Add Qwen2MoE (#29377)

* add support for qwen2 MoE models

* update docs

* add support for qwen2 MoE models

* update docs

* update model name & test

* update readme

* update class names & readme & model_doc of Qwen2MoE.

* update architecture name

* fix qwen2_moe tests

* use Qwen2Tokenizer instead of Qwen2MoeTokenizer

* update modeling_qwen2_moe.py

* fix model architecture

* fix qwen2_moe tests

* use Qwen2Tokenizer instead of Qwen2MoeTokenizer

* update modeling_qwen2_moe.py

* fix model architecture

* fix style

* fix test when there are sparse and non sparse layers

* fixup

* Update README.md
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fixup

* fixup

* add archive back

* add support for qwen2 MoE models

* update docs

* update model name & test

* update readme

* update class names & readme & model_doc of Qwen2MoE.

* update architecture name

* fix qwen2_moe tests

* use Qwen2Tokenizer instead of Qwen2MoeTokenizer

* update modeling_qwen2_moe.py

* fix model architecture

* fixup

* fix qwen2_moe tests

* use Qwen2Tokenizer instead of Qwen2MoeTokenizer

* fix style

* fix test when there are sparse and non sparse layers

* fixup

* add archive back

* fix integration test

* fixup

---------
Co-authored-by: bozheng-hit <dsoul0621@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/__init__.py

@@ -184,6 +184,7 @@ from . import (
     pvt_v2,
     qdqbert,
     qwen2,
+    qwen2_moe,
     rag,
     realm,
     reformer,

src/transformers/models/auto/configuration_auto.py

@@ -195,6 +195,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("pvt_v2", "PvtV2Config"),
         ("qdqbert", "QDQBertConfig"),
         ("qwen2", "Qwen2Config"),
+        ("qwen2_moe", "Qwen2MoeConfig"),
         ("rag", "RagConfig"),
         ("realm", "RealmConfig"),
         ("reformer", "ReformerConfig"),
@@ -467,6 +468,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("pvt_v2", "PVTv2"),
         ("qdqbert", "QDQBert"),
         ("qwen2", "Qwen2"),
+        ("qwen2_moe", "Qwen2MoE"),
         ("rag", "RAG"),
         ("realm", "REALM"),
         ("reformer", "Reformer"),

src/transformers/models/auto/modeling_auto.py

@@ -182,6 +182,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("pvt_v2", "PvtV2Model"),
         ("qdqbert", "QDQBertModel"),
         ("qwen2", "Qwen2Model"),
+        ("qwen2_moe", "Qwen2MoeModel"),
         ("reformer", "ReformerModel"),
         ("regnet", "RegNetModel"),
         ("rembert", "RemBertModel"),
@@ -467,6 +468,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("prophetnet", "ProphetNetForCausalLM"),
         ("qdqbert", "QDQBertLMHeadModel"),
         ("qwen2", "Qwen2ForCausalLM"),
+        ("qwen2_moe", "Qwen2MoeForCausalLM"),
         ("reformer", "ReformerModelWithLMHead"),
         ("rembert", "RemBertForCausalLM"),
         ("roberta", "RobertaForCausalLM"),
@@ -871,6 +873,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("plbart", "PLBartForSequenceClassification"),
         ("qdqbert", "QDQBertForSequenceClassification"),
         ("qwen2", "Qwen2ForSequenceClassification"),
+        ("qwen2_moe", "Qwen2MoeForSequenceClassification"),
         ("reformer", "ReformerForSequenceClassification"),
         ("rembert", "RemBertForSequenceClassification"),
         ("roberta", "RobertaForSequenceClassification"),

src/transformers/models/auto/tokenization_auto.py

@@ -354,6 +354,13 @@ else:
                     "Qwen2TokenizerFast" if is_tokenizers_available() else None,
                 ),
             ),
+            (
+                "qwen2_moe",
+                (
+                    "Qwen2Tokenizer",
+                    "Qwen2TokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             ("rag", ("RagTokenizer", None)),
             ("realm", ("RealmTokenizer", "RealmTokenizerFast" if is_tokenizers_available() else None)),
             (

src/transformers/models/qwen2_moe/__init__.py

+# 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 typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_qwen2_moe": ["QWEN2MOE_PRETRAINED_CONFIG_ARCHIVE_MAP", "Qwen2MoeConfig"],
+}
+
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_qwen2_moe"] = [
+        "Qwen2MoeForCausalLM",
+        "Qwen2MoeModel",
+        "Qwen2MoePreTrainedModel",
+        "Qwen2MoeForSequenceClassification",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_qwen2_moe import QWEN2MOE_PRETRAINED_CONFIG_ARCHIVE_MAP, Qwen2MoeConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_qwen2_moe import (
+            Qwen2MoeForCausalLM,
+            Qwen2MoeForSequenceClassification,
+            Qwen2MoeModel,
+            Qwen2MoePreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/qwen2_moe/configuration_qwen2_moe.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.
+""" Qwen2MoE model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+QWEN2MOE_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "Qwen/Qwen1.5-MoE-A2.7B": "https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B/resolve/main/config.json",
+}
+
+
+class Qwen2MoeConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2MoeModel`]. It is used to instantiate a
+    Qwen2MoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen1.5-MoE-A2.7B" [Qwen/Qwen1.5-MoE-A2.7B"](https://huggingface.co/Qwen/Qwen1.5-MoE-A2.7B").
+
+    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 Qwen2MoE model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2MoeModel`]
+        hidden_size (`int`, *optional*, defaults to 2048):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 5632):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 16):
+            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.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer.
+        moe_intermediate_size (`int`, *optional*, defaults to 1408):
+            Intermediate size of the routed expert.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 5632):
+            Intermediate size of the shared expert.
+        num_experts_per_tok (`int`, *optional*, defaults to 4):
+            Number of selected experts.
+        num_experts (`int`, *optional*, defaults to 60):
+            Number of routed experts.
+        norm_topk_prob (`bool`, *optional*, defaults to `False`):
+            Whether to normalize the topk probabilities.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabeling this will also
+            allow the model to output the auxiliary loss, including load balancing loss and router z-loss.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+    ```python
+    >>> from transformers import Qwen2MoeModel, Qwen2MoeConfig
+
+    >>> # Initializing a Qwen2MoE style configuration
+    >>> configuration = Qwen2MoeConfig()
+
+    >>> # Initializing a model from the Qwen1.5-MoE-A2.7B" style configuration
+    >>> model = Qwen2MoeModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2_moe"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=2048,
+        intermediate_size=5632,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        num_key_value_heads=16,
+        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,
+        decoder_sparse_step=1,
+        moe_intermediate_size=1408,
+        shared_expert_intermediate_size=5632,
+        num_experts_per_tok=4,
+        num_experts=60,
+        norm_topk_prob=False,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        **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
+
+        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
+
+        # MoE arguments
+        self.decoder_sparse_step = decoder_sparse_step
+        self.moe_intermediate_size = moe_intermediate_size
+        self.shared_expert_intermediate_size = shared_expert_intermediate_size
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_experts = num_experts
+        self.norm_topk_prob = norm_topk_prob
+        self.output_router_logits = output_router_logits
+        self.router_aux_loss_coef = router_aux_loss_coef
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -6939,6 +6939,34 @@ class Qwen2PreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class Qwen2MoeForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2MoeForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2MoeModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Qwen2MoePreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class RagModel(metaclass=DummyObject):
     _backends = ["torch"]
 

utils/not_doctested.txt

@@ -201,6 +201,7 @@ 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/qwen2_moe.md
 docs/source/en/model_doc/rag.md
 docs/source/en/model_doc/realm.md
 docs/source/en/model_doc/reformer.md
@@ -759,6 +760,8 @@ 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/qwen2_moe/configuration_qwen2_moe.py
+src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
 src/transformers/models/rag/configuration_rag.py
 src/transformers/models/rag/modeling_rag.py
 src/transformers/models/rag/modeling_tf_rag.py