vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/transformers_utils/configs/dbrx.py

@@ -0,0 +1,280 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# yapf: disable
+# ruff: noqa: E501
+# coding=utf-8
+# Copied from
+# https://huggingface.co/databricks/dbrx-base/blob/main/configuration_dbrx.py
+"""Dbrx configuration."""
+
+from typing import Any, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+DBRX_PRETRAINED_CONFIG_ARCHIVE_MAP = {} # type: ignore
+
+
+class DbrxAttentionConfig(PretrainedConfig):
+    """Configuration class for Dbrx Attention.
+
+    [`DbrxAttention`] class. It is used to instantiate attention layers
+    according to the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        attn_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the attention layers.
+        clip_qkv (`float`, *optional*, defaults to None):
+            If not `None`, clip the queries, keys, and values in the attention layer to this value.
+        kv_n_heads (Optional[int]): For grouped_query_attention only, allow user to specify number of kv heads.
+        rope_theta (float): The base frequency for rope.
+    """
+
+    def __init__(
+        self,
+        attn_pdrop: float = 0,
+        clip_qkv: Optional[float] = None,
+        kv_n_heads: int = 1,
+        rope_theta: float = 10000.0,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.attn_pdrop = attn_pdrop
+        self.clip_qkv = clip_qkv
+        self.kv_n_heads = kv_n_heads
+        self.rope_theta = rope_theta
+
+        for k in ["model_type"]:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f"Found unknown {kwargs=}")
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: str, **kwargs: Any
+    ) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        if config_dict.get("model_type") == "dbrx":
+            config_dict = config_dict["attn_config"]
+
+        if (
+            "model_type" in config_dict
+            and hasattr(cls, "model_type")
+            and config_dict["model_type"] != cls.model_type
+        ):
+            logger.warning(
+                "You are using a model of type %s to instantiate a model of "
+                "type %s. This is not supported for all configurations of "
+                "models and can yield errors.",
+                config_dict["model_type"], cls.model_type)
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxFFNConfig(PretrainedConfig):
+    """Configuration class for Dbrx FFN.
+
+    [`DbrxFFN`] class. It is used to instantiate feedforward layers according to
+    the specified arguments, defining the layers architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        ffn_act_fn (dict, optional): A dict specifying activation function for the FFN.
+            The dict should have a key 'name' with the value being the name of
+            the activation function along with any additional keyword arguments.
+        ffn_hidden_size (int, optional): The hidden size of the feedforward network.
+        moe_num_experts (int, optional): The number of experts in the mixture of experts layer.
+        moe_top_k (int, optional): The number of experts to use in the mixture of experts layer.
+        moe_jitter_eps (float, optional): The jitter epsilon for the mixture of experts layer.
+        moe_loss_weight (float, optional): The loss weight for the mixture of experts layer.
+        moe_normalize_expert_weights (float, optional): The normalization factor for the expert weights.
+        uniform_expert_assignment (bool, optional): Whether to use uniform expert assignment.
+            This should only be used for benchmarking purposes.
+    """
+
+    def __init__(
+        self,
+        ffn_act_fn: Optional[dict] = None,
+        ffn_hidden_size: int = 3584,
+        moe_num_experts: int = 4,
+        moe_top_k: int = 1,
+        moe_jitter_eps: Optional[float] = None,
+        moe_loss_weight: float = 0.01,
+        moe_normalize_expert_weights: Optional[float] = 1,
+        uniform_expert_assignment: bool = False,
+        **kwargs: Any,
+    ):
+        super().__init__()
+        if ffn_act_fn is None:
+            ffn_act_fn = {"name": "silu"}
+        self.ffn_act_fn = ffn_act_fn
+        self.ffn_hidden_size = ffn_hidden_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.moe_jitter_eps = moe_jitter_eps
+        self.moe_loss_weight = moe_loss_weight
+        self.moe_normalize_expert_weights = moe_normalize_expert_weights
+        self.uniform_expert_assignment = uniform_expert_assignment
+
+        for k in ["model_type"]:
+            if k in kwargs:
+                kwargs.pop(k)
+        if len(kwargs) != 0:
+            raise ValueError(f"Found unknown {kwargs=}")
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: str, **kwargs: Any
+    ) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        if config_dict.get("model_type") == "dbrx":
+            config_dict = config_dict["ffn_config"]
+
+        if (
+            "model_type" in config_dict
+            and hasattr(cls, "model_type")
+            and config_dict["model_type"] != cls.model_type
+        ):
+            logger.warning(
+                "You are using a model of type %s to instantiate a model of "
+                "type %s. This is not supported for all "
+                "configurations of models and can yield errors.", config_dict["model_type"], cls.model_type)
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class DbrxConfig(PretrainedConfig):
+    """Configuration class for Dbrx.
+
+    [`DbrxModel`]. It is used to instantiate a Dbrx model according to the
+    specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        d_model (`int`, *optional*, defaults to 6144):
+            Dimensionality of the embeddings and hidden states.
+        n_heads (`int`, *optional*, defaults to 48):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        n_layers (`int`, *optional*, defaults to 40):
+            Number of hidden layers in the Transformer encoder.
+        max_seq_len (`int`, *optional*, defaults to 32768):
+            The maximum sequence length of the model.
+        vocab_size (`int`, *optional*, defaults to 100352):
+            Vocabulary size of the Dbrx model. Defines the maximum number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`DbrxModel`].
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability applied to the attention output before combining with residual.
+        emb_pdrop (`float`, *optional*, defaults to 0.0):
+            The dropout probability for the embedding layer.
+        attn_config (`dict`, *optional*):
+            A dictionary used to configure the model's attention module.
+        ffn_config (`dict`, *optional*):
+            A dictionary used to configure the model's FFN module.
+        use_cache (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        output_router_logits (`bool`, *optional*, defaults to `False`):
+            Whether or not the router logits should be returned by the model. Enabling this will also allow the model to output the auxiliary loss.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+
+    Example:
+    ```python
+    >>> from transformers import DbrxConfig, DbrxModel
+
+    >>> # Initializing a Dbrx configuration
+    >>> configuration = DbrxConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = DbrxModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "dbrx"
+    attribute_map = {
+        "num_attention_heads": "n_heads",
+        "hidden_size": "d_model",
+        "num_hidden_layers": "n_layers",
+        "max_position_embeddings": "max_seq_len",
+    }
+
+    def __init__(
+        self,
+        d_model: int = 2048,
+        n_heads: int = 16,
+        n_layers: int = 24,
+        max_seq_len: int = 2048,
+        vocab_size: int = 32000,
+        resid_pdrop: float = 0.0,
+        emb_pdrop: float = 0.0,
+        attn_config: Optional[DbrxAttentionConfig] = None,
+        ffn_config: Optional[DbrxFFNConfig] = None,
+        use_cache: bool = True,
+        initializer_range: float = 0.02,
+        output_router_logits: bool = False,
+        router_aux_loss_coef: float = 0.05,
+        **kwargs: Any,
+    ):
+        if attn_config is None:
+            self.attn_config = DbrxAttentionConfig()
+        elif isinstance(attn_config, dict):
+            self.attn_config = DbrxAttentionConfig(**attn_config)
+        else:
+            self.attn_config = attn_config
+
+        if ffn_config is None:
+            self.ffn_config = DbrxFFNConfig()
+        elif isinstance(ffn_config, dict):
+            self.ffn_config = DbrxFFNConfig(**ffn_config)
+        else:
+            self.ffn_config = ffn_config
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.resid_pdrop = resid_pdrop
+        self.emb_pdrop = emb_pdrop
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+        self.output_router_logits = output_router_logits
+        self.router_aux_loss_coef = router_aux_loss_coef
+
+        tie_word_embeddings = kwargs.pop("tie_word_embeddings", False)
+        if tie_word_embeddings:
+            raise ValueError(
+                "tie_word_embeddings is not supported for Dbrx models."
+            )
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

vllm/transformers_utils/configs/deepseek_vl2.py

@@ -0,0 +1,216 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py#L115-L268
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class VisionEncoderConfig(PretrainedConfig):
+    model_type: str = "vision"
+
+    model_name: str = "vit_so400m_patch14_siglip_384.webli"
+    image_size: int = 384
+    patch_size: int = 16
+    width: int = 1024
+    layers: int = 24
+    heads: int = 16
+    mlp_ratio: int = 4
+    global_pool: str = "map"
+    ignore_head: bool = True
+    class_token: bool = False
+    num_classes: int = 0
+    use_checkpoint: bool = False
+    weight_init: str = "skip"
+    deterministic: bool = False
+    num_recomputing_layers: int = 0
+
+    def __init__(self,
+                 model_name: str = "vit_so400m_patch14_siglip_384.webli",
+                 image_size: int = 384,
+                 patch_size: int = 16,
+                 width: int = 1024,
+                 layers: int = 24,
+                 heads: int = 16,
+                 mlp_ratio: int = 4,
+                 global_pool: str = "map",
+                 ignore_head: bool = True,
+                 class_token: bool = False,
+                 num_classes: int = 0,
+                 use_checkpoint: bool = False,
+                 **kwargs):
+        self.model_name = model_name
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.width = width
+        self.layers = layers
+        self.heads = heads
+        self.mlp_ratio = mlp_ratio
+        self.global_pool = global_pool
+        self.ignore_head = ignore_head
+        self.class_token = class_token
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+
+        super().__init__(**kwargs)
+
+
+class MlpProjectorConfig(PretrainedConfig):
+    model_type = "mlp_projector"
+    projector_type: str = "downsample_mlp_gelu"
+    input_dim: int = 1152
+    n_embed: int = 2048
+    depth: int = 2
+    mlp_ratio: int = 1
+    downsample_ratio: int = 2
+    token_pooling: bool = False
+
+    def __init__(self,
+                 projector_type: str = "downsample_mlp_gelu",
+                 input_dim: int = 1152,
+                 n_embed: int = 2048,
+                 depth: int = 2,
+                 mlp_ratio: int = 1,
+                 downsample_ratio: int = 2,
+                 **kwargs):
+        self.projector_type = projector_type
+        self.input_dim = input_dim
+        self.n_embed = n_embed
+        self.depth = depth
+        self.mlp_ratio = mlp_ratio
+        self.downsample_ratio = downsample_ratio
+
+        super().__init__(**kwargs)
+
+
+class DeepseekV2Config(PretrainedConfig):
+
+    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,
+        )
+
+
+class DeepseekVLV2Config(PretrainedConfig):
+    model_type = "deepseek_vl_v2"
+    vision_config: VisionEncoderConfig
+    projector_config: MlpProjectorConfig
+
+    tile_tag: str = "2D"
+    global_view_pos: str = "head"
+    candidate_resolutions: tuple[tuple[int, int]] = ((384, 384), )
+
+    def __init__(self,
+                 tile_tag: str = "tile_tag",
+                 global_view_pos: str = "head",
+                 candidate_resolutions: tuple[tuple[int,
+                                                    int]] = ((384, 384), ),
+                 **kwargs):
+        super().__init__(**kwargs)
+
+        vision_config = kwargs.get("vision_config", {})
+        self.vision_config = VisionEncoderConfig(**vision_config)
+
+        projector_config = kwargs.get("projector_config", {})
+        self.projector_config = MlpProjectorConfig(**projector_config)
+
+        language_config = kwargs.get("language_config", {})
+        self.text_config = DeepseekV2Config(**language_config)
+
+        self.tile_tag = tile_tag
+        self.global_view_pos = global_view_pos
+        self.candidate_resolutions = candidate_resolutions
+        self.vocab_size = self.text_config.vocab_size

vllm/transformers_utils/configs/eagle.py

@@ -0,0 +1,85 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+from typing import Optional, Union
+
+from transformers import AutoConfig, PretrainedConfig
+
+import vllm.envs as envs
+from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
+
+
+class EAGLEConfig(PretrainedConfig):
+    model_type = "eagle"
+
+    def __init__(self,
+                 model: Union[PretrainedConfig, dict, None] = None,
+                 truncated_vocab_size: Optional[int] = None,
+                 method: Optional[str] = 'eagle',
+                 **kwargs):
+
+        model_config: Union[PretrainedConfig, DeepseekV2Config, None]
+        if isinstance(model, dict):
+            archs = model.get("architectures", [])
+            target_archs = ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]
+            if any(target_arch in archs for target_arch in target_archs):
+                # AutoConfig does not support DeepSeek MoE models yet
+                model_config = DeepseekV2Config(**model)
+            else:
+                model_config = AutoConfig.for_model(**model)
+        else:
+            model_config = model
+
+        for k, v in kwargs.items():
+            if k != "architectures" and k != "model_type" and hasattr(
+                    model_config, k):
+                setattr(model_config, k, v)
+
+        self.model = model_config
+
+        if self.model is None:
+            self.truncated_vocab_size = None
+        else:
+            self.truncated_vocab_size = self.model.vocab_size if \
+                truncated_vocab_size is None else truncated_vocab_size
+
+        if not envs.VLLM_USE_V1:
+            kwargs["architectures"] = ["EAGLEModel"]
+        else:
+            # Eagle model name should follow naming convention of
+            # LlamaForCausalLM -> EagleLlamaForCausalLM
+            if method == "eagle":
+                assert self.model is not None, \
+                    "model should not be None when method is eagle"
+                kwargs["architectures"] = [
+                    f"Eagle{arch}" if not arch.startswith("Eagle") \
+                        else arch for arch in self.model.architectures
+                ]
+            elif method == "eagle3":
+                assert self.model is not None, \
+                    "model should not be None when method is eagle3"
+                kwargs["architectures"] = [
+                    f"Eagle3{arch}" if not arch.startswith("Eagle3") \
+                        else arch for arch in self.model.architectures
+                ]
+            else:
+                raise ValueError(f"Invalid method {method}. \
+                    Supported methods are eagle and eagle3.")
+
+        super().__init__(**kwargs)
+
+        if self.model is not None:
+            for k, v in self.model.to_dict().items():
+                if k not in kwargs:
+                    setattr(self, k, v)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        **kwargs,
+    ) -> "EAGLEConfig":
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs)
+        return cls.from_dict(config_dict, **kwargs)