sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/configs/model_config.py

@@ -64,12 +64,28 @@ class ModelConfig:
         is_draft_model: bool = False,
         hybrid_kvcache_ratio: Optional[float] = None,
         model_impl: Union[str, ModelImpl] = ModelImpl.AUTO,
+        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,
     ) -> None:
         # Parse args
         self.model_path = model_path
         self.revision = revision
         self.quantization = quantization
         self.model_impl = model_impl
+        self.tp_rank = tp_rank
+        self.remote_instance_weight_loader_seed_instance_ip = (
+            remote_instance_weight_loader_seed_instance_ip
+        )
+        self.remote_instance_weight_loader_seed_instance_service_port = (
+            remote_instance_weight_loader_seed_instance_service_port
+        )
+        self.remote_instance_weight_loader_send_weights_group_ports = (
+            remote_instance_weight_loader_send_weights_group_ports
+        )
 
         self.maybe_pull_model_tokenizer_from_remote()
         self.model_override_args = json.loads(model_override_args)
@@ -141,12 +157,21 @@ class ModelConfig:
 
         if is_draft_model and self.hf_config.architectures[0] == "MiMoForCausalLM":
             self.hf_config.architectures[0] = "MiMoMTP"
+        if is_draft_model and self.hf_config.architectures[0] in [
+            "BailingMoeV2ForCausalLM",
+            "BailingMoeForCausalLM",
+        ]:
+            self.hf_config.architectures[0] = "BailingMoeForCausalLMNextN"
         if (
             is_draft_model
             and self.hf_config.architectures[0] == "Ernie4_5_MoeForCausalLM"
         ):
             self.hf_config.architectures[0] = "Ernie4_5_MoeForCausalLMMTP"
 
+        if is_draft_model and self.hf_config.architectures[0] == "Qwen3NextForCausalLM":
+            self.hf_config.architectures[0] = "Qwen3NextForCausalLMMTP"
+            self.hf_config.num_nextn_predict_layers = 1
+
         # Check model type
         self.is_generation = is_generation_model(
             self.hf_config.architectures, is_embedding
@@ -208,6 +233,7 @@ class ModelConfig:
             or "DeepseekV3ForCausalLMNextN" in self.hf_config.architectures
             or "LongcatFlashForCausalLM" in self.hf_config.architectures
             or "LongcatFlashForCausalLMNextN" in self.hf_config.architectures
+            or "DotsVLMForCausalLM" in self.hf_config.architectures
         ):
             self.head_dim = 256
             self.attention_arch = AttentionArch.MLA
@@ -302,11 +328,16 @@ class ModelConfig:
         ) or getattr(self.hf_config, "image_token_index", None)
 
     @staticmethod
-    def from_server_args(server_args: ServerArgs, model_path: str = None, **kwargs):
+    def from_server_args(
+        server_args: ServerArgs,
+        model_path: str = None,
+        model_revision: str = None,
+        **kwargs,
+    ):
         return ModelConfig(
             model_path=model_path or server_args.model_path,
             trust_remote_code=server_args.trust_remote_code,
-            revision=server_args.revision,
+            revision=model_revision or server_args.revision,
             context_length=server_args.context_length,
             model_override_args=server_args.json_model_override_args,
             is_embedding=server_args.is_embedding,
@@ -315,6 +346,9 @@ class ModelConfig:
             quantization=server_args.quantization,
             hybrid_kvcache_ratio=server_args.hybrid_kvcache_ratio,
             model_impl=server_args.model_impl,
+            remote_instance_weight_loader_seed_instance_ip=server_args.remote_instance_weight_loader_seed_instance_ip,
+            remote_instance_weight_loader_seed_instance_service_port=server_args.remote_instance_weight_loader_seed_instance_service_port,
+            remote_instance_weight_loader_send_weights_group_ports=server_args.remote_instance_weight_loader_send_weights_group_ports,
             **kwargs,
         )
 
@@ -412,11 +446,20 @@ class ModelConfig:
             is_local = os.path.exists(self.model_path)
             modelopt_quant_config = {"quant_method": "modelopt"}
             if not is_local:
-                from huggingface_hub import HfApi
+                import huggingface_hub
+
+                try:
+                    from huggingface_hub import HfApi
+
+                    hf_api = HfApi()
+                    if hf_api.file_exists(self.model_path, "hf_quant_config.json"):
+                        quant_cfg = modelopt_quant_config
+                except huggingface_hub.errors.OfflineModeIsEnabled:
+                    logger.warning(
+                        "Offline mode is enabled, skipping hf_quant_config.json check"
+                    )
+                    pass
 
-                hf_api = HfApi()
-                if hf_api.file_exists(self.model_path, "hf_quant_config.json"):
-                    quant_cfg = modelopt_quant_config
             elif os.path.exists(os.path.join(self.model_path, "hf_quant_config.json")):
                 quant_config_file = os.path.join(
                     self.model_path, "hf_quant_config.json"
@@ -712,6 +755,7 @@ multimodal_model_archs = [
     "Phi4MMForCausalLM",
     "VILAForConditionalGeneration",
     "Step3VLForConditionalGeneration",
+    "DotsVLMForCausalLM",
 ]
 
 

sglang/srt/configs/qwen3_next.py

@@ -0,0 +1,326 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen3Hybrid model configuration"""
+
+import enum
+import os
+
+import numpy as np
+import torch
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+from sglang.srt.distributed.utils import divide
+from sglang.srt.layers.dp_attention import get_attention_tp_size
+
+logger = logging.get_logger(__name__)
+
+
+# NOTE: HybridLayerType
+class HybridLayerType(enum.Enum):
+    full_attention = "attention"
+    swa_attention = "swa_attention"
+    linear_attention = "linear_attention"
+    mamba2 = "mamba"
+
+
+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*, defaults to None):
+            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
+    ```
+    """
+
+    model_type = "qwen3_next"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    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=[],
+        layer_types=None,
+        **kwargs,
+    ):
+        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)
+
+        # linear attention (gdn now 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
+
+    @property
+    def layers_block_type(self):
+        layer_type_list = []
+
+        for l in range(self.num_hidden_layers):
+            if (l + 1) % self.full_attention_interval == 0:
+                layer_type_list.append(HybridLayerType.full_attention.value)
+            else:
+                layer_type_list.append(HybridLayerType.linear_attention.value)
+
+        return layer_type_list
+
+    @property
+    def linear_layer_ids(self):
+        return [
+            i
+            for i, type_value in enumerate(self.layers_block_type)
+            if type_value == HybridLayerType.linear_attention.value
+        ]
+
+    @property
+    def full_attention_layer_ids(self):
+        return [
+            i
+            for i, type_value in enumerate(self.layers_block_type)
+            if type_value == HybridLayerType.full_attention.value
+        ]
+
+    @property
+    def hybrid_gdn_params(self):
+        world_size = get_attention_tp_size()
+        conv_dim = (
+            self.linear_key_head_dim * self.linear_num_key_heads * 2
+            + self.linear_value_head_dim * self.linear_num_value_heads
+        )
+        conv_state_shape = (
+            divide(conv_dim, world_size),
+            self.linear_conv_kernel_dim - 1,
+        )
+
+        temporal_state_shape = (
+            divide(self.linear_num_value_heads, world_size),
+            self.linear_key_head_dim,
+            self.linear_value_head_dim,
+        )
+        conv_dtype = torch.bfloat16
+        dtype_map = {
+            "float32": torch.float32,
+            "bfloat16": torch.bfloat16,
+        }
+        ssm_dtype = dtype_map[os.environ["SGLANG_MAMBA_SSM_DTYPE"]]
+        mamba_layers = self.linear_layer_ids
+        return (
+            conv_state_shape,
+            temporal_state_shape,
+            conv_dtype,
+            ssm_dtype,
+            mamba_layers,
+        )
+
+    @property
+    def mamba_cache_per_req(self):
+        conv_state_shape, temporal_state_shape, conv_dtype, ssm_dtype, mamba_layers = (
+            self.hybrid_gdn_params
+        )
+        mamba_layers_len = len(mamba_layers)
+
+        return (
+            int(np.prod(conv_state_shape)) * conv_dtype.itemsize
+            + int(np.prod(temporal_state_shape)) * ssm_dtype.itemsize
+        ) * mamba_layers_len

sglang/srt/connector/__init__.py

@@ -9,6 +9,7 @@ from sglang.srt.connector.base_connector import (
     BaseKVConnector,
 )
 from sglang.srt.connector.redis import RedisConnector
+from sglang.srt.connector.remote_instance import RemoteInstanceConnector
 from sglang.srt.connector.s3 import S3Connector
 from sglang.srt.utils import parse_connector_type
 
@@ -18,14 +19,17 @@ logger = logging.getLogger(__name__)
 class ConnectorType(str, enum.Enum):
     FS = "filesystem"
     KV = "KV"
+    INSTANCE = "instance"
 
 
-def create_remote_connector(url, **kwargs) -> BaseConnector:
+def create_remote_connector(url, device, **kwargs) -> BaseConnector:
     connector_type = parse_connector_type(url)
     if connector_type == "redis":
         return RedisConnector(url)
     elif connector_type == "s3":
         return S3Connector(url)
+    elif connector_type == "instance":
+        return RemoteInstanceConnector(url, device)
     else:
         raise ValueError(f"Invalid connector type: {url}")
 
@@ -35,6 +39,8 @@ def get_connector_type(client: BaseConnector) -> ConnectorType:
         return ConnectorType.KV
     if isinstance(client, BaseFileConnector):
         return ConnectorType.FS
+    if isinstance(client, RemoteInstanceConnector):
+        return ConnectorType.INSTANCE
 
     raise ValueError(f"Invalid connector type: {client}")
 
@@ -44,6 +50,7 @@ __all__ = [
     "BaseFileConnector",
     "BaseKVConnector",
     "RedisConnector",
+    "RemoteInstanceConnector",
     "S3Connector",
     "ConnectorType",
     "create_remote_connector",

sglang/srt/connector/remote_instance.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import logging
+from typing import Generator, List, Optional, Tuple
+from urllib.parse import urlparse
+
+import torch
+import torch.distributed as dist
+
+from sglang.srt.connector import BaseConnector
+from sglang.srt.utils import init_custom_process_group
+
+logger = logging.getLogger(__name__)
+
+
+class RemoteInstanceConnector(BaseConnector):
+
+    def __init__(self, url: str, device: torch.device = "cpu"):
+        assert (
+            device.type == "cuda"
+        ), "RemoteInstanceConnector only supports cuda device."
+        super().__init__(url)
+        self.url = url
+        self.device = device
+
+    def build_group(
+        self,
+        gpu_id: int = -1,
+        tp_rank: int = -1,
+        instance_ip: str = None,
+        group_rank: int = 1,
+        world_size: int = 2,
+    ):
+        assert (
+            self.device.type == "cuda"
+        ), "RemoteInstanceConnector only supports cuda device."
+        assert (
+            gpu_id != -1 and tp_rank != -1
+        ), "gpu_id and tp_rank must be specified for RemoteInstanceConnector. "
+
+        self.device_id = torch.device(self.device.type, gpu_id)
+
+        parsed_url = urlparse(self.url)
+        master_address = parsed_url.hostname
+        master_port = parsed_url.port
+        group_name = f"send_weights_{instance_ip}_{master_port}_{tp_rank}"
+        backend = "nccl"
+
+        logger.info(
+            f"init custom process group: master_address={master_address}, master_port={master_port}, "
+            f"rank_offset={group_rank}, world_size={world_size}, group_name={group_name}, backend={backend}"
+        )
+
+        try:
+            self._model_update_group = init_custom_process_group(
+                backend=backend,
+                init_method=f"tcp://{master_address}:{master_port}",
+                world_size=world_size,
+                rank=group_rank,
+                group_name=group_name,
+                device_id=self.device_id,
+            )
+            dist.barrier(group=self._model_update_group)
+            return True, "Succeeded to initialize custom process group."
+        except Exception as e:
+            message = f"Failed to initialize custom process group: {e}."
+            logger.error(message)
+            return False, message
+
+    # Implemented as a no-op to make BaseConnector interface consistent.
+    def pull_files(
+        self,
+        allow_pattern: Optional[list[str]] = None,
+        ignore_pattern: Optional[list[str]] = None,
+    ) -> None:
+        return
+
+    # Implemented as a no-op to make BaseConnector interface consistent.
+    def weight_iterator(
+        self, rank: int = 0
+    ) -> Generator[Tuple[str, torch.Tensor], None, None]:
+        return