init

vllm/transformers_utils/configs/__init__.py

@@ -0,0 +1,63 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Model configs may be defined in this directory for the following reasons:
+
+- There is no configuration file defined by HF Hub or Transformers library.
+- There is a need to override the existing config to support vLLM.
+"""
+
+from vllm.transformers_utils.configs.chatglm import ChatGLMConfig
+from vllm.transformers_utils.configs.deepseek_v3 import DeepseekV3Config
+from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
+from vllm.transformers_utils.configs.dotsocr import DotsOCRConfig
+from vllm.transformers_utils.configs.eagle import EAGLEConfig
+# RWConfig is for the original tiiuae/falcon-40b(-instruct) and
+# tiiuae/falcon-7b(-instruct) models. Newer Falcon models will use the
+# `FalconConfig` class from the official HuggingFace transformers library.
+from vllm.transformers_utils.configs.falcon import RWConfig
+from vllm.transformers_utils.configs.jais import JAISConfig
+from vllm.transformers_utils.configs.kimi_vl import KimiVLConfig
+from vllm.transformers_utils.configs.medusa import MedusaConfig
+from vllm.transformers_utils.configs.midashenglm import MiDashengLMConfig
+from vllm.transformers_utils.configs.mlp_speculator import MLPSpeculatorConfig
+from vllm.transformers_utils.configs.moonvit import MoonViTConfig
+from vllm.transformers_utils.configs.nemotron import NemotronConfig
+from vllm.transformers_utils.configs.nemotron_h import NemotronHConfig
+from vllm.transformers_utils.configs.nemotron_vl import Nemotron_Nano_VL_Config
+from vllm.transformers_utils.configs.olmo3 import Olmo3Config
+from vllm.transformers_utils.configs.ovis import OvisConfig
+from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig
+from vllm.transformers_utils.configs.radio import RadioConfig
+from vllm.transformers_utils.configs.speculators.base import SpeculatorsConfig
+from vllm.transformers_utils.configs.step3_vl import (Step3TextConfig,
+                                                      Step3VisionEncoderConfig,
+                                                      Step3VLConfig)
+from vllm.transformers_utils.configs.ultravox import UltravoxConfig
+
+__all__ = [
+    "ChatGLMConfig",
+    "DeepseekVLV2Config",
+    "DeepseekV3Config",
+    "DotsOCRConfig",
+    "EAGLEConfig",
+    "RWConfig",
+    "JAISConfig",
+    "MedusaConfig",
+    "MiDashengLMConfig",
+    "MLPSpeculatorConfig",
+    "MoonViTConfig",
+    "KimiVLConfig",
+    "NemotronConfig",
+    "NemotronHConfig",
+    "Nemotron_Nano_VL_Config",
+    "Olmo3Config",
+    "OvisConfig",
+    "RadioConfig",
+    "SpeculatorsConfig",
+    "UltravoxConfig",
+    "Step3VLConfig",
+    "Step3VisionEncoderConfig",
+    "Step3TextConfig",
+    "Qwen3NextConfig",
+]

vllm/transformers_utils/configs/arctic.py

@@ -0,0 +1,207 @@
+# 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/Snowflake/snowflake-arctic-instruct/blob/main/configuration_arctic.py
+""" Arctic model configuration"""
+
+from dataclasses import asdict, dataclass
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+ARCTIC_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "arctic": "https://huggingface.co/Snowflake/snowflake-arctic-instruct/tree/main/config.json",
+}
+
+
+@dataclass
+class ArcticLoRAConfig:
+    lora_r: int = 64
+    lora_alpha: float = 16
+    shard_base_weights: bool = False
+
+
+@dataclass
+class ArcticQuantizationConfig:
+    q_bits: int = 8
+    rounding: str = "nearest"
+    mantissa_bits: int = 3
+    group_size: int = 128
+
+
+class ArcticConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ArcticModel`]. It is used to instantiate an
+    Arctic 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 #TODO(rsamdani): add what model has the default config..
+
+
+    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 32000):
+            Vocabulary size of the Arctic model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`ArcticModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 14336):
+            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 8):
+            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 `8`.
+        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 `4096*32`):
+            The maximum sequence length that this model might ever be used with. Arctic's sliding window attention
+            allows sequence of up to 4096*32 tokens.
+        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-05):
+            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*):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        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 1000000.0):
+            The base period of the RoPE embeddings.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If not specified, will default to `4096`.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        num_experts_per_tok (`int`, *optional*, defaults to 2):
+            The number of experts to root per-token, can be also interpreted as the `top-p` routing
+            parameter
+        num_local_experts (`int`, *optional*, defaults to 8):
+            Number of experts per Sparse MLP layer.
+        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
+            The aux loss factor for the total loss.
+
+    ```python
+    >>> from transformers import ArcticModel, ArcticConfig
+
+    >>> # Initializing a Arctic 7B style configuration TODO(rsamdani): verify which model does the default configuration correspond to.
+    >>> configuration = ArcticConfig()
+
+    >>> # Initializing a model from the Arctic 7B style configuration
+    >>> model = ArcticModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "arctic"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=4096,
+        intermediate_size=14336,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        rope_theta=1e6,
+        sliding_window=None,
+        attention_dropout=0.0,
+        num_experts_per_tok=1,
+        num_local_experts=8,
+        router_aux_loss_coef=0.001,
+        moe_layer_frequency=2,
+        parallel_attn_mlp_res=False,
+        moe_train_capacity_factor=1,
+        moe_eval_capacity_factor=1,
+        enable_expert_tensor_parallelism=False,
+        moe_min_capacity=0,
+        moe_token_dropping=True,
+        quantization=None,
+        **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.sliding_window = sliding_window
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_local_experts = num_local_experts
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.moe_layer_frequency = moe_layer_frequency
+        self.moe_train_capacity_factor = moe_train_capacity_factor
+        self.moe_eval_capacity_factor = moe_eval_capacity_factor
+        self.enable_expert_tensor_parallelism = enable_expert_tensor_parallelism
+        self.moe_min_capacity = moe_min_capacity
+        self.moe_token_dropping = moe_token_dropping
+        self.parallel_attn_mlp_res = parallel_attn_mlp_res
+        if isinstance(quantization, dict):
+            self.quantization = ArcticQuantizationConfig(**quantization)
+        else:
+            self.quantization = quantization
+
+        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,
+        )
+
+    @classmethod
+    def from_dict(cls, config_dict: dict[str, Any], **kwargs) -> "ArcticConfig":
+        result = super().from_dict(config_dict, **kwargs)
+        config = result[0] if isinstance(result, tuple) else result
+        if isinstance(config.quantization, dict):
+            config.quantization = ArcticQuantizationConfig(**config.quantization)
+        return result
+
+    def to_dict(self) -> dict[str, Any]:
+        ret = super().to_dict()
+        if isinstance(ret["quantization"], ArcticQuantizationConfig):
+            ret["quantization"] = asdict(ret["quantization"])
+        return ret

vllm/transformers_utils/configs/chatglm.py

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://github.com/zai-org/ChatGLM2-6B
+from transformers import PretrainedConfig
+
+
+class ChatGLMConfig(PretrainedConfig):
+    model_type = "chatglm"
+    attribute_map = {
+        "num_hidden_layers": "num_layers",
+        "n_head_kv": "multi_query_group_num",
+    }
+
+    def __init__(self,
+                 num_layers=28,
+                 padded_vocab_size=65024,
+                 hidden_size=4096,
+                 ffn_hidden_size=13696,
+                 kv_channels=128,
+                 num_attention_heads=32,
+                 seq_length=2048,
+                 hidden_dropout=0.0,
+                 attention_dropout=0.0,
+                 layernorm_epsilon=1e-5,
+                 rmsnorm=True,
+                 apply_residual_connection_post_layernorm=False,
+                 post_layer_norm=True,
+                 add_bias_linear=False,
+                 add_qkv_bias=False,
+                 interleaved_qkv=False,
+                 bias_dropout_fusion=True,
+                 multi_query_attention=False,
+                 multi_query_group_num=1,
+                 apply_query_key_layer_scaling=True,
+                 attention_softmax_in_fp32=True,
+                 fp32_residual_connection=False,
+                 quantization_bit=0,
+                 pre_seq_len=None,
+                 prefix_projection=False,
+                 **kwargs):
+        self.num_layers = num_layers
+        self.vocab_size = padded_vocab_size
+        self.padded_vocab_size = padded_vocab_size
+        self.hidden_size = hidden_size
+        self.ffn_hidden_size = ffn_hidden_size
+        self.kv_channels = kv_channels
+        self.num_attention_heads = num_attention_heads
+        self.seq_length = seq_length
+        # It is to be compatible with long lora.
+        self.max_position_embeddings = seq_length
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.layernorm_epsilon = layernorm_epsilon
+        self.rmsnorm = rmsnorm
+        self.apply_residual_connection_post_layernorm = (
+            apply_residual_connection_post_layernorm)
+        self.post_layer_norm = post_layer_norm
+        self.add_bias_linear = add_bias_linear
+        self.add_qkv_bias = add_qkv_bias
+        self.bias_dropout_fusion = bias_dropout_fusion
+        self.multi_query_attention = multi_query_attention
+        self.multi_query_group_num = multi_query_group_num
+        self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
+        self.attention_softmax_in_fp32 = attention_softmax_in_fp32
+        self.fp32_residual_connection = fp32_residual_connection
+        self.quantization_bit = quantization_bit
+        self.pre_seq_len = pre_seq_len
+        self.prefix_projection = prefix_projection
+        self.interleaved_qkv = interleaved_qkv
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/deepseek_v3.py

@@ -0,0 +1,101 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class DeepseekV3Config(PretrainedConfig):
+
+    model_type = "deepseek_v3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=129280,
+        hidden_size=7168,
+        intermediate_size=18432,
+        moe_intermediate_size=2048,
+        num_hidden_layers=61,
+        num_nextn_predict_layers=1,
+        num_attention_heads=128,
+        num_key_value_heads=128,
+        n_shared_experts=1,
+        n_routed_experts=256,
+        ep_size=1,
+        routed_scaling_factor=2.5,
+        kv_lora_rank=512,
+        q_lora_rank=1536,
+        qk_rope_head_dim=64,
+        v_head_dim=128,
+        qk_nope_head_dim=128,
+        topk_method='noaux_tc',
+        n_group=8,
+        topk_group=4,
+        num_experts_per_tok=8,
+        moe_layer_freq=1,
+        first_k_dense_replace=3,
+        norm_topk_prob=True,
+        scoring_func='sigmoid',
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=0,
+        eos_token_id=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_nextn_predict_layers = num_nextn_predict_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
+        # 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.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        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,
+        )

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/dotsocr.py

@@ -0,0 +1,69 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.models.qwen2 import Qwen2Config
+
+
+class DotsVisionConfig(PretrainedConfig):
+    model_type: str = "dots_vit"
+
+    def __init__(
+        self,
+        embed_dim: int = 1536,  # vision encoder embed size
+        hidden_size: int = 1536,  # after merger hidden size
+        intermediate_size: int = 4224,
+        num_hidden_layers: int = 42,
+        num_attention_heads: int = 12,
+        num_channels: int = 3,
+        patch_size: int = 14,
+        spatial_merge_size: int = 2,
+        temporal_patch_size: int = 1,
+        rms_norm_eps: float = 1e-5,
+        use_bias: bool = False,
+        attn_implementation="flash_attention_2",
+        initializer_range=0.02,
+        init_merger_std=0.02,
+        is_causal=False,  # ve causal forward
+        post_norm=True,
+        gradient_checkpointing=False,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+        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.num_channels = num_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+        self.rms_norm_eps = rms_norm_eps
+        self.use_bias = use_bias
+        self.attn_implementation = attn_implementation
+        self.initializer_range = initializer_range
+        self.init_merger_std = init_merger_std
+        self.is_causal = is_causal
+        self.post_norm = post_norm
+        self.gradient_checkpointing = gradient_checkpointing
+
+
+class DotsOCRConfig(Qwen2Config):
+    model_type = "dots_ocr"
+
+    def __init__(self,
+                 image_token_id=151665,
+                 video_token_id=151656,
+                 vision_config: Optional[dict] = None,
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.image_token_id = image_token_id
+        self.video_token_id = video_token_id
+        self.vision_config = DotsVisionConfig(**(vision_config or {}))
+
+    def save_pretrained(self, save_directory, **kwargs):
+        self._auto_class = None
+        super().save_pretrained(save_directory, **kwargs)

vllm/transformers_utils/configs/eagle.py

@@ -0,0 +1,84 @@
+# 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
+
+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
+
+        # Eagle model name should follow naming convention of
+        # LlamaForCausalLM -> EagleLlamaForCausalLM
+        # LlamaForCausalLM -> Eagle3LlamaForCausalLM
+        # LlamaForCausalLMEagle3 -> LlamaForCausalLMEagle3
+        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"] = [
+                arch if arch.startswith("Eagle3") or arch.endswith("Eagle3")
+                else f"Eagle3{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)

vllm/transformers_utils/configs/falcon.py

@@ -0,0 +1,90 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from
+# https://huggingface.co/tiiuae/falcon-7b/blob/main/configuration_RW.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 the Big Science Workshop and 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.
+"""Falcon configuration"""
+from transformers.configuration_utils import PretrainedConfig
+
+
+class RWConfig(PretrainedConfig):
+    model_type = "falcon"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_hidden_layers": "n_layer",
+        "num_attention_heads": "n_head",
+        "num_kv_heads": "n_head_kv",
+    }
+
+    def __init__(
+        self,
+        vocab_size=250880,
+        hidden_size=64,
+        n_layer=2,
+        n_head=8,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        multi_query=True,
+        n_head_kv=None,
+        alibi=False,
+        bias=False,
+        parallel_attn=False,
+        new_decoder_architecture=False,
+        **kwargs,
+    ) -> None:
+        self.vocab_size = vocab_size
+        # Backward compatibility with n_embed kwarg
+        n_embed = kwargs.pop("n_embed", None)
+        self.hidden_size = hidden_size if n_embed is None else n_embed
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.multi_query = multi_query
+        self.n_head_kv = 1 if n_head_kv is None else n_head_kv
+        self.alibi = alibi
+        self.bias = bias
+        self.parallel_attn = parallel_attn
+        self.new_decoder_architecture = new_decoder_architecture
+
+        if self.hidden_size == 8192:
+            # Hack for falcon-40b
+            self.new_decoder_architecture = True
+
+        super().__init__(bos_token_id=bos_token_id,
+                         eos_token_id=eos_token_id,
+                         **kwargs)
+
+    @property
+    def head_dim(self):
+        return self.hidden_size // self.n_head
+
+    @property
+    def rotary(self):
+        return not self.alibi

vllm/transformers_utils/configs/jais.py

@@ -0,0 +1,237 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+# Copyright 2023 Cerebras Systems.
+#
+# 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.
+"""JAIS configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class JAISConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a
+    [`JAISModel`]. It is used to instantiate a JAIS 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:
+        vocab_size (`int`, *optional*, defaults to 50257):
+            Vocabulary size of the JAIS model. Defines the number of different
+            tokens that can be represented by the
+            `inputs_ids` passed when calling [`JAISModel`].
+        n_positions (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used
+            with. Typically set this to something large just in case
+            (e.g., 512 or 1024 or 2048).
+        n_embd (`int`, *optional*, defaults to 768):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        n_head (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the
+            Transformer encoder.
+        n_inner (`int`, *optional*, defaults to None):
+            Dimensionality of the inner feed-forward layers. `None` will set
+            it to 4 times n_embd
+        activation_function (`str`, *optional*, defaults to `"gelu"`):
+            Activation function, to be selected in the list
+            `["relu", "silu", "gelu", "tanh", "gelu_new", "swiglu"]`.
+        resid_pdrop (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in
+            the embeddings, encoder, and pooler.
+        embd_pdrop (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the embeddings.
+        attn_pdrop (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for
+            initializing all weight matrices.
+        scale_attn_weights (`bool`, *optional*, defaults to `True`):
+            Scale attention weights by dividing by sqrt(hidden_size)..
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values
+            attentions (not used by all models).
+        scale_attn_by_inverse_layer_idx (`bool`, *optional*, default `True`):
+            Whether to additionally scale attention weights 
+            by `1 / layer_idx + 1`.
+        reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
+            Whether to scale keys (K) prior to computing attention
+            (dot-product)
+            and upcast attention dot-product/softmax to float() when training
+            with mixed precision.
+        position_embedding_type (`str`, *optional*, defaults to `"learned"`):
+            Positional embedding can be either `"alibi"` or `"learned"`.
+        mup_width_scale (`float`, *optional*, defaults to 1.0):
+            muP parameter to scale learning rate and initializers. Calculated
+            as (`d_model,0 / d_model`), where
+            `d_model` is the model's width and `d_model,0` is the proxy
+            model's width.
+        mup_embeddings_scale (`float`, *optional*, defaults to 1.0):
+            muP parameter to scale token and position embeddings.
+        mup_output_alpha (`float`, *optional*, defaults to 1.0):
+            muP parameter to scale output logits
+            (`output_logits_scale = mup_output_alpha * mup_width_scale`).
+        mup_scale_qk_dot_by_d (`bool`, *optional*, defaults to `False`):
+            Scale attention weights by dividing by hidden_size instead of
+            sqrt(hidden_size). Need to set scale_attn_weights to `True` as
+            well.
+        alibi_scaling (`dict`, *optional*):
+            Dictionary containing the scaling configuration for ALiBi
+            embeddings. Currently only supports linear
+            scaling strategy. Can specify either the scaling `factor` (must be
+            a float greater than 1) for fixed scaling
+            or `train_seq_len` for dynamic scaling on input samples with
+            sequence length > `train_seq_len`. The expected
+            formats are `{"type": strategy name, "factor": scaling factor}` or
+            `{"type": strategy name,
+            "train_seq_len": training sequence length}`.
+        architectures (`list`, *optional*, defaults to ['JAISLMHeadModel']):
+            architecture names for Jais.
+
+    Example:
+
+    ```python
+    >>> from transformers import JAISConfig, JAISModel
+
+    >>> # Initializing a JAIS configuration
+    >>> configuration = JAISConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = JAISModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "jais"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "n_embd",
+        "max_position_embeddings": "n_positions",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50257,
+        n_positions=1024,
+        n_embd=768,
+        n_layer=12,
+        n_head=12,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.1,
+        embd_pdrop=0.1,
+        attn_pdrop=0.1,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        scale_attn_weights=True,
+        use_cache=True,
+        bos_token_id=50256,
+        eos_token_id=50256,
+        scale_attn_by_inverse_layer_idx=False,
+        reorder_and_upcast_attn=False,
+        position_embedding_type="learned",
+        mup_width_scale=1.0,
+        mup_embeddings_scale=1.0,
+        mup_output_alpha=1.0,
+        mup_scale_qk_dot_by_d=False,
+        alibi_scaling=None,
+        architectures=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+        self.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
+        self.reorder_and_upcast_attn = reorder_and_upcast_attn
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        self.position_embedding_type = position_embedding_type
+        self.mup_width_scale = mup_width_scale
+        self.mup_embeddings_scale = mup_embeddings_scale
+        self.mup_output_alpha = mup_output_alpha
+        self.mup_scale_qk_dot_by_d = mup_scale_qk_dot_by_d
+
+        self.alibi_scaling = alibi_scaling
+        self._alibi_scaling_validation()
+        if architectures is None:
+            architectures = ["JAISLMHeadModel"]
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            architectures=architectures,
+            **kwargs,
+        )
+
+    def _alibi_scaling_validation(self):
+        """
+        Validate the `alibi_scaling` configuration.
+        """
+        if self.alibi_scaling is None:
+            return
+
+        if (not isinstance(self.alibi_scaling, dict)
+                or len(self.alibi_scaling) != 2):
+            raise ValueError(
+                "`alibi_scaling` must be a dictionary with two fields, "
+                "`type` and `factor` or `type` and `train_seq_len`, "
+                f"got {self.alibi_scaling}")
+        alibi_scaling_type = self.alibi_scaling.get("type", None)
+        alibi_scaling_factor = self.alibi_scaling.get("factor", None)
+        alibi_dynamic_scaling = self.alibi_scaling.get("train_seq_len", None)
+        if alibi_scaling_type is None or alibi_scaling_type != "linear":
+            raise ValueError(f"`alibi_scaling`'s type field must be 'linear', "
+                             f"got {alibi_scaling_type}")
+        if (alibi_scaling_factor is not None
+                and not isinstance(alibi_scaling_factor, float)
+                or (alibi_scaling_factor is not None
+                    and alibi_scaling_factor <= 1.0)):
+            raise ValueError(
+                f"`alibi_scaling`'s factor field must be a float > 1.0, "
+                f"got {alibi_scaling_factor}")
+        if (alibi_dynamic_scaling is not None
+                and not isinstance(alibi_dynamic_scaling, int)
+                or (alibi_dynamic_scaling is not None
+                    and alibi_dynamic_scaling <= 1)):
+            raise ValueError(
+                f"`alibi_scaling`'s `train_seq_len` field must be an "
+                f"integer > 1, got {alibi_dynamic_scaling}")

vllm/transformers_utils/configs/kimi_vl.py

@@ -0,0 +1,37 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
+from typing import Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
+from vllm.transformers_utils.configs.moonvit import MoonViTConfig
+
+
+class KimiVLConfig(PretrainedConfig):
+    model_type = "kimi_vl"
+
+    def __init__(self,
+                 vision_config: Optional[Union[dict, MoonViTConfig]] = None,
+                 text_config: Optional[Union[dict, DeepseekV2Config]] = None,
+                 ignore_index: int = -100,
+                 media_placeholder_token_id: int = 163605,
+                 pad_token_id: int = 0,
+                 **kwargs):
+        if vision_config is None:
+            vision_config = MoonViTConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = MoonViTConfig(**vision_config)
+        self.vision_config = vision_config
+
+        if text_config is None:
+            text_config = DeepseekV2Config()
+        elif isinstance(text_config, dict):
+            text_config = DeepseekV2Config(**text_config)
+        self.text_config = text_config
+
+        self.ignore_index = ignore_index
+        self.media_placeholder_token_id = media_placeholder_token_id
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)

vllm/transformers_utils/configs/medusa.py

@@ -0,0 +1,63 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+from typing import Optional, Union
+
+from transformers import PretrainedConfig
+
+
+class MedusaConfig(PretrainedConfig):
+    model_type = "medusa"
+
+    def __init__(self,
+                 hidden_size: int = 4096,
+                 vocab_size: int = 32001,
+                 num_heads: int = 5,
+                 num_hidden_layers: int = 1,
+                 max_paths: int = 64,
+                 topk: int = 10,
+                 truncated_vocab_size: Optional[int] = None,
+                 **kwargs):
+
+        self.hidden_size = hidden_size
+        self.vocab_size = vocab_size
+        self.num_heads = num_heads
+        self.num_hidden_layers = num_hidden_layers
+        self.max_paths = max_paths
+        self.topk = topk
+        self.max_seq_len = int(2**20)
+        self.truncated_vocab_size = vocab_size if truncated_vocab_size is None\
+            else truncated_vocab_size
+        if "architectures" not in kwargs:
+            kwargs["architectures"] = ["MedusaModel"]
+
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        **kwargs,
+    ) -> "MedusaConfig":
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs)
+        for k in list(config_dict.keys()):
+            if 'num' in k:
+                if 'heads' in k:
+                    config_dict["num_heads"] = config_dict.pop(k)
+                elif 'layers' in k:
+                    config_dict["num_hidden_layers"] = config_dict.pop(k)
+        return cls.from_dict(config_dict, **kwargs)
+
+    @property
+    def num_attention_heads(self):
+        return 0
+
+    @property
+    def num_lookahead_tokens(self):
+        return self.num_heads
+
+    @num_lookahead_tokens.setter
+    def num_lookahead_tokens(self, num_lookahead_tokens: int):
+        self.num_heads = num_lookahead_tokens

vllm/transformers_utils/configs/midashenglm.py

@@ -0,0 +1,101 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright 2025 Horizon team, Xiaomi MiLM Plus.
+# Copyright 2024 The Qwen team.
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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 Optional, Union
+
+from transformers import PretrainedConfig
+from transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import (
+    Qwen2_5OmniTextConfig)
+
+
+class DashengConfig(PretrainedConfig):
+    model_type = "midashenglm_dasheng_encoder"
+
+    def __init__(
+        self,
+        embed_dim: int = 768,
+        outputdim: int = 527,
+        patch_size: Union[int, tuple[int, int]] = 16,
+        patch_stride: Union[int, tuple[int, int]] = 16,
+        input_channels: int = 1,
+        target_length: int = 1012,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        init_values: Optional[float] = None,
+        drop_rate: float = 0.0,
+        attn_drop_rate: float = 0.0,
+        f_min: float = 0.0,
+        f_max: float = 8000.0,
+        center: bool = True,
+        win_length: int = 512,
+        hop_length: int = 160,
+        sample_rate: int = 16000,
+        n_fft: int = 512,
+        n_mels: int = 64,
+        **kwargs,
+    ):
+        self.embed_dim = embed_dim
+        self.outputdim = outputdim
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.input_channels = input_channels
+        self.target_length = target_length
+        self.depth = depth
+        self.num_heads = num_heads
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.init_values = init_values
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.f_min = f_min
+        self.f_max = f_max
+        self.center = center
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.sample_rate = sample_rate
+        self.n_fft = n_fft
+        self.n_mels = n_mels
+        super().__init__(**kwargs)
+
+
+class MiDashengLMConfig(PretrainedConfig):
+    model_type = "midashenglm"
+
+    def __init__(
+        self,
+        audio_encoder_config: Optional[dict] = None,
+        subsample_factor: int = 5,
+        text_config: Optional[dict] = None,
+        audio_token_id: Optional[int] = None,
+        **kwargs,
+    ):
+        self.audio_encoder_config = DashengConfig(
+            **(audio_encoder_config or {}))
+        self.subsample_factor = subsample_factor
+        self.text_config = (Qwen2_5OmniTextConfig(
+            **text_config) if text_config else Qwen2_5OmniTextConfig())
+        self.text_config.rope_scaling = None  # uses_mrope is false
+        self.audio_token_id = audio_token_id
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/mistral.py

@@ -0,0 +1,165 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any
+
+from transformers import PretrainedConfig, WhisperConfig
+
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def adapt_config_dict(config_dict: dict[str, Any],
+                      **kwargs) -> PretrainedConfig:
+    config_dict.update(kwargs)
+    config_dict = _remap_general_mistral_args(config_dict)
+
+    if bool(config_dict.get("quantization")):
+        config_dict = _remap_mistral_quantization_args(config_dict)
+
+    if bool(config_dict.get("moe")):
+        config_dict["architectures"] = ["MixtralForCausalLM"]
+    else:
+        config_dict["architectures"] = ["MistralForCausalLM"]
+
+    if bool(config_dict.get("yarn")):
+        config_dict = _remap_mistral_yarn_args(config_dict)
+
+    is_vision = ((config_dict.get("multimodal")
+                  or {}).get("vision_encoder_args")
+                 or config_dict.get("vision_encoder"))
+    is_audio = bool(
+        ((config_dict.get("multimodal") or {}).get("whisper_model_args")
+         or {}).get("encoder_args"))
+
+    assert not (is_vision and is_audio), \
+        "Vision and audio are mutually exclusive"
+
+    if is_vision:
+        config_dict = _remap_mistral_vision_args(config_dict)
+    if is_audio:
+        config_dict = _remap_mistral_audio_args(config_dict)
+
+    config = PretrainedConfig.from_dict(config_dict)
+
+    logger.debug("Initialized config %s", config)
+
+    return config
+
+
+def _remap_mistral_vision_args(config: dict) -> dict:
+    if config.get("multimodal"):
+        vision_config = config.pop("multimodal")
+    else:
+        vision_config = config.pop("vision_encoder")
+
+    quant_config = config.get("quantization_config")
+    config = {
+        "model_type": "pixtral",
+        "architectures": ["PixtralForConditionalGeneration"],
+        "text_config": PretrainedConfig.from_dict(config),
+        "vision_config": PretrainedConfig.from_dict(vision_config),
+    }
+    if quant_config:
+        config["quantization_config"] = quant_config
+    return config
+
+
+def _remap_mistral_yarn_args(config: dict) -> dict:
+    # Direct remaps: yarn.X -> rope_scaling.Y
+    # Source keys are from mistral.model.args.YarnArgs
+    _map = {
+        "beta": "beta_fast",
+        "alpha": "beta_slow",
+    }
+    yarn_config = config.get("yarn") or {}
+    renamed_yarn_config = {_map.get(k, k): v for k, v in yarn_config.items()}
+    config["rope_scaling"] = {
+        "rope_type": "yarn",
+        "mscale_all_dim": 1,  # We hardcoded this to 1
+        **renamed_yarn_config
+    }
+    return config
+
+
+def _remap_general_mistral_args(config: dict) -> dict:
+    # Mistral key -> HF key
+    config_mapping = {
+        "dim": "hidden_size",
+        "norm_eps": "rms_norm_eps",
+        "n_kv_heads": "num_key_value_heads",
+        "n_layers": "num_hidden_layers",
+        "n_heads": "num_attention_heads",
+        "hidden_dim": "intermediate_size",
+    }
+    # HF key -> (Mistral key, default value)
+    top_level_mapping_with_default = {
+        "model_type": ("model_type", "transformer"),
+        "hidden_act": ("activation", "silu"),
+        "tie_word_embeddings": ("tied_embeddings", False),
+        "max_seq_len": ("max_seq_len", 128_000),
+        "max_position_embeddings": ("max_position_embeddings", 128_000),
+    }
+
+    for key, new_key in config_mapping.items():
+        if key in config:
+            config[new_key] = config.pop(key)
+
+    for new_key, (key,
+                  default_value) in top_level_mapping_with_default.items():
+        config[new_key] = config.pop(key, default_value)
+
+    return config
+
+
+def _remap_mistral_quantization_args(config: dict) -> dict:
+    quantization = config.get("quantization", {})
+    if quantization.get("qformat_weight") == "fp8_e4m3":
+        # This maps to the FP8 static per-tensor quantization scheme
+        quantization_config = {
+            "quant_method": "fp8",
+            "activation_scheme": "static"
+        }
+    elif quantization.get("quant_method") == "compressed-tensors":
+        # Pass through the quantization config to compressed-tensors
+        quantization_config = quantization
+    else:
+        raise ValueError(
+            f"Found unknown quantization='{quantization}' in config")
+
+    config["quantization_config"] = quantization_config
+
+    return config
+
+
+def _remap_mistral_audio_args(config: dict) -> dict:
+    whisper_args = config["multimodal"].pop("whisper_model_args")
+    encoder_args = whisper_args["encoder_args"]
+    downsample_args = whisper_args["downsample_args"]
+
+    quant_config = config.get("quantization_config")
+    config = {
+        "model_type":
+        "whixtral",
+        "architectures": ["VoxtralForConditionalGeneration"],
+        "text_config":
+        PretrainedConfig.from_dict(config),
+        "audio_config":
+        WhisperConfig(
+            num_mel_bins=encoder_args["audio_encoding_args"]["num_mel_bins"],
+            window_size=encoder_args["audio_encoding_args"]["window_size"],
+            sampling_rate=encoder_args["audio_encoding_args"]["sampling_rate"],
+            hop_length=encoder_args["audio_encoding_args"]["hop_length"],
+            downsample_factor=downsample_args["downsample_factor"],
+            d_model=encoder_args["dim"],
+            encoder_layers=encoder_args["n_layers"],
+            encoder_ffn_dim=encoder_args["hidden_dim"],
+            encoder_attention_heads=encoder_args["n_heads"],
+            vocab_size=encoder_args["vocab_size"],
+            max_source_positions=encoder_args["max_source_positions"],
+            is_encoder_decoder=False,  # Override WhisperConfig default
+        )
+    }
+    if quant_config:
+        config["quantization_config"] = quant_config
+    return config

vllm/transformers_utils/configs/mlp_speculator.py

@@ -0,0 +1,68 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Optional
+
+from transformers import PretrainedConfig
+
+
+class MLPSpeculatorConfig(PretrainedConfig):
+    model_type = "mlp_speculator"
+
+    attribute_map = {
+        "hidden_size": "emb_dim",
+    }
+
+    def __init__(self,
+                 vocab_size: int = 32000,
+                 emb_dim: int = 4096,
+                 inner_dim: int = 0,
+                 n_predict: int = 3,
+                 top_k_tokens_per_head: Optional[list[int]] = None,
+                 n_candidates: int = 5,
+                 tie_weights: bool = False,
+                 scale_input: bool = False,
+                 **kwargs):
+        """
+        Initialize an MLPSpeculatorConfig
+
+        Args:
+            vocab_size: int
+                the model vocab size
+            emb_dim: int
+                the model embedding dimension
+            inner_dim: int
+                the inner dimension of the model. If 0, will be the emb_dim.
+            n_predict: int
+                the number of lookaheads for the speculator
+            top_k_tokens_per_head: list[int]
+                Number of tokens to consider from each head when forming the
+                candidate tree.
+                For each candidate branch in the tree, head n produces topk[n]
+                additional sub-branches.
+                NOTE: This parameter is currently unused.
+            n_candidates: int
+                number of child candidates to create per sequence
+            tie_weights: bool
+                If true, use a single set of weights for every model
+                head/stage after the first. The initial projection
+                from the base model may have a different size, so that
+                stays separate.
+            scale_input: bool
+                if True, will scale the initial hidden states from
+                the base model.
+        """
+        if top_k_tokens_per_head is None:
+            top_k_tokens_per_head = [5, 4, 3]
+        assert len(top_k_tokens_per_head) == n_predict
+        self.vocab_size = vocab_size
+        self.emb_dim = emb_dim
+        self.inner_dim = inner_dim
+        self.n_predict = n_predict
+        self.top_k_tokens_per_head = top_k_tokens_per_head
+        self.n_candidates = n_candidates
+        self.num_lookahead_tokens = n_predict
+        self.tie_weights = tie_weights
+        self.scale_input = scale_input
+
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/moonvit.py

@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
+from transformers.configuration_utils import PretrainedConfig
+
+
+class MoonViTConfig(PretrainedConfig):
+    model_type = "moonvit"
+
+    def __init__(
+            self,
+            patch_size: int = 14,
+            init_pos_emb_height: int = 64,
+            init_pos_emb_width: int = 64,
+            num_attention_heads: int = 16,
+            num_hidden_layers: int = 27,
+            hidden_size: int = 1152,
+            intermediate_size: int = 4304,
+            merge_kernel_size: tuple[int, int] = (2, 2),
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        # Positional embedding config
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        # Transformer config
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        # Patch merger config
+        self.merge_kernel_size = merge_kernel_size

vllm/transformers_utils/configs/nemotron.py

@@ -0,0 +1,205 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2024, NVIDIA CORPORATION. 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.
+"""Nemotron model configuration"""
+
+from transformers import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a
+    [`NemotronModel`]. It is used to instantiate a Nemotron 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 Nemotron-8B.
+
+    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 256000):
+            Vocabulary size of the Nemotron model. Defines the number of
+            different tokens that can be represented by the
+            `inputs_ids` passed when calling [`NemotronModel`]
+        hidden_size (`int`, *optional*, defaults to 6144):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 24576):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 48):
+            Number of attention heads for each attention layer in the
+            Transformer decoder.
+        head_dim (`int`, *optional*):
+            Projection weights dimension in multi-head attention. Set to
+            hidden_size // num_attention_heads if None
+        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 `"relu2"`):
+            The non-linear activation function (function or string) in the
+            decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used
+            with.
+        initializer_range (`float`, *optional*, defaults to 0.0134):
+            The standard deviation of the truncated_normal_initializer for
+            initializing all weight matrices.
+        norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the 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 2):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 3):
+            End of stream token id.
+        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.
+        partial_rotary_factor (`float`, *optional*, defaults to 0.5):
+            Percentage of the query and keys which will have rotary embedding.
+        attention_bias (`bool`, *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.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj and down_proj layers in the MLP
+            layers.
+
+    ```python
+    >>> from transformers import NemotronModel, NemotronConfig
+    >>> # Initializing a Nemotron nemotron-15b style configuration
+    >>> configuration = NemotronConfig()
+    >>> # Initializing a model from the nemotron-15b style configuration
+    >>> model = NemotronModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "nemotron"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=256000,
+        hidden_size=6144,
+        intermediate_size=24576,
+        num_hidden_layers=32,
+        num_attention_heads=48,
+        head_dim=None,
+        num_key_value_heads=None,
+        hidden_act="relu2",
+        max_position_embeddings=4096,
+        initializer_range=0.0134,
+        norm_eps=1e-5,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=2,
+        eos_token_id=3,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        partial_rotary_factor=0.5,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        **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
+        head_dim = head_dim or kwargs.get("kv_channels")
+        self.head_dim = head_dim if head_dim is not None else (
+            hidden_size // num_attention_heads)
+
+        # 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.norm_eps = norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        # for backward compatibility
+        partial_rotary_factor = kwargs.get("rope_percent") or kwargs.get(
+            "rope_percentage") or partial_rotary_factor
+        self.partial_rotary_factor = partial_rotary_factor
+        self._rope_scaling_validation()
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+
+        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,
+        )
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(
+                self.rope_scaling) != 2:
+            raise ValueError(
+                "`rope_scaling` must be a dictionary with two fields, "
+                f"`type` and `factor`, got {self.rope_scaling}")
+        rope_scaling_type = self.rope_scaling.get("type", None)
+        rope_scaling_factor = self.rope_scaling.get("factor", None)
+        if rope_scaling_type is None or rope_scaling_type not in [
+                "linear", "dynamic"
+        ]:
+            raise ValueError(
+                "`rope_scaling`'s type field must be one of ['linear', "
+                f"'dynamic'], got {rope_scaling_type}")
+        if rope_scaling_factor is None or not isinstance(
+                rope_scaling_factor, float) or rope_scaling_factor <= 1.0:
+            raise ValueError(
+                "`rope_scaling`'s factor field must be a float > 1, got "
+                f"{rope_scaling_factor}")

vllm/transformers_utils/configs/nemotron_h.py

@@ -0,0 +1,259 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2024 HuggingFace Inc. team. All rights reserved.
+# Copyright (c) 2025, NVIDIA CORPORATION. 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.
+"""NemotronH model configuration"""
+
+import regex as re
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class NemotronHConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a
+    [`NemotronHModel`]. It is used to instantiate a NemotronH 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 NemotronH-v0.1 model.
+    Args:
+        vocab_size (`int`, *optional*, defaults to 131072):
+            Vocabulary size of the NemotronH model. Defines the number of
+            different tokens that can be represented by the `inputs_ids`
+            passed when calling [`NemotronHModel`]
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be
+            tied. Note that this is only relevant if the model has an output
+            word embedding layer.
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 21504):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 52):
+            Number of hidden layers in the Transformer encoder.
+        hybrid_override_pattern (`str`, *optional*, defaults to
+            `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
+            The pattern of the hybrid model. The pattern is a string of
+            characters where each character represents
+            M: Mamba2, *: Attention, -: MLP
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the
+            Transformer encoder.
+        attention_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each attention head.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            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.
+        mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
+            The non-linear activation function in the MLP layers.
+        attention_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in attention layers.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in MLP layers.
+        use_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the model.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for
+            initializing all weight matrices.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        residual_in_fp32 (`bool`, *optional*, defaults to `False`):
+            Whether or not residuals should be in `float32`. If set to `False`
+            residuals will keep the same `dtype` as the rest of the model.
+        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`.
+        num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
+            Number of prompt logits to calculate during generation. If `None`,
+            all logits will be calculated. If an integer value, only last
+            `num_logits_to_keep` logits will be calculated.
+        pad_token_id (`int`, *optional*, defaults to 0):
+            The id of the padding token.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the "end-of-sequence" token.
+        sliding_window (`int`, *optional*, defaults to None):
+            Sliding window attention window size.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used
+            with.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        hidden_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the hidden states.
+        use_mamba_kernels (`bool`, *optional*, defaults to `True`):
+            Flag indicating whether or not to use the fast mamba kernels.
+            These are available only if `mamba-ssm` and `causal-conv1d`
+            are installed, and the mamba modules are running on a CUDA device.
+        ssm_state_size (`int`, *optional*, defaults to 128):
+            The dimension of the mamba state space latents.
+        mamba_num_heads (`int`, *optional*, defaults to 128):
+            Number of heads in Mamba layers.
+        mamba_n_groups (`int`, *optional*, defaults to 8):
+            Number of groups in Mamba layers.
+        mamba_head_dim (`int`, *optional*, defaults to 64):
+            Dimension of each Mamba head.
+        mamba_d_conv (`int`, *optional*, defaults to 4):
+            The size of the mamba convolution kernel.
+        mamba_expand (`int`, *optional*, defaults to 2):
+            Expanding factor used to determine the mamba intermediate size.
+        mamba_hidden_act (`str`, *optional*, defaults to "silu"):
+            The non-linear activation function in the Mamba layers.
+        mamba_dt_min (`float`, *optional*, defaults to 0.001):
+            Minimum value for the time step in Mamba.
+        mamba_dt_max (`float`, *optional*, defaults to 0.1):
+            Maximum value for the time step in Mamba.
+        mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
+            Limits for the time step in Mamba.
+        mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
+            Floor value for time step initialization in Mamba.
+        mamba_conv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to use bias in the convolution layer of the mamba mixer
+            block.
+        mamba_proj_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use bias in the input and output projections of the
+            mamba mixer block.
+        mamba_chunk_size (`int`, *optional*, defaults to 256):
+            Size of chunks for Mamba processing.
+        rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the pre-normalization residual connections.
+    """
+
+    model_type = "nemotron_h"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=131072,
+        tie_word_embeddings=False,
+        hidden_size=4096,
+        intermediate_size=21504,
+        num_hidden_layers=52,
+        hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
+        num_attention_heads=32,
+        head_dim=128,
+        num_key_value_heads=8,  # nemo: num_query_groups
+        mlp_hidden_act="relu2",
+        attention_bias=False,
+        mlp_bias=False,
+        use_bias=False,
+        initializer_range=0.02,  # nemo: init_method_std
+        layer_norm_epsilon=1e-5,  # nemo: layernorm_epsilon
+        residual_in_fp32=False,  #  Megatron Core default value
+        use_cache=True,
+        num_logits_to_keep=1,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        sliding_window=None,
+        max_position_embeddings=4096,
+        attention_dropout=0.0,
+        hidden_dropout=0.0,  # * ADDED
+        use_mamba_kernels=True,
+        ssm_state_size=128,  # mamba_state_size
+        mamba_num_heads=128,
+        mamba_n_groups=8,  # nemo: mamba_ssm_ngroups = num_heads
+        mamba_head_dim=64,
+        mamba_d_conv=4,
+        mamba_expand=2,
+        mamba_hidden_act="silu",
+        mamba_dt_min=0.001,
+        mamba_dt_max=0.1,
+        mamba_dt_limit=(0.0, float("inf")),
+        mamba_dt_init_floor=1e-4,
+        mamba_conv_bias=True,
+        mamba_proj_bias=False,
+        mamba_chunk_size=256,
+        rescale_prenorm_residual=True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.tie_word_embeddings = tie_word_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hybrid_override_pattern = hybrid_override_pattern
+        self.num_attention_heads = num_attention_heads
+        self.head_dim = head_dim
+        self.sliding_window = sliding_window
+        self.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.hidden_dropout = hidden_dropout
+
+        # Validate hybrid_override_pattern
+        # M: Mamba2, *: Attention, -: MLP
+        assert len(self.hybrid_override_pattern) == self.num_hidden_layers, (
+            "hybrid_override_pattern must have same length as "
+            "num_hidden_layers")
+        assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), (
+            "hybrid_override_pattern must only contain characters "
+            "'M', '*', or '-'")
+
+        # 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.mlp_hidden_act = mlp_hidden_act
+        self.attention_bias = attention_bias
+        self.mlp_bias = mlp_bias
+        self.use_bias = use_bias
+        self.initializer_range = initializer_range
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.residual_in_fp32 = residual_in_fp32
+
+        self.use_cache = use_cache
+        self.num_logits_to_keep = num_logits_to_keep
+
+        self.use_mamba_kernels = use_mamba_kernels
+        self.n_groups = mamba_n_groups
+        self.mamba_head_dim = mamba_head_dim
+        self.ssm_state_size = ssm_state_size
+        self.mamba_num_heads = mamba_num_heads
+        self.conv_kernel = mamba_d_conv
+        self.expand = mamba_expand
+        self.mamba_hidden_act = mamba_hidden_act
+        self.time_step_min = mamba_dt_min
+        self.time_step_max = mamba_dt_max
+        self.time_step_limit = mamba_dt_limit
+        self.time_step_floor = mamba_dt_init_floor
+        self.use_conv_bias = mamba_conv_bias
+        self.mamba_proj_bias = mamba_proj_bias
+        self.chunk_size = mamba_chunk_size
+        self.rescale_prenorm_residual = rescale_prenorm_residual
+
+        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,
+        )
+
+    @property
+    def layers_block_type(self):
+        return [
+            "mamba" if self.hybrid_override_pattern[i] == "M" else
+            "attention" if self.hybrid_override_pattern[i] == "*" else "mlp"
+            for i in range(self.num_hidden_layers)
+        ]

vllm/transformers_utils/configs/nemotron_vl.py

@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# yapf: disable
+# ruff: noqa: E501
+# Adapted from
+# https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/blob/main/configuration.py
+# --------------------------------------------------------
+# Adapted from https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B under MIT License
+#     LICENSE is in incl_licenses directory.
+# --------------------------------------------------------
+
+from transformers import LlamaConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.dynamic_module_utils import get_class_from_dynamic_module
+
+
+class Nemotron_Nano_VL_Config(PretrainedConfig):
+    model_type = 'Llama_Nemotron_Nano_VL'
+    is_composition = True
+
+    def __init__(
+        self,
+        vision_config=None,
+        llm_config=None,
+        force_image_size=None,
+        downsample_ratio=0.5,
+        template=None,
+        ps_version='v1',
+        image_tag_type="internvl",
+        projector_hidden_size=4096,
+        vit_hidden_size=1280,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if vision_config is not None:
+            assert "auto_map" in vision_config and "AutoConfig" in vision_config["auto_map"]
+            vision_auto_config = get_class_from_dynamic_module(*vision_config["auto_map"]["AutoConfig"].split("--")[::-1])
+            self.vision_config = vision_auto_config(**vision_config)
+        else:
+            self.vision_config = PretrainedConfig()
+
+        if llm_config is None:
+            self.text_config = LlamaConfig()
+        else:
+            self.text_config = LlamaConfig(**llm_config)
+
+        # Assign configuration values
+        self.force_image_size = force_image_size
+        self.downsample_ratio = downsample_ratio
+        self.template = template  # TODO move out of here and into the tokenizer
+        self.ps_version = ps_version  # Pixel shuffle version
+        self.image_tag_type = image_tag_type # TODO: into the tokenizer too?
+        self.projector_hidden_size = projector_hidden_size
+        self.vit_hidden_size = vit_hidden_size

vllm/transformers_utils/configs/olmo3.py

@@ -0,0 +1,80 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class Olmo3Config(PretrainedConfig):
+
+    model_type = "olmo3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=50304,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        use_cache=True,
+        pad_token_id=1,
+        bos_token_id=None,
+        eos_token_id=50279,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        rms_norm_eps=1e-5,
+        sliding_window=4096,
+        layer_types=None,
+        **kwargs,
+    ):
+        # This model uses Olmo3ForCausalLM in transformers but Olmo2ForCausalLM
+        # in vLLM.
+        if "architectures" not in kwargs:
+            kwargs["architectures"] = ["Olmo2ForCausalLM"]
+        elif "Olmo3ForCausalLM" in kwargs["architectures"]:
+            kwargs["architectures"].remove("Olmo3ForCausalLM")
+            kwargs["architectures"].append("Olmo2ForCausalLM")
+
+        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,
+        )
+        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
+
+        # 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.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.rms_norm_eps = rms_norm_eps
+
+        self.sliding_window = sliding_window
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention" if (i + 1) % 4 != 0 else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]

vllm/transformers_utils/configs/ovis.py

@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# yapf: disable
+# ruff: noqa: E501
+# adapted from https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/configuration_aimv2.py
+# and https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/configuration_ovis.py
+# Ovis Config with AimV2 config registration removed for Transformers compatibility
+from typing import Any, Optional, Union
+
+from transformers import AutoConfig, PretrainedConfig
+
+
+class AIMv2Config(PretrainedConfig):
+    """This is the configuration class to store the configuration of an [`AIMv2Model`].
+    Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
+    Args:
+        hidden_size: Dimension of the hidden representations.
+        intermediate_size: Dimension of the SwiGLU representations.
+        num_hidden_layers: Number of hidden layers in the Transformer.
+        num_attention_heads: Number of attention heads for each attention layer
+            in the Transformer.
+        num_channels: Number of input channels.
+        image_size: Image size.
+        patch_size: Patch size.
+        rms_norm_eps: Epsilon value used for the RMS normalization layer.
+        attention_dropout: Dropout ratio for attention probabilities.
+        projection_dropout: Dropout ratio for the projection layer after the attention.
+        qkv_bias: Whether to add a bias to the queries, keys and values.
+        use_bias: Whether to add a bias in the feed-forward and projection layers.
+        kwargs: Keyword arguments for the [`PretrainedConfig`].
+    """
+
+    model_type: str = "aimv2"
+
+    def __init__(
+        self,
+        hidden_size: int = 1024,
+        intermediate_size: int = 2816,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 8,
+        num_channels: int = 3,
+        image_size: int = 224,
+        patch_size: int = 14,
+        rms_norm_eps: float = 1e-5,
+        attention_dropout: float = 0.0,
+        projection_dropout: float = 0.0,
+        qkv_bias: bool = False,
+        use_bias: bool = False,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        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.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.rms_norm_eps = rms_norm_eps
+
+        self.projection_dropout = projection_dropout
+        self.qkv_bias = qkv_bias
+        self.use_bias = use_bias
+
+
+# ----------------------------------------------------------------------
+#                     Visual Tokenizer Configuration
+# ----------------------------------------------------------------------
+class BaseVisualTokenizerConfig(PretrainedConfig):
+
+    def __init__(self,
+                 vocab_size=16384,
+                 tokenize_function="softmax",
+                 tau=1.0,
+                 depths=None,
+                 drop_cls_token=False,
+                 backbone_config: Optional[Union[PretrainedConfig,
+                                                 dict]] = None,
+                 hidden_stride: int = 1,
+                 **kwargs):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.tokenize_function = tokenize_function
+        self.tau = tau
+        if isinstance(depths, str):
+            depths = [int(x) for x in depths.split('|')]
+        self.depths = depths
+        self.backbone_kwargs = dict[str, Any]()
+        self.drop_cls_token = drop_cls_token
+        if backbone_config is not None:
+            assert isinstance(backbone_config, (PretrainedConfig, dict)), \
+                f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"
+            if not isinstance(backbone_config, PretrainedConfig):
+                model_type = backbone_config['model_type']
+                if model_type != "aimv2":
+                    backbone_config.pop('model_type')
+                    backbone_config = AutoConfig.for_model(model_type, **backbone_config)
+                else:
+                    backbone_config = AIMv2Config(**backbone_config)
+        self.backbone_config = backbone_config
+        self.hidden_stride = hidden_stride
+
+
+class Aimv2VisualTokenizerConfig(BaseVisualTokenizerConfig):
+    model_type = "aimv2_visual_tokenizer"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        if self.drop_cls_token:
+            self.drop_cls_token = False
+        if self.depths:
+            assert len(self.depths) == 1
+            self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
+
+
+class SiglipVisualTokenizerConfig(BaseVisualTokenizerConfig):
+    model_type = "siglip_visual_tokenizer"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        if self.drop_cls_token:
+            self.drop_cls_token = False
+        if self.depths:
+            assert len(self.depths) == 1
+            self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
+
+
+AutoConfig.register("siglip_visual_tokenizer", SiglipVisualTokenizerConfig)
+AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)
+
+
+# ----------------------------------------------------------------------
+#                           Ovis Configuration
+# ----------------------------------------------------------------------
+class OvisConfig(PretrainedConfig):
+    model_type = "ovis"
+
+    def __init__(self,
+                 llm_config: Optional[Union[PretrainedConfig, dict]] = None,
+                 visual_tokenizer_config: Optional[Union[PretrainedConfig,
+                                                         dict]] = None,
+                 multimodal_max_length=8192,
+                 hidden_size=None,
+                 conversation_formatter_class=None,
+                 llm_attn_implementation=None,
+                 disable_tie_weight=False,
+                 **kwargs):
+        super().__init__(**kwargs)
+        if llm_config is not None:
+            assert isinstance(llm_config, (PretrainedConfig, dict)), \
+                f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"
+            if not isinstance(llm_config, PretrainedConfig):
+                model_type = llm_config['model_type']
+                llm_config.pop('model_type')
+                llm_config = AutoConfig.for_model(model_type, **llm_config)
+
+        # map llm_config to text_config
+        self.text_config = llm_config
+        if visual_tokenizer_config is not None:
+            assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), \
+                f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
+            if not isinstance(visual_tokenizer_config, PretrainedConfig):
+                model_type = visual_tokenizer_config['model_type']
+                visual_tokenizer_config.pop('model_type')
+                visual_tokenizer_config = AutoConfig.for_model(
+                    model_type, **visual_tokenizer_config)
+
+        self.visual_tokenizer_config = visual_tokenizer_config
+        self.multimodal_max_length = multimodal_max_length
+        self.hidden_size = hidden_size
+        self.conversation_formatter_class = conversation_formatter_class
+        self.llm_attn_implementation = llm_attn_implementation
+        self.disable_tie_weight = disable_tie_weight

vllm/transformers_utils/configs/qwen3_next.py

@@ -0,0 +1,275 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright 2025 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.
+"""Qwen3-Next model configuration"""
+
+from transformers.configuration_utils import (PretrainedConfig,
+                                              layer_type_validation)
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen3NextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3NextModel`]. It is used to instantiate a
+    Qwen3-Next model according to the specified arguments, defining the model architecture.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of
+    Qwen3-Next-80B-A3B-Instruct [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct).
+
+    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 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids`.
+        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 48):
+            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 2):
+            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`, *optional*, defaults to `"silu"`):
+            The non-linear activation function 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.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        partial_rotary_factor (`float`, *optional*, defaults to 0.25):
+            Percentage of the query and keys which will have rotary embedding.
+        attention_bias (`bool`, *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.
+        head_dim (`int`, *optional*, defaults to 256):
+            Projection weights dimension in multi-head attention.
+        linear_conv_kernel_dim (`int`, *optional*, defaults to 4):
+            Kernel size of the convolution used in linear attention layers.
+        linear_key_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each key head in linear attention.
+        linear_value_head_dim (`int`, *optional*, defaults to 128):
+            Dimension of each value head in linear attention.
+        linear_num_key_heads (`int`, *optional*, defaults to 16):
+            Number of key heads used in linear attention layers.
+        linear_num_value_heads (`int`, *optional*, defaults to 32):
+            Number of value heads used in linear attention layers.
+        decoder_sparse_step (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer.
+        moe_intermediate_size (`int`, *optional*, defaults to 512):
+            Intermediate size of the routed expert.
+        shared_expert_intermediate_size (`int`, *optional*, defaults to 512):
+            Intermediate size of the shared expert.
+        num_experts_per_tok (`int`, *optional*, defaults to 10):
+            Number of selected experts.
+        num_experts (`int`, *optional*, defaults to 512):
+            Number of routed experts.
+        norm_topk_prob (`bool`, *optional*, defaults to `True`):
+            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. Enabling 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.
+        mlp_only_layers (`list[int]`, *optional*, defaults to `[]`):
+            Indicate which layers use Qwen3NextMLP rather than Qwen3NextSparseMoeBlock
+            The list contains layer index, from 0 to num_layers-1 if we have num_layers layers
+            If `mlp_only_layers` is empty, `decoder_sparse_step` is used to determine the sparsity.
+        layer_types (`list[str]`, *optional*):
+            Types of each layer (attention or linear).
+
+    ```python
+    >>> from transformers import Qwen3NextModel, Qwen3NextConfig
+
+    >>> # Initializing a Qwen3Next style configuration
+    >>> configuration =  Qwen3NextConfig()
+
+    >>> # Initializing a model from the Qwen3-Next-80B-A3B style configuration
+    >>> model = Qwen3NextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """  # noqa: E501
+
+    model_type = "qwen3_next"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.experts.*.gate_proj": "colwise",
+        "layers.*.mlp.experts.*.up_proj": "colwise",
+        "layers.*.mlp.experts.*.down_proj": "rowwise",
+        "layers.*.mlp.shared_experts.gate_proj": "colwise",
+        "layers.*.mlp.shared_experts.up_proj": "colwise",
+        "layers.*.mlp.shared_experts.down_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=2048,
+        intermediate_size=5632,
+        num_hidden_layers=48,
+        num_attention_heads=16,
+        num_key_value_heads=2,
+        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,
+        rope_scaling=None,
+        partial_rotary_factor=0.25,
+        attention_bias=False,
+        attention_dropout=0.0,
+        head_dim=256,
+        linear_conv_kernel_dim=4,
+        linear_key_head_dim=128,
+        linear_value_head_dim=128,
+        linear_num_key_heads=16,
+        linear_num_value_heads=32,
+        decoder_sparse_step=1,
+        moe_intermediate_size=512,
+        shared_expert_intermediate_size=512,
+        num_experts_per_tok=10,
+        num_experts=512,
+        norm_topk_prob=True,
+        output_router_logits=False,
+        router_aux_loss_coef=0.001,
+        mlp_only_layers=None,
+        layer_types=None,
+        **kwargs,
+    ):
+        if mlp_only_layers is None:
+            mlp_only_layers = []
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **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.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.rope_scaling = rope_scaling
+        self.partial_rotary_factor = partial_rotary_factor
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.head_dim = head_dim
+        rope_config_validation(self)
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "linear_attention" if bool((i + 1) % 4) else "full_attention"
+                for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        # linear attention part
+        self.linear_conv_kernel_dim = linear_conv_kernel_dim
+        self.linear_key_head_dim = linear_key_head_dim
+        self.linear_value_head_dim = linear_value_head_dim
+        self.linear_num_key_heads = linear_num_key_heads
+        self.linear_num_value_heads = linear_num_value_heads
+
+        # 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
+        self.mlp_only_layers = mlp_only_layers
+
+
+__all__ = ["Qwen3NextConfig"]

vllm/transformers_utils/configs/radio.py

@@ -0,0 +1,91 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Radio vision model configuration"""
+
+from typing import Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+VIT_TIMM_DIM_BY_NAME: dict[str, tuple[int, int, int, int]] = {
+    "vit_small_patch16_224": (384, 12, 6, 1536),
+    "vit_base_patch16_224": (768, 12, 12, 3072),
+    "vit_large_patch16_224": (1024, 24, 16, 4096),
+    "vit_huge_patch16_224": (1280, 32, 16, 5120),
+}
+
+OPENAI_CLIP_MEAN = (0.48145466, 0.4578275, 0.40821073)
+OPENAI_CLIP_STD = (0.26862954, 0.26130258, 0.27577711)
+
+
+class RadioConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a Radio
+    vision model. It is used to instantiate a Radio model according to the
+    specified arguments, defining the model architecture.
+
+    Args:
+        model_name: Name of the vision transformer model
+            (e.g., "vit_base_patch16_224"). Used to determine architecture
+            dimensions from `VIT_TIMM_DIM_BY_NAME`.
+        image_size: The size (resolution) of each image.
+        patch_size: The size (resolution) of each patch.
+        qkv_bias: Whether to add a bias to the queries, keys and values.
+        qk_normalization: Whether to apply normalization to queries and keys.
+        norm_type: The normalization type to use.
+        layer_norm_eps: The epsilon used by the layer normalization layers.
+        initializer_factor: A factor for initializing all weight matrices.
+        hidden_act: The non-linear activation function in the encoder.
+        max_img_size: Maximum image size for position embeddings.
+        norm_mean: Mean values for image normalization (RGB channels).
+            Defaults to (0.48145466, 0.4578275, 0.40821073)).
+        norm_std: Standard deviation values for image normalization
+            (RGB channels). Defaults to (0.26862954, 0.26130258, 0.27577711)).
+        reg_tokens: Number of register tokens to use.
+    """
+
+    model_type = "radio"
+
+    def __init__(
+        self,
+        model_name: str,
+        image_size: int = 224,
+        patch_size: int = 16,
+        qkv_bias: bool = True,
+        qk_normalization: bool = False,
+        norm_type: str = "layer_norm",
+        layer_norm_eps: float = 1e-6,
+        initializer_factor: float = 1.0,
+        hidden_act: str = "gelu",
+        max_img_size: int = 2048,
+        norm_mean: Union[tuple[float, float, float], list] = OPENAI_CLIP_MEAN,
+        norm_std: Union[tuple[float, float, float], list] = OPENAI_CLIP_STD,
+        reg_tokens: Optional[int] = None,
+        **kwargs,
+    ):
+        self.model_name = model_name
+        (
+            self.hidden_size,
+            self.num_hidden_layers,
+            self.num_attention_heads,
+            self.intermediate_size,
+        ) = VIT_TIMM_DIM_BY_NAME[model_name]
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.qkv_bias = qkv_bias
+        self.qk_normalization = qk_normalization
+        self.norm_type = norm_type
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_factor = initializer_factor
+        self.hidden_act = hidden_act
+        self.max_img_size = max_img_size
+        self.norm_mean = list(norm_mean) if isinstance(norm_mean,
+                                                       (tuple,
+                                                        list)) else norm_mean
+        self.norm_std = list(norm_std) if isinstance(norm_std,
+                                                     (tuple,
+                                                      list)) else norm_std
+        self.reg_tokens = reg_tokens
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/speculators/__init__.py

@@ -0,0 +1,2 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

vllm/transformers_utils/configs/speculators/algos.py

@@ -0,0 +1,32 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+SUPPORTED_SPECULATORS_TYPES = {}
+
+
+def register_speculator(name):
+
+    def decorator(fn):
+        SUPPORTED_SPECULATORS_TYPES[name] = fn
+        return fn
+
+    return decorator
+
+
+@register_speculator("eagle3")
+def update_eagle3(config_dict: dict, vllm_config: dict) -> None:
+    """
+    Apply Eagle-3 specific configuration transformations.
+    
+    Eagle-3 specific fields:
+    - draft_vocab_size: Size of the draft model's vocabulary
+    - target_hidden_size: Hidden size of the target model
+    - norm_before_residual: Whether to apply norm before residual connection
+    """
+
+    vllm_config["draft_vocab_size"] = config_dict.get("draft_vocab_size")
+    if config_dict.get("target_hidden_size") is not None:
+        vllm_config["target_hidden_size"] = config_dict["target_hidden_size"]
+    vllm_config["norm_before_residual"] = config_dict.get(
+        "norm_before_residual", True)
+    vllm_config["architectures"] = ["Eagle3LlamaForCausalLM"]

vllm/transformers_utils/configs/speculators/base.py

@@ -0,0 +1,111 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
+from typing import Any, Union
+
+from transformers import PretrainedConfig
+
+from vllm.transformers_utils.configs.speculators.algos import (
+    SUPPORTED_SPECULATORS_TYPES)
+
+__all__ = ["SpeculatorsConfig"]
+
+
+class SpeculatorsConfig(PretrainedConfig):
+    model_type = "speculators"
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        **kwargs,
+    ) -> "SpeculatorsConfig":
+        """Load speculators Eagle config and convert to vLLM format."""
+        config_dict, _ = cls.get_config_dict(pretrained_model_name_or_path,
+                                             **kwargs)
+
+        vllm_config = cls.extract_vllm_speculative_config(config_dict)
+        return cls(**vllm_config)
+
+    @classmethod
+    def extract_vllm_speculative_config(
+            cls, config_dict: dict[str, Any]) -> dict[str, Any]:
+        speculators_model_type = config_dict.get("speculators_model_type")
+        if speculators_model_type not in SUPPORTED_SPECULATORS_TYPES:
+            raise ValueError(
+                f"Expected one of: {SUPPORTED_SPECULATORS_TYPES}. "
+                "Please ensure you're loading a speculators-format model.")
+
+        # validate fields
+        # TODO: @dsikka - use speculators pydantic model to validate
+        cls.validate_speculators_config(config_dict=config_dict)
+        # Convert from speculators config -> format that can be ingested by vLLM
+        vllm_config = cls.build_vllm_speculative_config(
+            config_dict=config_dict)
+        # Apply anything specific to the supported algorithm
+        algo_updater = SUPPORTED_SPECULATORS_TYPES[speculators_model_type]
+        algo_updater(config_dict=config_dict, vllm_config=vllm_config)
+        return vllm_config
+
+    @classmethod
+    def validate_speculators_config(cls, config_dict: dict[str, Any]) -> None:
+        try:
+            spec_config = config_dict["speculators_config"]
+            methods = spec_config["proposal_methods"]
+            first_method = methods[0]
+            _ = first_method["speculative_tokens"]
+            _ = spec_config["verifier"]["name_or_path"]
+            _ = config_dict["speculators_model_type"]
+        except (KeyError, IndexError, TypeError) as e:
+            raise ValueError("Invalid speculators config structure") from e
+
+        if "transformer_layer_config" not in config_dict:
+            raise ValueError("Must provide transformer_layer_config")
+
+        if not isinstance(config_dict["transformer_layer_config"], dict):
+            raise TypeError(
+                "'transformer_layer_config' must be a dictionary if provided")
+
+    @classmethod
+    def build_vllm_speculative_config(
+            cls, config_dict: dict[str, Any]) -> dict[str, Any]:
+        """
+        Build vLLM-compatible speculative configuration from speculators format.
+
+        This method extracts and transforms speculative configuration from the
+        speculators format into the structure expected by vLLM.
+
+        Args:
+            config_dict: Configuration dictionary in speculators format
+
+        Returns:
+            Dictionary with vLLM-compatible speculative configuration
+        """
+        # Extract speculators configuration
+        spec_config = config_dict["speculators_config"]
+
+        # Currently we only support one proposal method
+        proposal_methods = spec_config.get("proposal_methods")
+        if not proposal_methods:
+            raise ValueError("No proposal methods found in speculators config")
+
+        first_method = proposal_methods[0]
+        num_speculative_tokens = first_method.get("speculative_tokens")
+
+        if num_speculative_tokens is None:
+            raise ValueError(
+                "Missing 'speculative_tokens' in proposal method. "
+                f"Got: {first_method}")
+
+        # Build base vLLM speculative configuration
+        vllm_config = {
+            "method": config_dict.get("speculators_model_type"),
+            "num_speculative_tokens": num_speculative_tokens,
+            "target_model": spec_config.get("verifier")["name_or_path"]
+        }
+
+        # Merge transformer layer configuration if present
+        transformer_config = config_dict.get("transformer_layer_config", {})
+        vllm_config.update(transformer_config)
+
+        return vllm_config

vllm/transformers_utils/configs/step3_vl.py

@@ -0,0 +1,123 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class Step3VisionEncoderConfig(PretrainedConfig):
+    model_type = "step3_vision_encoder"
+
+    def __init__(
+        self,
+        hidden_size=1792,
+        intermediate_size=3072,
+        output_hidden_size=4096,
+        num_hidden_layers=63,
+        num_attention_heads=16,
+        num_channels=3,
+        image_size=728,
+        patch_size=14,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.output_hidden_size = output_hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        super().__init__(**kwargs)
+
+
+class Step3TextConfig(PretrainedConfig):
+    model_type = "step3_text"
+    architectures = ["Step3TextForCausalLM"]
+
+    def __init__(
+        self,
+        hidden_size: int = 7168,
+        intermediate_size: int = 18432,
+        num_attention_heads: int = 64,
+        num_attention_groups: int = 1,
+        num_hidden_layers: int = 61,
+        max_seq_len: int = 65536,
+        vocab_size: int = 128815,
+        rms_norm_eps: float = 1e-5,
+        moe_intermediate_size: int = 5120,
+        moe_num_experts: int = 48,
+        moe_top_k: int = 3,
+        rope_theta: float = 500000,
+        rope_scaling: Optional[dict[str, Any]] = None,
+        max_position_embedding: int = 65536,
+        share_expert_dim: int = 5120,
+        share_q_dim: int = 2048,
+        head_dim: int = 256,
+        norm_expert_weight: bool = False,
+        moe_layers_enum: tuple[int,
+                               ...] = (4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+                                       15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
+                                       25, 26, 27, 28, 29, 30, 31, 32, 33, 34,
+                                       35, 36, 37, 38, 39, 40, 41, 42, 43, 44,
+                                       45, 46, 47, 48, 49, 50, 51, 52, 53, 54,
+                                       55, 56, 57, 58, 59),
+        **kwargs,
+    ) -> None:
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_groups = num_attention_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.rms_norm_eps = rms_norm_eps
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.max_position_embedding = max_position_embedding
+        self.share_expert_dim = share_expert_dim
+        self.share_q_dim = share_q_dim
+        self.head_dim = head_dim
+        self.norm_expert_weight = norm_expert_weight
+        self.moe_layers_enum = moe_layers_enum
+
+        super().__init__(**kwargs)
+
+
+class Step3VLConfig(PretrainedConfig):
+    model_type = "step3_vl"
+
+    def __init__(
+        self,
+        vision_config: Optional[Union[dict, Step3VisionEncoderConfig]] = None,
+        text_config: Optional[Union[dict, Step3TextConfig]] = None,
+        understand_projector_stride: int = 1,
+        projector_bias: bool = True,
+        image_token_id: int = 128001,
+        **kwargs,
+    ) -> None:
+        if vision_config is None:
+            vision_config = Step3VisionEncoderConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = Step3VisionEncoderConfig(**vision_config)
+        self.vision_config = vision_config
+
+        if text_config is None:
+            text_config = Step3TextConfig()
+        elif isinstance(text_config, dict):
+            text_config = Step3TextConfig(**text_config)
+        self.text_config = text_config
+
+        self.understand_projector_stride = understand_projector_stride
+        self.projector_bias = projector_bias
+        self.hidden_size = text_config.hidden_size
+        self.image_token_id = image_token_id
+
+        super().__init__(**kwargs)

vllm/transformers_utils/configs/ultravox.py

@@ -0,0 +1,116 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Adapted from https://github.com/fixie-ai/ultravox/blob/ecd58c4041030bae2ad15aa6bcf04ab43199ea02/ultravox/model/ultravox_config.py
+from typing import Any, Optional
+
+import transformers
+
+
+class UltravoxConfig(transformers.PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a
+    [`UltravoxForConditionalGeneration`]. It is used to instantiate an
+    Ultravox 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:
+        audio_config (`Union[AutoConfig, dict]`,  *optional*):
+            Custom audio config or dict.
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object of the text backbone.
+        audio_model_id (`str`, *optional*):
+            The model ID of the audio backbone.
+        text_model_id (`str`, *optional*):
+            The model ID of the text backbone.
+        ignore_index (`int`, *optional*, defaults to -100):
+            The ignore index for the loss function.
+        audio_token_index (`int`, *optional*, defaults to 32000):
+            The audio token index to encode the audio prompt.
+        stack_factor (`int`, *optional*, defaults to 8):
+            Audio downsampling factor for the multimodal projector.
+        norm_init (`float`, *optional*, defaults to 0.4):
+            The initialization value for the layer normalization.
+        projector_act (`str`, *optional*, defaults to `"swiglu"`):
+            The activation function used by the multimodal projector.
+        projector_ln_mid (`bool`, *optional*, defaults to `False`):
+            Whether to apply layer normalization at the middle of the
+            projector or at the end. Versions v0.4.1 and below
+            use `False`, but v0.5 and above use `True`.
+    """
+    wrapped_model_config: transformers.PretrainedConfig
+    model_type = "ultravox"
+    audio_token = "<|audio|>"
+    is_composition = False
+
+    def __init__(
+        self,
+        audio_config: Optional[dict[str, Any]] = None,
+        text_config: Optional[dict[str, Any]] = None,
+        audio_model_id: Optional[str] = None,
+        text_model_id: Optional[str] = None,
+        ignore_index: int = -100,
+        audio_token_index: int = 32000,
+        hidden_size: int = 4096,
+        stack_factor: int = 8,
+        norm_init: float = 0.4,
+        projector_act: str = "swiglu",
+        projector_ln_mid: bool = False,
+        **kwargs,
+    ):
+        self.ignore_index = ignore_index
+        self.audio_token_index = audio_token_index
+
+        self.hidden_size = hidden_size
+        self.stack_factor = stack_factor
+        self.norm_init = norm_init
+        self.projector_act = projector_act
+        self.projector_ln_mid = projector_ln_mid
+
+        # N.B. May set the wrapped_model_config below.
+        self.text_model_id = text_model_id
+        if text_model_id is None:
+            text_config = text_config or {}
+            self.wrapped_model_config = transformers.CONFIG_MAPPING[
+                text_config.get("model_type", "llama")](**text_config)
+
+        # N.B. May set the audio_config below.
+        self.audio_model_id = audio_model_id
+        if audio_model_id is None:
+            self.audio_model_id = None
+            audio_config = audio_config or {}
+            self.audio_config = transformers.CONFIG_MAPPING[audio_config.get(
+                "model_type", "whisper")](**audio_config)
+
+        super().__init__(**kwargs)
+
+    def __setattr__(self, key, value):
+        # Since --hf-overrides are applied _after_ the UltravoxConfig is
+        # instantiated, load the configs implicitly when assigning text_model_id
+        # or audio_model_id. This allows:
+        #
+        #   --hf-overrides.text_model_id=<quantized variant>
+        #
+        # to behave as intended.
+        if key == "text_model_id" and value is not None:
+            from vllm.transformers_utils.config import get_config
+
+            self.wrapped_model_config = get_config(value,
+                                                   trust_remote_code=False)
+        elif key == "audio_model_id" and value is not None:
+            from vllm.transformers_utils.config import get_config
+
+            self.audio_config = get_config(value, trust_remote_code=False)
+
+        return super().__setattr__(key, value)
+
+    @property
+    def text_config(self) -> transformers.PretrainedConfig:
+        # When Ultravox wraps a multi-modal model (e.g. Gemma), we instantiate
+        # the full model, but the text config is the text config of the inner
+        # model.
+        return self.wrapped_model_config.get_text_config()