Revert RuntimeContext config-namespace reads/roles (#31813–#31817) (#32100)

python/sglang/srt/eplb/eplb_manager.py

@@ -16,7 +16,7 @@ from sglang.srt.eplb.expert_location import (
     get_global_expert_location_metadata,
 )
 from sglang.srt.eplb.expert_location_updater import ExpertLocationUpdater
-from sglang.srt.runtime_context import get_model
+from sglang.srt.runtime_context import get_server_args
 
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
@@ -274,8 +274,8 @@ def update_expert_location_with_recovery(
         else:
             # Load the missing weights from disk
             update_weights_from_disk_callable(
-                get_model().model_path,
-                get_model().load_format,
+                get_server_args().model_path,
+                get_server_args().load_format,
                 weight_name_filter=weight_name_filter,
             )
 

python/sglang/srt/eplb/expert_location_dispatch.py

@@ -18,7 +18,7 @@ from typing import Literal, Optional
 import torch
 
 from sglang.srt.eplb.expert_location import get_global_expert_location_metadata
-from sglang.srt.runtime_context import get_exec
+from sglang.srt.runtime_context import get_server_args
 
 
 @dataclass
@@ -34,7 +34,7 @@ class ExpertLocationDispatchInfo:
 
     @classmethod
     def init_new(cls, layer_id: int):
-        ep_dispatch_algorithm = get_exec().moe.ep_dispatch_algorithm
+        ep_dispatch_algorithm = get_server_args().ep_dispatch_algorithm
         expert_location_metadata = get_global_expert_location_metadata()
         assert expert_location_metadata is not None
 

python/sglang/srt/eplb/expert_location_updater.py

@@ -26,7 +26,7 @@ from sglang.srt.eplb.expert_location import (
     ExpertLocationMetadata,
     get_global_expert_location_metadata,
 )
-from sglang.srt.runtime_context import get_device
+from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import get_bool_env_var
 
 logger = logging.getLogger(__name__)
@@ -107,7 +107,7 @@ def _update_expert_weights_with_canary(
         canary_tensor = (
             _get_canary_value(old_expert_location_metadata, layer_id)
             .clone()
-            .to(device=get_device().device, non_blocking=True)
+            .to(device=get_server_args().device, non_blocking=True)
         )
         routed_experts_weights_of_layer[layer_id].append(canary_tensor)