xinference 1.3.1.post1__py3-none-any.whl → 1.4.1__py3-none-any.whl

xinference/model/llm/vllm/distributed_executor.py

@@ -0,0 +1,314 @@
+# Copyright 2022-2025 XProbe Inc.
+#
+# 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 asyncio
+import logging
+import os
+from functools import partial
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union
+
+import xoscar as xo
+from vllm.executor.executor_base import DistributedExecutorBase
+from vllm.utils import _run_task_with_lock, get_distributed_init_method
+from vllm.worker.worker_base import WorkerWrapperBase
+from xoscar.utils import get_next_port
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+    from vllm.model_executor.layers.sampler import SamplerOutput
+    from vllm.sequence import ExecuteModelRequest
+
+logger = logging.getLogger(__name__)
+
+
+class WorkerActor(xo.StatelessActor):
+    def __init__(self, vllm_config: "VllmConfig", rpc_rank: int = 0, **kwargs):
+        super().__init__(**kwargs)
+        self._worker = WorkerWrapperBase(vllm_config, rpc_rank=rpc_rank)
+
+    async def __post_create__(self):
+        try:
+            # Change process title for model
+            import setproctitle
+
+            setproctitle.setproctitle(f"Xinf vLLM worker: {self._worker.rpc_rank}")
+        except ImportError:
+            pass
+
+    def __getattr__(self, item):
+        return getattr(self._worker, item)
+
+    @classmethod
+    def gen_uid(cls, rank):
+        return f"VllmWorker_{rank}"
+
+    def execute_method(self, method: Union[str, Callable], *args, **kwargs):
+        logger.debug(
+            "Calling method %s in vllm worker %s, args: %s, kwargs: %s",
+            method,
+            self.uid,
+            args,
+            kwargs,
+        )
+        if isinstance(method, str):
+            return getattr(self._worker, method)(*args, **kwargs)
+        else:
+            return method(self._worker, *args, **kwargs)
+
+
+class WorkerWrapper:
+    def __init__(
+        self,
+        loop: asyncio.AbstractEventLoop,
+        worker_actor_ref: xo.ActorRefType[WorkerActor],
+    ):
+        self._loop = loop
+        self._worker_actor_ref = worker_actor_ref
+
+    def execute_method(self, method: Union[str, Callable], *args, **kwargs):
+        coro = self._worker_actor_ref.execute_method(method, *args, **kwargs)
+        return asyncio.run_coroutine_threadsafe(coro, self._loop)
+
+    async def execute_method_async(self, method: Union[str, Callable], *args, **kwargs):
+        return await self._worker_actor_ref.execute_method(method, *args, **kwargs)
+
+    def kill(self):
+        coro = xo.kill_actor(self._worker_actor_ref)
+        return asyncio.run_coroutine_threadsafe(coro, self._loop)
+
+
+class XinferenceDistributedExecutor(DistributedExecutorBase):
+    """Xoscar based distributed executor"""
+
+    use_ray: bool = False
+    _loop: asyncio.AbstractEventLoop
+    _pool_addresses: List[str]
+    _n_worker: int
+
+    def __init__(
+        self,
+        vllm_config: "VllmConfig",
+        pool_addresses: List[str],
+        n_worker: int,
+        loop: asyncio.AbstractEventLoop,
+        *args,
+        **kwargs,
+    ):
+        self._pool_addresses = pool_addresses
+        self._loop = loop
+        self._n_worker = n_worker
+        super().__init__(vllm_config, *args, **kwargs)
+
+    def _init_executor(self) -> None:
+        # Create the parallel GPU workers.
+        world_size = self.parallel_config.world_size
+        tensor_parallel_size = self.parallel_config.tensor_parallel_size
+
+        self.driver_worker: Optional[WorkerActor] = None
+        # The remaining workers are Xoscar actors
+        self.workers: List[WorkerWrapper] = []
+
+        assert (
+            self._pool_addresses and len(self._pool_addresses) == world_size
+        ), f"Pool addresses(#{len(self._pool_addresses or [])} must be equal to worldsize(#{world_size})"
+
+        futures = []
+        for rank in range(world_size):
+            coro = xo.create_actor(
+                WorkerActor,
+                self.vllm_config,
+                rpc_rank=rank,
+                address=self._pool_addresses[rank],
+                uid=WorkerActor.gen_uid(rank),
+            )
+            futures.append(asyncio.run_coroutine_threadsafe(coro, self._loop))
+        refs = [fut.result() for fut in futures]
+        self.workers = [WorkerWrapper(self._loop, ref) for ref in refs[1:]]
+        self.driver_worker = WorkerActor(self.vllm_config, rpc_rank=0)
+
+        def driver_execute_method(*args, **kwargs):
+            func = partial(self.driver_worker.execute_method, *args, **kwargs)
+            return self._loop.run_in_executor(None, func)
+
+        self.driver_exec_method = driver_execute_method
+
+        # Set environment variables for the driver and workers.
+        all_args_to_update_environment_variables: List[Dict[str, str]] = [
+            dict() for _ in range(world_size)
+        ]
+
+        for args in all_args_to_update_environment_variables:
+            # some carry-over env vars from the driver
+            # TODO: refactor platform-specific env vars
+            for name in [
+                "VLLM_ATTENTION_BACKEND",
+                "TPU_CHIPS_PER_HOST_BOUNDS",
+                "TPU_HOST_BOUNDS",
+                "VLLM_USE_V1",
+                "VLLM_TRACE_FUNCTION",
+            ]:
+                if name in os.environ:
+                    args[name] = os.environ[name]
+
+        self._env_vars_for_all_workers = all_args_to_update_environment_variables
+
+        self._run_workers(
+            "update_environment_variables", self._env_vars_for_all_workers
+        )
+
+        all_kwargs = []
+        distributed_init_method = get_distributed_init_method(
+            self._pool_addresses[0].split(":", 1)[0], get_next_port()
+        )
+        for rank in range(world_size):
+            local_rank = rank % (world_size // self._n_worker)
+            kwargs = dict(
+                vllm_config=self.vllm_config,
+                local_rank=local_rank,
+                rank=rank,
+                distributed_init_method=distributed_init_method,
+                is_driver_worker=not self.parallel_config
+                or (rank % tensor_parallel_size == 0),
+            )
+            all_kwargs.append(kwargs)
+        self._run_workers("init_worker", all_kwargs)
+        self._run_workers("init_device")
+        self._run_workers(
+            "load_model",
+            max_concurrent_workers=self.parallel_config.max_parallel_loading_workers,
+        )
+
+        # This is the list of workers that are rank 0 of each TP group EXCEPT
+        # global rank 0. These are the workers that will broadcast to the
+        # rest of the workers.
+        self.tp_driver_workers: List[WorkerWrapper] = []
+        # This is the list of workers that are not drivers and not the first
+        # worker in a TP group. These are the workers that will be
+        # broadcasted to.
+        self.non_driver_workers: List[WorkerWrapper] = []
+
+        # Enforce rank order for correct rank to return final output.
+        for index, worker in enumerate(self.workers):
+            # The driver worker is rank 0 and not in self.workers.
+            rank = index + 1
+            if rank % self.parallel_config.tensor_parallel_size == 0:
+                self.tp_driver_workers.append(worker)
+            else:
+                self.non_driver_workers.append(worker)
+
+        self.pp_locks: Optional[List[asyncio.Lock]] = None
+
+    def _run_workers(
+        self,
+        method: Union[str, Callable],
+        *args,
+        async_run_tensor_parallel_workers_only: bool = False,
+        max_concurrent_workers: Optional[int] = None,
+        **kwargs,
+    ) -> Any:
+        if max_concurrent_workers:
+            raise NotImplementedError("max_concurrent_workers is not supported yet.")
+
+        workers = self.workers
+        if async_run_tensor_parallel_workers_only:
+            workers = self.non_driver_workers
+        worker_outputs = [
+            worker.execute_method(method, *args, **kwargs) for worker in workers
+        ]
+
+        if async_run_tensor_parallel_workers_only:
+            return worker_outputs
+
+        driver_worker_outputs = [
+            self.driver_worker.execute_method(method, *args, **kwargs)  # type: ignore
+        ]
+        return driver_worker_outputs + [output.result() for output in worker_outputs]
+
+    def _wait_for_tasks_completion(self, parallel_worker_tasks: Any) -> None:
+        """Wait for futures returned from _run_workers() with
+        async_run_remote_workers_only to complete."""
+        for result in parallel_worker_tasks:
+            result.get()
+
+    def check_health(self) -> None:
+        # Assume that the workers are healthy.
+        # TODO: check the health by checking if the workers all alive
+        return
+
+    def shutdown(self) -> None:
+        try:
+            futs = [worker.kill() for worker in self.workers]
+            _ = [fut.result() for fut in futs]
+        except (RuntimeError, ConnectionError):
+            # event loop closed already, ignore
+            pass
+
+    def __del__(self):
+        return self.shutdown()
+
+    def _driver_execute_model(
+        self, execute_model_req: Optional["ExecuteModelRequest"]
+    ) -> Optional[List["SamplerOutput"]]:
+        return self.driver_worker.execute_method("execute_model", execute_model_req)  # type: ignore
+
+    async def _driver_execute_model_async(
+        self,
+        execute_model_req: Optional["ExecuteModelRequest"] = None,
+    ) -> List["SamplerOutput"]:
+        if not self.tp_driver_workers:
+            return await self.driver_exec_method("execute_model", execute_model_req)
+
+        if self.pp_locks is None:
+            # This locks each pipeline parallel stage so multiple virtual
+            # engines can't execute on the same stage at the same time
+            # We create the locks here to avoid creating them in the constructor
+            # which uses a different asyncio loop.
+            self.pp_locks = [
+                asyncio.Lock()
+                for _ in range(self.parallel_config.pipeline_parallel_size)
+            ]
+
+        tasks = [
+            asyncio.create_task(
+                _run_task_with_lock(
+                    self.driver_exec_method,
+                    self.pp_locks[0],
+                    "execute_model",
+                    execute_model_req,
+                )
+            )
+        ]
+        for pp_rank, driver_worker in enumerate(self.tp_driver_workers, start=1):
+            tasks.append(
+                asyncio.create_task(
+                    _run_task_with_lock(
+                        driver_worker.execute_method_async,
+                        self.pp_locks[pp_rank],
+                        "execute_model",
+                        execute_model_req,
+                    )
+                )
+            )
+
+        results = await asyncio.gather(*tasks)
+
+        # Only the last PP stage has the final results.
+        return results[-1]
+
+    async def _start_worker_execution_loop(self):
+        coros = [
+            worker.execute_method_async("start_worker_execution_loop")
+            for worker in self.non_driver_workers
+        ]
+        return await asyncio.gather(*coros)

xinference/model/rerank/core.py

@@ -106,9 +106,10 @@ def generate_rerank_description(model_spec: RerankModelSpec) -> Dict[str, List[D
     return res
 
 
-class _ModelWrapper:
+class _ModelWrapper(nn.Module):
     def __init__(self, module: nn.Module):
-        self._module = module
+        super().__init__()
+        self.model = module
         self._local_data = threading.local()
 
     @property
@@ -116,18 +117,22 @@ class _ModelWrapper:
         return getattr(self._local_data, "n_tokens", 0)
 
     @n_tokens.setter
-    def n_tokens(self, new_n_tokens):
-        self._local_data.n_tokens = new_n_tokens
+    def n_tokens(self, value):
+        self._local_data.n_tokens = value
 
-    def __getattr__(self, attr):
-        return getattr(self._module, attr)
-
-    def __call__(self, **kwargs):
-        attention_mask = kwargs["attention_mask"]
+    def forward(self, **kwargs):
+        attention_mask = kwargs.get("attention_mask")
         # when batching, the attention mask 1 means there is a token
         # thus we just sum up it to get the total number of tokens
-        self.n_tokens += attention_mask.sum().item()
-        return self._module(**kwargs)
+        if attention_mask is not None:
+            self.n_tokens += attention_mask.sum().item()
+        return self.model(**kwargs)
+
+    def __getattr__(self, attr):
+        try:
+            return super().__getattr__(attr)
+        except AttributeError:
+            return getattr(self.model, attr)
 
 
 class RerankModel:

xinference/thirdparty/deepseek_vl2/__init__.py

@@ -0,0 +1,31 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+
+# check if python version is above 3.10
+import sys
+
+if sys.version_info >= (3, 10):
+    print("Python version is above 3.10, patching the collections module.")
+    # Monkey patch collections
+    import collections
+    import collections.abc
+
+    for type_name in collections.abc.__all__:
+        setattr(collections, type_name, getattr(collections.abc, type_name))

xinference/thirdparty/deepseek_vl2/models/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from .processing_deepseek_vl_v2 import DeepseekVLV2Processor
+from .modeling_deepseek_vl_v2 import DeepseekVLV2ForCausalLM
+
+__all__ = [
+    "DeepseekVLV2Processor",
+    "DeepseekVLV2ForCausalLM",
+]

xinference/thirdparty/deepseek_vl2/models/configuration_deepseek.py

@@ -0,0 +1,210 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+class DeepseekV2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV2Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V2 with multi-latent attention.
+
+    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 102400):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1407):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        n_shared_experts (`int`, *optional*, defaults to None):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to None):
+            Number of routed experts, None means dense model.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
+            Scaling factor or routed experts.
+        topk_method (`str`, *optional*, defaults to `gready`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to None):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to None):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to None):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 0):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to False):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to 'softmax'):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+        seq_aux = (`bool`, *optional*, defaults to True):
+            Whether to compute the auxiliary loss for each individual sample.
+        num_key_value_heads (`int`, *optional*):
+            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
+            `num_attention_heads`.
+        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 2048):
+            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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        use_mla (`bool`, *optional*, defaults to `True`): Use multi-latent attention or multi-head attention. If True,
+            the model will use multi-latent attention, otherwise, it will use multi-head attention.
+
+    ```python
+    >>> from transformers import DeepseekV2Model, DeepseekV2Config
+
+    >>> # Initializing a Deepseek-V2 style configuration
+    >>> configuration = DeepseekV2Config()
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=11008,
+        moe_intermediate_size = 1407,
+        num_hidden_layers=30,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        n_shared_experts = None,
+        n_routed_experts = None,
+        ep_size = 1,
+        routed_scaling_factor = 1.0,
+        kv_lora_rank = 512,
+        q_lora_rank = 1536,
+        qk_rope_head_dim = 64,
+        v_head_dim = 128,
+        qk_nope_head_dim = 128,
+        topk_method = 'gready',
+        n_group = None,
+        topk_group = None,
+        num_experts_per_tok = None,
+        moe_layer_freq = 1,
+        first_k_dense_replace = 0,
+        norm_topk_prob = False,
+        scoring_func = 'softmax',
+        aux_loss_alpha = 0.001,
+        seq_aux = True,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=100000,
+        eos_token_id=100001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        use_mla=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # 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 = float(rms_norm_eps)
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.use_mla = use_mla
+
+        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,
+        )