sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

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)