, 260, 268, 269, 26]
+        sequence = " lower<|endoftext|><|im_start|>;"
+        sequence_with_space = " lower<|endoftext|> <|im_start|> ;"
+
+        tokenizer = self.get_tokenizer()
+        # let's add a legacy_added_tokens
+        im_start = AddedToken(
+            "<|im_start|>", single_word=False, lstrip=False, rstrip=False, special=True, normalized=False
+        )
+        tokenizer.add_tokens([im_start])
+
+        # `spaces_between_special_tokens` defaults to False
+        self.assertEqual(tokenizer.decode(token_ids), sequence)
+
+        # but it can be set to True
+        self.assertEqual(tokenizer.decode(token_ids, spaces_between_special_tokens=True), sequence_with_space)
+
+    @slow
+    def test_tokenizer_integration(self):
+        sequences = [
+            "Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
+            "general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
+            "Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained "
+            "models in 100+ languages and deep interoperability between Jax, PyTorch and TensorFlow.",
+            "🤗 Transformers 提供了可以轻松地下载并且训练先进的预训练模型的 API 和工具。使用预训练模型可以减少计算消耗和碳排放，并且节省从头训练所需要的时间和资源。",
+            """```python\ntokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-tokenizer")\n"""
+            """tokenizer("世界，你好！")```""",
+        ]
+
+        expected_encoding = {'input_ids': [[8963, 388, 320, 69514, 3881, 438, 4510, 27414, 32852, 388, 323, 4510, 27414, 21334, 35722, 1455, 529, 8, 5707, 4586, 58238, 77235, 320, 61437, 11, 479, 2828, 12, 17, 11, 11830, 61437, 64, 11, 1599, 10994, 11, 27604, 321, 33, 529, 11, 29881, 6954, 32574, 369, 18448, 11434, 45451, 320, 45, 23236, 8, 323, 18448, 11434, 23470, 320, 30042, 38, 8, 448, 916, 220, 18, 17, 10, 80669, 4119, 304, 220, 16, 15, 15, 10, 15459, 323, 5538, 94130, 2897, 1948, 619, 706, 11, 5355, 51, 21584, 323, 94986, 13], [144834, 80532, 93685, 83744, 34187, 73670, 104261, 29490, 62189, 103937, 104034, 102830, 98841, 104034, 104949, 9370, 5333, 58143, 102011, 1773, 37029, 98841, 104034, 104949, 73670, 101940, 100768, 104997, 33108, 100912, 105054, 90395, 100136, 106831, 45181, 64355, 104034, 113521, 101975, 33108, 85329, 1773, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643], [73594, 12669, 198, 85593, 284, 8979, 37434, 6387, 10442, 35722, 445, 48, 16948, 45274, 16948, 34841, 3135, 1138, 85593, 445, 99489, 3837, 108386, 6313, 899, 73594, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]}  # fmt: off
+
+        self.tokenizer_integration_test_util(
+            expected_encoding=expected_encoding,
+            model_name="Qwen/Qwen-tokenizer",
+            revision="5909c8222473b2c73b0b73fb054552cd4ef6a8eb",
+            sequences=sequences,
+        )

utils/not_doctested.txt

@@ -198,6 +198,7 @@ docs/source/en/model_doc/pop2piano.md
 docs/source/en/model_doc/prophetnet.md
 docs/source/en/model_doc/pvt.md
 docs/source/en/model_doc/qdqbert.md
+docs/source/en/model_doc/qwen2.md
 docs/source/en/model_doc/rag.md
 docs/source/en/model_doc/realm.md
 docs/source/en/model_doc/reformer.md
@@ -745,6 +746,10 @@ src/transformers/models/pvt/image_processing_pvt.py
 src/transformers/models/pvt/modeling_pvt.py
 src/transformers/models/qdqbert/configuration_qdqbert.py
 src/transformers/models/qdqbert/modeling_qdqbert.py
+src/transformers/models/qwen2/configuration_qwen2.py
+src/transformers/models/qwen2/modeling_qwen2.py
+src/transformers/models/qwen2/tokenization_qwen2.py
+src/transformers/models/qwen2/tokenization_qwen2_fast.py
 src/transformers/models/rag/configuration_rag.py
 src/transformers/models/rag/modeling_rag.py
 src/transformers/models/rag/modeling_tf_rag.py