chore: import upstream snapshot with attribution

scripts/codestyle/convert/example/qwen2/__init__.py

@@ -0,0 +1,20 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# Copyright 2024 The Qwen Team 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.
+
+from .configuration import *
+from .modeling import *
+from .modeling_pp import *
+from .tokenizer import *
+from .tokenizer_fast import *

scripts/codestyle/convert/example/qwen2/configuration.py

@@ -0,0 +1,213 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 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
+
+
+__all__ = [
+    "QWEN2_PRETRAINED_INIT_CONFIGURATION",
+    "Qwen2Config",
+    "QWEN2_PRETRAINED_RESOURCE_FILES_MAP",
+]
+QWEN2_PRETRAINED_INIT_CONFIGURATION = {
+    # Hypothetical model weights (tiny-random-llama & micro-random-llama) for test only
+    "__internal_testing__/micro-random-llama": {
+        "architectures": ["LlamaForCausalLM"],
+        "hidden_size": 64,
+        "initializer_range": 0.02,
+        "intermediate_size": 1000,
+        "max_position_embeddings": 2048,
+        "model_type": "llama",
+        "num_attention_heads": 8,
+        "num_hidden_layers": 1,
+        "rms_norm_eps": 1e-06,
+        "vocab_size": 32000,
+        "bos_token_id": 1,
+        "eos_token_id": 2,
+        "pad_token_id": 0,
+    },
+    "__internal_testing__/tiny-random-llama": {
+        "architectures": ["LlamaForCausalLM"],
+        "hidden_size": 768,
+        "initializer_range": 0.02,
+        "intermediate_size": 11008,
+        "max_position_embeddings": 2048,
+        "model_type": "llama",
+        "num_attention_heads": 8,
+        "num_hidden_layers": 2,
+        "rms_norm_eps": 1e-06,
+        "vocab_size": 32000,
+        "bos_token_id": 1,
+        "eos_token_id": 2,
+        "pad_token_id": 0,
+    },
+}
+
+# Hypothetical model weights (tiny-random-llama) for test only
+QWEN2_PRETRAINED_RESOURCE_FILES_MAP = {
+    "model_state": {
+        "__internal_testing__/micro-random-llama": "https://bj.bcebos.com/paddlenlp/models/community/__internal_testing__/micro-random-llama/model_state.pdparams",
+        "__internal_testing__/tiny-random-llama": "https://bj.bcebos.com/paddlenlp/models/community/__internal_testing__/tiny-random-llama/model_state.pdparams",
+    },
+}
+
+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,
+        # seq_length=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        pad_token_id=151643,
+        bos_token_id=151643,
+        eos_token_id=151643,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        use_flash_attention_for_generation=False,
+        alibi=False,
+        use_last_token_for_generation=False,
+        attention_bias=True,
+        attention_dropout=0.0,
+        rope_scaling_factor=1.0,
+        rope_scaling_type=None,
+        dpo_config=None,
+        use_fused_head_and_loss_fn=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        # self.seq_length = seq_length
+        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_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        self.use_cache = use_cache
+        self.rope_scaling_factor = rope_scaling_factor
+        self.rope_scaling_type = rope_scaling_type
+
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.dpo_config = dpo_config
+        self.use_flash_attention_for_generation = use_flash_attention_for_generation
+        self.alibi = alibi
+        self.use_fused_head_and_loss_fn = use_fused_head_and_loss_fn
+        self.use_last_token_for_generation = use_last_token_for_generation
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

scripts/codestyle/convert/example/qwen2/modeling_pp.py

@@ -0,0 +1,364 @@
+# Copyright (c) 2023 PaddlePaddle Authors. 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.
+
+
+from typing import OrderedDict
+
+import paddle
+import paddle.distributed.fleet as fleet
+import paddle.nn as nn
+from paddle.distributed.fleet.meta_parallel import (
+    LayerDesc,
+    PipelineLayer,
+    SharedLayerDesc,
+)
+from paddle.distributed.fleet.recompute.recompute import recompute
+
+from paddlenlp.transformers.refined_recompute import get_skip_recompute_ops
+from paddlenlp.transformers.refined_recompute import recompute as rr_recompute
+
+from ...utils.tools import get_env_device
+from ..dpo_criterion import DPOCriterion
+from ..model_utils import PipelinePretrainedModel
+from .modeling import (
+    Qwen2Config,
+    Qwen2DecoderLayer,
+    Qwen2LMHead,
+    Qwen2Model,
+    Qwen2PretrainedModel,
+    Qwen2PretrainingCriterion,
+    Qwen2RMSNorm,
+)
+
+__all__ = [
+    "Qwen2ForCausalLMPipe",
+]
+
+
+def parse_args(args):
+    if isinstance(args, tuple):
+        if len(args) == 4:
+            hidden_states, attention_mask, attn_mask_startend_row_indices, position_ids = args
+        elif len(args) == 3:
+            hidden_states, attention_mask, attn_mask_startend_row_indices = args
+            position_ids = None
+        elif len(args) == 2:
+            hidden_states, attention_mask = args
+            attn_mask_startend_row_indices, position_ids = None, None
+    else:
+        hidden_states = args
+        attention_mask, attn_mask_startend_row_indices, position_ids = None, None, None
+
+    if position_ids is not None:
+        position_ids.stop_gradient = True
+
+    if attention_mask is not None:
+        attention_mask.stop_gradient = True
+
+    if attn_mask_startend_row_indices is not None:
+        attn_mask_startend_row_indices.stop_gradient = True
+
+    return hidden_states, attention_mask, attn_mask_startend_row_indices, position_ids
+
+
+def return_args(hidden_states, attention_mask=None, attn_mask_startend_row_indices=None, position_ids=None):
+    ret = (hidden_states,)
+
+    if attention_mask is not None:
+        ret += (attention_mask.clone(),)
+    if attn_mask_startend_row_indices is not None:
+        ret += (attn_mask_startend_row_indices.clone(),)
+    if position_ids is not None:
+        ret += (position_ids.clone(),)
+    if len(ret) == 1:
+        ret = ret[0]
+
+    return ret
+
+
+def get_attr(layer, name):
+    if getattr(layer, name, None) is not None:
+        return getattr(layer, name, None)
+    else:
+        return get_attr(layer._layer, name)
+
+
+class Qwen2EmbeddingPipe(nn.Layer):
+    """Extends QWenEmbeddings to forward attention_mask through the pipeline."""
+
+    def __init__(self, config: Qwen2Config):
+        super(Qwen2EmbeddingPipe, self).__init__()
+        self.config = config
+        self.sequence_parallel = config.sequence_parallel
+        self.hidden_size = config.hidden_size
+        if config.tensor_parallel_degree > 1 and config.vocab_size % config.tensor_parallel_degree == 0:
+            self.embed_tokens = fleet.meta_parallel.VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                weight_attr=paddle.ParamAttr(initializer=nn.initializer.XavierNormal()),
+            )
+        else:
+            self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+
+    @property
+    def embedding_weight(self):
+        return get_attr(self.embed_tokens, "weight")
+
+    def forward(self, args):
+        """_summary_
+
+        Args:
+            input (_type_): _description_
+
+        Returns:
+            _type_: _description_
+        """
+        input_ids, attention_mask, attn_mask_startend_row_indices, position_ids = parse_args(args)
+        input_embeds = self.embed_tokens(input_ids)
+        if self.config.sequence_parallel:
+            from paddlenlp.transformers import ScatterOp
+
+            # [bs, seq_len, num_head * head_dim] -> [bs * seq_len, num_head * head_dim]
+            bs, seq_len, hidden_size = input_embeds.shape
+            input_embeds = paddle.reshape_(input_embeds, [bs * seq_len, hidden_size])
+            # [seq_len * bs / n, num_head * head_dim] (n is mp parallelism)
+            input_embeds = ScatterOp.apply(input_embeds)
+
+        batch_size, seq_length = input_ids.shape
+
+        if attention_mask is not None:
+            assert (
+                attn_mask_startend_row_indices is None
+            ), "attention_mask and attn_mask_startend_row_indices can not be set at same time"
+
+            attention_mask = Qwen2Model._prepare_decoder_attention_mask(
+                attention_mask, (batch_size, seq_length), 0, input_embeds.dtype
+            )
+            attention_mask.stop_gradient = True
+            if get_env_device() == "npu":
+                attention_mask = attention_mask.astype("bool")
+        elif get_env_device() == "npu":
+            attention_mask = paddle.tril(paddle.ones((seq_length, seq_length), dtype="bool"))
+            attention_mask.stop_gradient = True
+
+        return return_args(input_embeds, attention_mask, attn_mask_startend_row_indices, position_ids)
+
+
+class Qwen2DecoderLayerPipe(Qwen2DecoderLayer):
+    def forward(self, args):
+        hidden_states, attention_mask, attn_mask_startend_row_indices, position_ids = parse_args(args)
+
+        has_gradient = not hidden_states.stop_gradient
+
+        if attention_mask is not None and attention_mask.dtype == paddle.int32:
+            attention_mask, attn_mask_startend_row_indices, position_ids = (
+                None,
+                attention_mask,
+                attn_mask_startend_row_indices,
+            )
+        elif attention_mask is not None and attention_mask.dtype == paddle.int64:
+            attention_mask, attn_mask_startend_row_indices, position_ids = None, None, attention_mask
+        elif attn_mask_startend_row_indices is not None and attn_mask_startend_row_indices.dtype == paddle.int64:
+            attn_mask_startend_row_indices, position_ids = None, attn_mask_startend_row_indices
+
+        if self.enable_recompute and self.config.recompute_granularity == "full" and has_gradient:
+            recompute_fn = rr_recompute if any(self.skip_recompute_ops.values()) else recompute
+            if attention_mask is not None or attn_mask_startend_row_indices is not None:
+                hidden_states = recompute_fn(
+                    super().forward,
+                    hidden_states,
+                    position_ids=position_ids,
+                    attention_mask=attention_mask,
+                    attn_mask_startend_row_indices=attn_mask_startend_row_indices,
+                    use_reentrant=False,
+                )
+            else:
+                # for pretrain
+                hidden_states = recompute_fn(
+                    super().forward,
+                    hidden_states,
+                    position_ids=position_ids,
+                    attn_mask_startend_row_indices=attn_mask_startend_row_indices,
+                    use_reentrant=self.config.recompute_use_reentrant,
+                )
+        else:
+            hidden_states = super().forward(
+                hidden_states,
+                position_ids=position_ids,
+                attention_mask=attention_mask,
+                attn_mask_startend_row_indices=attn_mask_startend_row_indices,
+            )
+
+        return return_args(hidden_states, attention_mask, attn_mask_startend_row_indices, position_ids)
+
+
+class Qwen2RMSNormPipe(nn.Layer):
+    def __init__(self, config):
+        super().__init__()
+        self.norm = Qwen2RMSNorm(config)
+
+    def forward(self, args):
+        hidden_states, attention_mask, attn_mask_startend_row_indices, position_ids = parse_args(args)
+        return self.norm(hidden_states)
+
+
+class Qwen2LMHeadPipe(Qwen2LMHead):
+    def __init__(self, config, transpose_y=False):
+        super(Qwen2LMHeadPipe, self).__init__(config, transpose_y=transpose_y)
+
+    @property
+    def embedding_weight(self):
+        return get_attr(self, "weight")
+
+
+class Qwen2ForCausalLMPipe(PipelinePretrainedModel, PipelineLayer):
+    """QWenForPretraining adapted for pipeline parallelism.
+
+    The largest change is flattening the QWenModel class so we can express it as a
+    sequence of layers including embedding, transformer layers, and output.
+    """
+
+    config_class = Qwen2Config
+
+    _get_tensor_parallel_mappings = Qwen2PretrainedModel._get_tensor_parallel_mappings
+    _init_weights = Qwen2PretrainedModel._init_weights
+    _keys_to_ignore_on_load_unexpected = Qwen2PretrainedModel._keys_to_ignore_on_load_unexpected
+    _get_model_flops = Qwen2PretrainedModel._get_model_flops
+    _get_hardware_flops = Qwen2PretrainedModel._get_hardware_flops
+
+    _tied_weights_keys = ["lm_head.weight"]
+
+    # DONOT Add base_model_prefix !!!!
+
+    @classmethod
+    def _prepare_pipeline_inputs_func(cls, inputs):
+
+        first_stage_keys = ["input_ids", "attention_mask", "attn_mask_startend_row_indices", "position_ids"]
+        last_stage_keys = ["labels"]
+
+        def get_expected_keys(inputs, keys):
+            ret = tuple([inputs.pop(k) if k in inputs else None for k in keys])
+            if len(ret) == 1:
+                ret = ret[0]
+            return ret
+
+        if type(inputs) is dict or type(inputs) is OrderedDict:
+            return [
+                get_expected_keys(inputs, first_stage_keys),
+                get_expected_keys(inputs, last_stage_keys),
+            ]
+
+        keys = list(inputs[0].keys())
+        inputs_batch = {key: [data.pop(key) for data in inputs] for key in keys}
+        return [
+            get_expected_keys(inputs_batch, first_stage_keys),
+            get_expected_keys(inputs_batch, last_stage_keys),
+        ]
+
+    def __init__(self, config: Qwen2Config):
+        self.config = config
+
+        # Note that we will actually perform a recompute only if both enable_recompute and layerwise_recompute are set to True
+        # Enable_recompute defaults to False and is controlled by Trainer
+        self.enable_recompute = False
+        self.recompute_granularity = self.config.recompute_granularity
+        self.pp_recompute_interval = self.config.pp_recompute_interval
+        self.no_recompute_layers = config.no_recompute_layers if config.no_recompute_layers is not None else []
+        if self.recompute_granularity == "full":
+            assert len(self.no_recompute_layers) == 0, "for pp with full recompute, no_recompute_layers is not support"
+
+        virtual_pp_degree = getattr(self.config, "virtual_pp_degree", 1)
+
+        def get_hcg():
+            return fleet.get_hybrid_communicate_group()
+
+        hcg = get_hcg()
+        tensor_parallel_degree = max(hcg.get_model_parallel_world_size(), 1)
+        tensor_parallel_rank = max(hcg.get_model_parallel_rank(), 0)
+
+        # TODO: fix tensor_parallel_degree rewrite in here
+        config.tensor_parallel_degree = tensor_parallel_degree
+        config.tensor_parallel_rank = tensor_parallel_rank
+
+        if config.tie_word_embeddings:
+            self.add_sequential_layer(
+                SharedLayerDesc(
+                    "qwen2_shared_weight", Qwen2EmbeddingPipe, shared_weight_attr="embedding_weight", config=config
+                ),
+                "qwen2",
+            )
+        else:
+            self.add_sequential_layer(LayerDesc(Qwen2EmbeddingPipe, config=config), "qwen2")
+
+        for i in range(config.num_hidden_layers):
+            self.add_sequential_layer(
+                LayerDesc(
+                    Qwen2DecoderLayerPipe,
+                    config=config,
+                    layerwise_recompute=i not in self.no_recompute_layers,
+                    skip_recompute_ops=get_skip_recompute_ops(config, i),
+                ),
+                f"qwen2.layers.{i}",
+            )
+        self.add_sequential_layer(LayerDesc(Qwen2RMSNormPipe, config=config), "qwen2")
+
+        if config.tie_word_embeddings:
+            self.add_sequential_layer(
+                SharedLayerDesc(
+                    "qwen2_shared_weight",
+                    Qwen2LMHeadPipe,
+                    shared_weight_attr="embedding_weight",
+                    config=config,
+                    **{"transpose_y": True},
+                ),
+                "lm_head",
+            )
+        else:
+            self.add_sequential_layer(LayerDesc(Qwen2LMHeadPipe, config=config), "lm_head")
+
+        recompute_interval = 0
+        if self.enable_recompute and self.recompute_granularity == "full":
+            assert self.config.pp_recompute_interval <= config.num_hidden_layers // (
+                virtual_pp_degree * get_hcg().topology().get_dim_size("pipe")
+            ), "pp recompute interval should smaller than num layers of each pp chunk"
+            recompute_interval = self.config.pp_recompute_interval
+
+        seg_method = "layer:Qwen2DecoderLayer"
+        if config.num_hidden_layers % get_hcg().topology().get_dim_size("pipe") != 0:
+            seg_method = "uniform"
+
+        PipelineLayer.__init__(
+            self,
+            layers=self.get_sequential_layers(),
+            loss_fn=self.get_loss_fn(config),
+            topology=get_hcg().topology(),
+            seg_method=seg_method,
+            recompute_interval=recompute_interval,
+            recompute_ctx={
+                "mp_group": get_hcg().get_model_parallel_group(),
+                "offload": False,
+                "partition": False,
+            },
+            num_virtual_pipeline_stages=virtual_pp_degree,
+        )
+        # You should call init here, since there is a  diamond inheritance problem
+        self.apply(self._init_weights)
+        # DON'T init PipelinePretrainedModel
+        # PipelinePretrainedModel.__init__(self.super(), config=config)
+
+    def get_loss_fn(self, config):
+        if config.dpo_config is not None:
+            return DPOCriterion(config, use_infohub=True)
+        else:
+            return Qwen2PretrainingCriterion(config)

scripts/codestyle/convert/example/qwen2/tokenizer.py

@@ -0,0 +1,448 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 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 __future__ import annotations
+
+import json
+import os
+import unicodedata
+from functools import lru_cache
+from typing import Any, Dict, List, Optional, Tuple
+
+import regex as re
+
+from ...utils.log import logger
+from .. import AddedToken, PretrainedTokenizer
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+__all__ = ["Qwen2Tokenizer"]
+
+MAX_MODEL_INPUT_SIZES = {"__internal_testing__/tiny-random-qwen2": 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()
+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))
+
+
+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 GPT2Tokenizer, 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.
+    """
+
+    resource_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask", "attn_mask_startend_row_indices"]
+    max_model_input_sizes = MAX_MODEL_INPUT_SIZES
+
+    pretrained_resource_files_map = {
+        "vocab_file": {
+            "__internal_testing__/tiny-random-qwen2": "https://bj.bcebos.com/paddlenlp/models/community/qwen2/vocab.json",
+        },
+    }
+
+    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,
+    ):
+
+        if unk_token is None:
+            logger.info("The `unk_token` parameter needs to be defined: we use `eos_token` by default.")
+            unk_token = eos_token
+
+        # 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 i, line in enumerate(merges_handle):
+                line = line.strip()
+                if (i == 0 and line.startswith("#version:")) or not line:
+                    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)
+
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    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
+
+    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
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.added_tokens_encoder.get(token, len(self.encoder)))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index, self.added_tokens_decoder.get(index, self.unk_token))
+
+    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,
+        )
+
+    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)
+
+    def _encode_chat_inputs(
+        self,
+        conversations: List[List[str, str]],
+        context_data: Dict[str, Any] = {},
+        system: str = None,
+        add_generation_prompt=True,
+    ):
+        result = {}
+
+        # Some template do not support system msg, so we need to check it first.
+        if system:
+            try:
+                self.chat_template.render(messages={"role": "system", "content": system})
+            except Exception as e:
+                raise ValueError("System is not supported in this tokenizer.", e)
+
+        # convert list msg to role dict msg
+        conversation_dict = []
+        origin_msg = []
+        for round in conversations:
+            round_role = [
+                {"role": "user", "content": round[0]},
+                {"role": "assistant", "content": round[1]},
+            ]
+            origin_msg.extend(round_role)
+            conversation_dict.append(round_role)
+
+        # Get system string in ChatTemplate
+        # ChatTemplate contains three parts: system, user, and assistant.
+        # However, the system string cannot be obtained directly with the chat_template.render() function.
+        # Thus, three steps are needed to extract the system string.
+        # Step 1: Obtain the combined system and user string in the first round.
+        # Step 2: Obtain the special system string.
+        # Step 3: Obtain the special combined system and user string in the first round.
+        # Then, user string = (special system and user string) - (special system string)
+        # And, system string = (initial system and user string) - (user string)
+
+        assert len(conversation_dict) > 0, "conversations is empty"
+
+        def replace_first_occurrence(original_string, to_find, to_replace):
+            index = original_string.find(to_find)
+            if index == -1:  # to_find not found in original_string
+                return original_string
+            else:
+                return original_string[:index] + to_replace + original_string[index + len(to_find) :]
+
+        if system:
+            system_str = self.chat_template.render([system])
+        else:
+            # get system and user str
+            round0_str = self.chat_template.render(
+                messages=conversation_dict[0][:1], add_generation_prompt=False, **self.special_tokens_map
+            )
+            # get special system str
+            round0_only_system_str = self.chat_template.render(
+                messages=[{"role": "system", "content": ""}], add_generation_prompt=False, **self.special_tokens_map
+            )
+            # get special system and user str
+            round0_system_user_str = self.chat_template.render(
+                messages=[{"role": "system", "content": ""}] + conversation_dict[0][:1],
+                add_generation_prompt=False,
+                **self.special_tokens_map,
+            )
+
+            # get user str = {special system and user str} - {special system str}
+            user_str = replace_first_occurrence(round0_system_user_str, round0_only_system_str, "")
+            # get system str = { system and user str} - {user str}
+            system_str = round0_str.replace(user_str, "")
+
+        no_ans = []
+        ans = []
+        for conv in conversation_dict:
+            roundi = [system] + conv if system else conv
+            roundi_str = self.chat_template.render(
+                messages=roundi, add_generation_prompt=False, **self.special_tokens_map
+            )
+
+            roundi_no_ans = [system] + [conv[0]] if system else [conv[0]]
+            roundi_no_ans_str = self.chat_template.render(
+                messages=roundi_no_ans, add_generation_prompt=add_generation_prompt, **self.special_tokens_map
+            )
+
+            roundi_ans_str = roundi_str[len(roundi_no_ans_str) :]
+            ans.append(roundi_ans_str)
+
+            roundi_no_ans_no_system_str = replace_first_occurrence(roundi_no_ans_str, system_str, "")
+            assert (
+                roundi_no_ans_str == system_str + roundi_no_ans_no_system_str
+            ), f"the src string contains system str: {system_str}"
+            no_ans.append(roundi_no_ans_no_system_str)
+
+        # the first round is special, we need to add system_str
+        no_ans[0] = system_str + no_ans[0]
+
+        conversation_ids = []
+        for i in range(len(no_ans)):
+            conversation_ids.append(
+                self.batch_encode(
+                    [no_ans[i], ans[i]],
+                    add_special_tokens=False,
+                    padding=False,
+                )["input_ids"]
+            )
+
+        result["conversations"] = conversation_ids
+        return result

scripts/codestyle/convert/example/qwen2/tokenizer_fast.py

@@ -0,0 +1,131 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+# 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 ..tokenizer_utils import AddedToken
+from ..tokenizer_utils_fast import PretrainedTokenizerFast
+from .tokenizer import Qwen2Tokenizer
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_file": "tokenizer.json",
+}
+
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+
+class Qwen2TokenizerFast(PretrainedTokenizerFast):
+    """
+    Construct a "fast" Qwen2 tokenizer (backed by PaddleNLP's *tokenizers* library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    Same with GPT2Tokenizer, 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
+    resource_files_names = VOCAB_FILES_NAMES
+    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=vocab_file,
+            merges_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)