sglang 0.5.2rc2__py3-none-any.whl → 0.5.3.post1__py3-none-any.whl

sglang/srt/configs/dots_ocr.py

@@ -0,0 +1,64 @@
+from typing import Optional
+
+from transformers import AutoProcessor, Qwen2_5_VLProcessor
+from transformers.image_processing_utils import BaseImageProcessor
+from transformers.models.qwen2 import Qwen2Config
+
+from sglang.srt.configs.dots_vlm import DotsVisionConfig
+
+
+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)
+
+
+class DummyVideoProcessor(BaseImageProcessor):
+    model_input_names = ["pixel_values"]
+
+    def __call__(self, *args, **kwargs):
+        return None
+
+
+class DotsVLProcessor(Qwen2_5_VLProcessor):
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        video_processor=None,
+        chat_template=None,
+        **kwargs
+    ):
+        if video_processor is None:
+            video_processor = DummyVideoProcessor()
+        super().__init__(
+            image_processor, tokenizer, video_processor, chat_template=chat_template
+        )
+        self.image_token = (
+            "<|imgpad|>"
+            if not hasattr(tokenizer, "image_token")
+            else tokenizer.image_token
+        )
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None) is not None
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+
+
+AutoProcessor.register(DotsOCRConfig, DotsVLProcessor)

sglang/srt/configs/dots_vlm.py

@@ -0,0 +1,139 @@
+from typing import Any, List, Optional, Union
+
+from transformers import AutoProcessor, LlamaTokenizerFast, PretrainedConfig
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessingKwargs, Unpack
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+try:
+    from transformers import Qwen2_5_VLProcessor
+except ImportError:
+    raise ImportError(
+        "Qwen2_5_VLProcessor can not be found. Please upgrade your transformers version."
+    )
+
+from sglang.srt.configs.deepseekvl2 import DeepseekV2Config
+
+
+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",  # "eager","sdpa","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,
+    ):
+        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 DotsVLMConfig(PretrainedConfig):
+    model_type = "dots_vlm"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        vision_config = kwargs.get("vision_config", {})
+        self.im_span_id = kwargs.get("image_token_id", 128815)
+        self.video_span_id = kwargs.get("video_token_id", 128836)
+        self.vision_config = DotsVisionConfig(**vision_config)
+        self.language_config = DeepseekV2Config(**kwargs)
+        self.architectures = ["DotsVLMForCausalLM"]
+
+
+class DotsVLMProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+    }
+
+
+class DotsVLMProcessor(Qwen2_5_VLProcessor):
+    r"""
+    Constructs a DotsVLM processor which derives from Qwen2_5_VLProcessor, but overrides the image and video token ids.
+    Besides, its tokenizer is a LlamaTokenizerFast instead of Qwen2TokenizerFast.
+    [`DotsVLMProcessor`] offers all the functionalities of [`DotsVisionConfig`] and [`LlamaTokenizerFast`]. See the
+    [`~DotsVLMProcessor.__call__`] and [`~DotsVLMProcessor.decode`] for more information.
+    Args:
+        image_processor ([`Qwen2VLImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`LlamaTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+
+    valid_kwargs = ["chat_template"]
+
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+
+    def __init__(
+        self, image_processor=None, tokenizer=None, chat_template=None, **kwargs
+    ):
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+        self.image_token = (
+            "<|imgpad|>"
+            if not hasattr(tokenizer, "image_token")
+            else tokenizer.image_token
+        )
+        self.video_token = (
+            "<|video_pad|>"
+            if not hasattr(tokenizer, "video_token")
+            else tokenizer.video_token
+        )
+        self.img_token = (
+            "<|img|>" if not hasattr(tokenizer, "img_token") else tokenizer.img_token
+        )
+        self.endofimg_token = (
+            "<|endofimg|>"
+            if not hasattr(tokenizer, "endofimg_token")
+            else tokenizer.endofimg_token
+        )
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.encode(self.image_token)[0]
+        )
+        self.video_token_id = (
+            tokenizer.video_token_id
+            if getattr(tokenizer, "video_token_id", None)
+            else tokenizer.encode(self.video_token)[0]
+        )
+
+
+AutoProcessor.register(DotsVLMConfig, DotsVLMProcessor)

sglang/srt/configs/falcon_h1.py

@@ -0,0 +1,314 @@
+# coding=utf-8
+# Copyright 2024 TII 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.
+"""Falcon-H1 model configuration"""
+
+import enum
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+from sglang.srt.configs.mamba_utils import Mamba2CacheParams, Mamba2StateShape
+from sglang.srt.layers.dp_attention import (
+    get_attention_tp_size,
+    get_tensor_model_parallel_world_size,
+)
+
+logger = logging.get_logger(__name__)
+
+
+class FalconH1Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`FalconH1Model`]. It is used to instantiate a
+    FalconH1Model model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with defaults taken from [ibm-fms/FalconH1-9.8b-2.2T-hf](https://huggingface.co/ibm-fms/FalconH1-9.8b-2.2T-hf).
+    The FalconH1Model is a hybrid [mamba2](https://github.com/state-spaces/mamba) architecture with SwiGLU.
+    The checkpoints are  jointly trained by IBM, Princeton, and UIUC.
+    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 128000):
+            Vocabulary size of the FalconH1 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`FalconH1Model`]
+        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 a output word embedding layer.
+        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, check out [this
+            paper](https://huggingface.co/papers/2305.13245). 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.
+        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`.
+        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. Default is 1 because only the
+            logits of the last prompt token are needed for generation. For long sequences, the logits for the entire
+            sequence may use a lot of memory so, setting `num_logits_to_keep=1` will reduce memory footprint
+            significantly.
+        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.
+        max_position_embeddings (`int`, *optional*, defaults to 8192):
+            Max cached sequence length for the model
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        mamba_d_ssm (`int`, *optional*, defaults to 1024):
+            The dimension of the SSM state space latents.
+        mamba_n_heads (`int`, *optional*, defaults to 128):
+            The number of mamba heads used in the v2 implementation.
+        mamba_d_head (`int`, *optional*, defaults to `"auto"`):
+            Head embedding dimension size
+        mamba_n_groups (`int`, *optional*, defaults to 1):
+            The number of the mamba groups used in the v2 implementation.
+        mamba_d_state (`int`, *optional*, defaults to 256):
+            The dimension the mamba state space latents
+        mamba_d_conv (`int`, *optional*, defaults to 4):
+            The size of the mamba convolution kernel
+        mamba_expand (`int`, *optional*, defaults to 2):
+            Expanding factor (relative to hidden_size) used to determine the mamba intermediate size
+        mamba_chunk_size (`int`, *optional*, defaults to 256):
+            The chunks in which to break the sequence when doing prefill/training
+        mamba_conv_bias (`bool`, *optional*, defaults to `True`):
+            Flag indicating whether or not to use bias in the convolution layer of the mamba mixer block.
+        mamba_proj_bias (`bool`, *optional*, defaults to `False`):
+            Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the mamba mixer block
+        mamba_norm_before_gate (`bool`, *optional*, defaults to `True`):
+            Whether to use RMSNorm before the gate in the Mamba block
+        mamba_rms_norm (`bool`, *optional*, defaults to `False`):
+            Whether to use RMSNorm instead of LayerNorm in the Mamba block
+        projectors_bias (`bool`, *optional*, defaults to `False`):
+            Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the attention block
+        rope_theta (`float`, *optional*, defaults to 100000.0):
+            The theta value used for the RoPE embeddings.
+        rope_scaling (`float`, *optional*):
+            The scaling value used for the RoPE embeddings. If `None`, no scaling is applied.
+        lm_head_multiplier (`float`, *optional*, defaults to 1.0):
+            The multiplier for the LM head. This is used to scale the output of the LM head.
+        embedding_multiplier (`float`, *optional*, defaults to 1.0):
+            The multiplier for the embedding layer. This is used to scale the output of the embedding layer.
+        mlp_multipliers (`list[float]`, *optional*):
+            The multipliers for the MLP layers. This is used to scale the output of the MLP layers. The first value is
+            the multiplier of gate layer, the second value is the multiplier of the down_proj layer.
+        key_multiplier (`float`, *optional*):
+            The multiplier for the key layer. This is used to scale the output of the key layer.
+        attention_out_multiplier (`float`, *optional*):
+            The multiplier for the attention output layer. This is used to scale the output of the attention output
+        attention_in_multiplier (`float`, *optional*):
+            The multiplier for the attention input layer. This is used to scale the output of the attention input layer.
+        ssm_multipliers (`list[float]`, *optional*):
+            The multipliers for the SSM layers. This is used to scale the output of the SSM layers.
+        ssm_in_multiplier (`float`, *optional*):
+            The multiplier for the SSM input layer. This is used to scale the output of the SSM input layer.
+        ssm_out_multiplier (`float`, *optional*):
+            The multiplier for the SSM output layer. This is used to scale the output of the SSM output layer.
+    """
+
+    model_type = "falcon_h1"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=128000,
+        tie_word_embeddings=False,
+        hidden_size=4096,
+        intermediate_size=14336,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        num_logits_to_keep=1,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        max_position_embeddings=8192,
+        attention_dropout=0.0,
+        mamba_d_ssm=1024,
+        mamba_n_heads=128,
+        mamba_d_head="auto",
+        mamba_n_groups=1,
+        mamba_d_state=256,
+        mamba_d_conv=4,
+        mamba_expand=2,
+        mamba_chunk_size=256,
+        mamba_conv_bias=True,
+        mamba_proj_bias=False,
+        mamba_norm_before_gate=True,
+        mamba_rms_norm=False,
+        projectors_bias=False,
+        rope_theta=100000.0,
+        rope_scaling=None,
+        lm_head_multiplier=1.0,
+        embedding_multiplier=1.0,
+        mlp_multipliers=None,
+        key_multiplier=None,
+        attention_out_multiplier=None,
+        attention_in_multiplier=None,
+        ssm_multipliers=None,
+        ssm_in_multiplier=None,
+        ssm_out_multiplier=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        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.max_position_embeddings = max_position_embeddings
+        self.attention_dropout = attention_dropout
+        self.attention_bias = False
+        self.mlp_bias = False
+
+        # 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.num_logits_to_keep = num_logits_to_keep
+
+        self.rope_theta = rope_theta
+        self.rope_scaling = None
+        self.rope_scaling = rope_scaling
+        self.projectors_bias = projectors_bias
+        self.mamba_intermediate = mamba_intermediate = (
+            mamba_expand * hidden_size if mamba_d_ssm is None else mamba_d_ssm
+        )
+
+        if mamba_intermediate % mamba_n_heads != 0:
+            raise ValueError("mamba_n_heads must divide mamba_expand * hidden_size")
+
+        # for the mamba_v2, must satisfy the following
+        if mamba_d_head == "auto":
+            mamba_d_head = mamba_intermediate // mamba_n_heads
+
+        if mamba_d_head * mamba_n_heads != mamba_intermediate:
+            raise ValueError(
+                "The dimensions for the Mamba head state do not match the model intermediate_size"
+            )
+
+        self.mamba_d_ssm = mamba_d_ssm
+        self.mamba_n_heads = mamba_n_heads
+        self.mamba_d_head = mamba_d_head
+        self.mamba_n_groups = mamba_n_groups
+        self.mamba_d_state = mamba_d_state
+        self.mamba_d_conv = mamba_d_conv
+        self.mamba_expand = mamba_expand
+        self.mamba_chunk_size = mamba_chunk_size
+        self.mamba_conv_bias = mamba_conv_bias
+        self.mamba_proj_bias = mamba_proj_bias
+
+        self.mamba_norm_before_gate = mamba_norm_before_gate
+        self.mamba_rms_norm = mamba_rms_norm
+
+        self.lm_head_multiplier = lm_head_multiplier
+        self.embedding_multiplier = embedding_multiplier
+
+        if mlp_multipliers is not None:
+            self.mlp_multipliers = mlp_multipliers
+        else:
+            self.mlp_multipliers = [1.0, 1.0]
+
+        if attention_out_multiplier is not None:
+            self.attention_out_multiplier = attention_out_multiplier
+        else:
+            self.attention_out_multiplier = 1.0
+
+        if attention_in_multiplier is not None:
+            self.attention_in_multiplier = attention_in_multiplier
+        else:
+            self.attention_in_multiplier = 1.0
+
+        if key_multiplier is not None:
+            self.key_multiplier = key_multiplier
+        else:
+            self.key_multiplier = 1.0
+
+        if ssm_multipliers is not None:
+            self.ssm_multipliers = ssm_multipliers
+        else:
+            self.ssm_multipliers = [1.0, 1.0, 1.0, 1.0, 1.0]
+
+        if ssm_in_multiplier is not None:
+            self.ssm_in_multiplier = ssm_in_multiplier
+        else:
+            self.ssm_in_multiplier = 1.0
+
+        if ssm_out_multiplier is not None:
+            self.ssm_out_multiplier = ssm_out_multiplier
+        else:
+            self.ssm_out_multiplier = 1.0
+
+        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 ["falcon_h1" for i in range(self.num_hidden_layers)]
+
+    @property
+    def full_attention_layer_ids(self):
+        # For Falcon-H1, we do have attention on all layers
+        return range(self.num_hidden_layers)
+
+    @property
+    def linear_layer_ids(self):
+        # For Falcon-H1, we do have mamba on all layers
+        return range(self.num_hidden_layers)
+
+    @property
+    def mamba2_cache_params(self):
+        shape = Mamba2StateShape.create(
+            tp_world_size=get_tensor_model_parallel_world_size(),
+            intermediate_size=self.mamba_intermediate,
+            n_groups=self.mamba_n_groups,
+            num_heads=self.mamba_n_heads,
+            head_dim=self.mamba_d_head,
+            state_size=self.mamba_d_state,
+            conv_kernel=self.mamba_d_conv,
+        )
+        return Mamba2CacheParams(shape=shape, layers=self.linear_layer_ids)

sglang/srt/configs/load_config.py

@@ -23,6 +23,9 @@ class LoadFormat(str, enum.Enum):
     LAYERED = "layered"
     JAX = "jax"
     REMOTE = "remote"
+    REMOTE_INSTANCE = "remote_instance"
+    RDMA = "rdma"
+    LOCAL_CACHED = "local_cached"
 
 
 @dataclass
@@ -46,6 +49,7 @@ class LoadConfig:
         checkpoints.
     decryption_key_file: If set, decrypts the output files with a password read
         from this file (after PBKDF2).
+    decrypt_max_concurrency: The maximum number of concurrent processes to decrypt the safetensor files. -1 means no limit.
     """
 
     load_format: Union[str, LoadFormat] = LoadFormat.AUTO
@@ -53,6 +57,11 @@ class LoadConfig:
     model_loader_extra_config: Optional[Union[str, dict]] = field(default_factory=dict)
     ignore_patterns: Optional[Union[List[str], str]] = None
     decryption_key_file: Optional[str] = None
+    decrypt_max_concurrency: int = -1
+    tp_rank: Optional[int] = None
+    remote_instance_weight_loader_seed_instance_ip: Optional[str] = None
+    remote_instance_weight_loader_seed_instance_service_port: Optional[int] = None
+    remote_instance_weight_loader_send_weights_group_ports: Optional[List[int]] = None
 
     def __post_init__(self):
         model_loader_extra_config = self.model_loader_extra_config or {}

sglang/srt/configs/mamba_utils.py

@@ -0,0 +1,117 @@
+# Copyright 2025 SGLang Team
+# 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.
+"""Common config utils for mamba2 - NemotronH, FalconH1, Qwen3Next, etc."""
+
+import os
+from dataclasses import dataclass, field
+
+import numpy as np
+import torch
+
+from sglang.srt.distributed.utils import divide
+
+
+def extra_groups_for_head_shards(ngroups: int, tp_size: int):
+    """Compute the increase in group numbers to account for
+    replication in order to accompany the head shards."""
+
+    # in the case ngoups % tp_size == 0, this will be zero
+    if ngroups % tp_size == 0:
+        return 0
+
+    # for n_groups == 1, this is exactly tp_size - n_groups
+    return tp_size - ngroups
+
+
+@dataclass(kw_only=True, frozen=True)
+class Mamba2StateShape:
+    conv: tuple[int, int]
+    temporal: tuple[int, int, int]
+
+    intermediate_size: int
+    conv_dim: int
+    ssm_state_size: int
+    num_heads: int
+    head_dim: int
+    state_size: int
+    conv_kernel: int
+
+    @staticmethod
+    def create(
+        *,
+        tp_world_size: int,
+        intermediate_size: int,
+        n_groups: int,
+        num_heads: int,
+        head_dim: int,
+        state_size: int,
+        conv_kernel: int,
+    ) -> "Mamba2StateShape":
+        # if n_groups is not divisible by world_size, need to extend the shards
+        # to ensure all groups needed by a head is sharded along with it
+        if n_groups % tp_world_size != 0:
+            extra_groups = extra_groups_for_head_shards(n_groups, tp_world_size)
+            n_groups += extra_groups
+        # heads and n_groups are TP-ed
+        conv_dim = intermediate_size + 2 * n_groups * state_size
+
+        # contiguous along 'dim' axis
+        conv_state_shape = divide(conv_dim, tp_world_size), conv_kernel - 1
+
+        # These are not TP-ed as they depend on A, dt_bias, D
+        # - they are typically small
+        #   e.g., (h_heads, head_dim, state_size) = (128, 64, 128)
+        temporal_state_shape = (divide(num_heads, tp_world_size), head_dim, state_size)
+        return Mamba2StateShape(
+            conv=conv_state_shape,
+            temporal=temporal_state_shape,
+            intermediate_size=intermediate_size,
+            conv_dim=conv_dim,
+            ssm_state_size=state_size,
+            num_heads=num_heads,
+            head_dim=head_dim,
+            state_size=state_size,
+            conv_kernel=conv_kernel,
+        )
+
+
+@dataclass(kw_only=True, frozen=True)
+class Mamba2StateDType:
+    conv: torch.dtype
+    temporal: torch.dtype
+
+
+CONV_DTYPE = torch.bfloat16
+
+
+def mamba2_state_dtype() -> Mamba2StateDType:
+    dtype_map = {
+        "float32": torch.float32,
+        "bfloat16": torch.bfloat16,
+    }
+    ssm_dtype = dtype_map[os.environ["SGLANG_MAMBA_SSM_DTYPE"]]
+    return Mamba2StateDType(conv=CONV_DTYPE, temporal=ssm_dtype)
+
+
+@dataclass(kw_only=True, frozen=True)
+class Mamba2CacheParams:
+    shape: Mamba2StateShape
+    dtype: Mamba2StateDType = field(default_factory=mamba2_state_dtype)
+    layers: list[int]
+
+    @property
+    def mamba_cache_per_req(self) -> int:
+        return (
+            int(np.prod(self.shape.conv)) * self.dtype.conv.itemsize
+            + int(np.prod(self.shape.temporal)) * self.dtype.temporal.itemsize
+        ) * len(self.layers)