move all get_stream in sgl_kernel to c++ to reduce the launch overhead (#12521)

python/sglang/srt/configs/falcon_h1.py

@@ -19,7 +19,6 @@ 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_tensor_model_parallel_world_size
 
 logger = logging.get_logger(__name__)
 
@@ -297,8 +296,10 @@ class FalconH1Config(PretrainedConfig):
 
     @property
     def mamba2_cache_params(self):
+        from sglang.srt.layers.dp_attention import get_attention_tp_size
+
         shape = Mamba2StateShape.create(
-            tp_world_size=get_tensor_model_parallel_world_size(),
+            tp_world_size=get_attention_tp_size(),
             intermediate_size=self.mamba_intermediate,
             n_groups=self.mamba_n_groups,
             num_heads=self.mamba_n_heads,

python/sglang/srt/configs/nemotron_h.py

@@ -20,7 +20,6 @@ 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__)
 
@@ -273,6 +272,8 @@ class NemotronHConfig(PretrainedConfig):
 
     @property
     def mamba2_cache_params(self) -> Mamba2CacheParams:
+        from sglang.srt.layers.dp_attention import get_attention_tp_size
+
         shape = Mamba2StateShape.create(
             tp_world_size=get_attention_tp_size(),
             intermediate_size=self.mamba_num_heads * self.mamba_head_dim,

python/sglang/srt/configs/qwen3_next.py

@@ -21,7 +21,6 @@ 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
 
 logger = logging.get_logger(__name__)
 
@@ -277,6 +276,8 @@ class Qwen3NextConfig(PretrainedConfig):
 
     @property
     def mamba2_cache_params(self) -> Mamba2CacheParams:
+        from sglang.srt.layers.dp_attention import get_attention_tp_size
+
         shape = Mamba2StateShape.create(
             tp_world_size=get_attention_tp_size(),
             intermediate_size=self.linear_value_head_dim * self.linear_num_value_heads,