sglang 0.5.3rc0__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/configs/nemotron_h.py

@@ -0,0 +1,286 @@
+# 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.
+# ==============================================================================
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/configs/nemotron_h.py
+
+"""NemotronH model configuration"""
+
+import regex as re
+from transformers.configuration_utils import PretrainedConfig
+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
+
+logger = logging.get_logger(__name__)
+
+MAMBA = "M"
+ATTENTION = "*"
+MLP = "-"
+
+
+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.mamba_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.mamba_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 mamba_layer_ids(self):
+        return [
+            i
+            for i in range(self.num_hidden_layers)
+            if self.hybrid_override_pattern[i] == MAMBA
+        ]
+
+    @property
+    def full_attention_layer_ids(self):
+        return [
+            i
+            for i in range(self.num_hidden_layers)
+            if self.hybrid_override_pattern[i] == ATTENTION
+        ]
+
+    @property
+    def mamba2_cache_params(self) -> Mamba2CacheParams:
+        shape = Mamba2StateShape.create(
+            tp_world_size=get_attention_tp_size(),
+            intermediate_size=self.mamba_num_heads * self.mamba_head_dim,
+            n_groups=self.n_groups,
+            num_heads=self.mamba_num_heads,
+            head_dim=self.mamba_head_dim,
+            state_size=self.ssm_state_size,
+            conv_kernel=self.conv_kernel,
+        )
+
+        return Mamba2CacheParams(shape=shape, layers=self.mamba_layer_ids)

sglang/srt/configs/olmo3.py

@@ -0,0 +1,105 @@
+# 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.
+"""Olmo3 model configuration"""
+
+import enum
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class Olmo3LayerType(enum.Enum):
+    full_attention = "full_attention"
+    sliding_attention = "sliding_attention"
+
+
+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 sglang.
+        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
+        rope_config_validation(self)
+        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)
+            ]

sglang/srt/configs/points_v15_chat.py

@@ -0,0 +1,29 @@
+from typing import Optional, Union
+
+from transformers import PretrainedConfig, Qwen2Config
+from transformers.models.qwen2_vl.configuration_qwen2_vl import Qwen2VLVisionConfig
+
+
+class POINTSV15ChatConfig(PretrainedConfig):
+    model_type = "pointsv1.5_chat"
+
+    def __init__(
+        self,
+        vision_config: Optional[Union[dict, Qwen2VLVisionConfig]] = None,
+        llm_config: Optional[Union[dict, Qwen2Config]] = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if vision_config is None:
+            vision_config = Qwen2VLVisionConfig()
+        elif isinstance(vision_config, dict):
+            vision_config = Qwen2VLVisionConfig(**vision_config)
+        self.vision_config = vision_config
+
+        if llm_config is None:
+            llm_config = Qwen2Config()
+        elif isinstance(llm_config, dict):
+            llm_config = Qwen2Config(**llm_config)
+
+        self.llm_config = llm_config
+        self.hidden_size = self.llm_config.hidden_size

sglang/srt/configs/qwen3_next.py

@@ -15,26 +15,20 @@
 """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.configs.mamba_utils import Mamba2CacheParams, Mamba2StateShape
 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):
@@ -282,45 +276,15 @@ class Qwen3NextConfig(PretrainedConfig):
         ]
 
     @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
+    def mamba2_cache_params(self) -> Mamba2CacheParams:
+        shape = Mamba2StateShape.create(
+            tp_world_size=get_attention_tp_size(),
+            intermediate_size=self.linear_value_head_dim * self.linear_num_value_heads,
+            n_groups=self.linear_num_key_heads,
+            num_heads=self.linear_num_value_heads,
+            head_dim=self.linear_value_head_dim,
+            state_size=self.linear_key_head_dim,
+            conv_kernel=self.linear_conv_kernel_dim,
         )
-        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
+        return Mamba2CacheParams(shape=shape, layers=self.linear_layer_ids)