diff --git a/python/sglang/srt/elastic_ep/expert_backup_client.py b/python/sglang/srt/elastic_ep/expert_backup_client.py
index 1b82477ba17e..0f217aca6591 100644
--- a/python/sglang/srt/elastic_ep/expert_backup_client.py
+++ b/python/sglang/srt/elastic_ep/expert_backup_client.py
@@ -14,8 +14,7 @@
 from sglang.srt.environ import envs
 from sglang.srt.eplb.expert_location import get_global_expert_location_metadata
 from sglang.srt.managers.io_struct import UpdateExpertBackupReq, sock_recv, sock_send
-from sglang.srt.runtime_context import get_exec
-from sglang.srt.server_args import ServerArgs
+from sglang.srt.runtime_context import get_exec, get_parallel
 from sglang.srt.utils.network import get_local_ip_auto
 
 PORT_BASE = envs.SGLANG_BACKUP_PORT_BASE.get()
@@ -34,16 +33,14 @@ class ExpertBackupClient:
     def __init__(
         self,
         *,
-        server_args: ServerArgs,
         model_config,
         moe_ep_size: int,
         moe_ep_rank: int,
         get_model: Callable[[], Any],
     ):
         context = zmq.Context(2)
-        self.server_args = server_args
-        self.engine_num = server_args.nnodes
-        self.engine_rank = server_args.node_rank
+        self.engine_num = get_parallel().nnodes
+        self.engine_rank = get_parallel().node_rank
         self.recv_list = [None] * self.engine_num
         self.ready_sockets = [None] * self.engine_num
         self._get_model = get_model
@@ -66,14 +63,14 @@ def __init__(
         for i in range(self.engine_num):
             self.recv_list[i] = context.socket(zmq.SUB)
             self.recv_list[i].connect(
-                f"tcp://{all_ips[i * get_world_size() // server_args.nnodes]}:{PORT_BASE + i * 2 + 1}"
+                f"tcp://{all_ips[i * get_world_size() // get_parallel().nnodes]}:{PORT_BASE + i * 2 + 1}"
             )
             self.recv_list[i].setsockopt(zmq.SUBSCRIBE, b"")
 
             # Synchronization channel to notify the manager when this client is ready.
             self.ready_sockets[i] = context.socket(zmq.PUSH)
             self.ready_sockets[i].connect(
-                f"tcp://{all_ips[i * get_world_size() // server_args.nnodes]}:{PORT_BASE + i * 2}"
+                f"tcp://{all_ips[i * get_world_size() // get_parallel().nnodes]}:{PORT_BASE + i * 2}"
             )
             sock_send(self.ready_sockets[i], UpdateExpertBackupReq())
 
diff --git a/python/sglang/srt/eplb/expert_distribution.py b/python/sglang/srt/eplb/expert_distribution.py
index d0cf374a27a8..9ab510d8eb2c 100644
--- a/python/sglang/srt/eplb/expert_distribution.py
+++ b/python/sglang/srt/eplb/expert_distribution.py
@@ -38,7 +38,6 @@
 import torch
 import torch.distributed
 
-from sglang.srt.arg_groups.overrides import should_report_expert_balancedness
 from sglang.srt.environ import envs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.observability.metrics_collector import (
@@ -50,8 +49,9 @@
 from sglang.srt.runtime_context import (
     get_exec,
     get_schedule,
+    logs_expert_balancedness_to_server_log,
+    reports_expert_balancedness,
 )
-from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import Withable, get_device, get_int_env_var
 
 if TYPE_CHECKING:
@@ -85,7 +85,6 @@ class ExpertDistributionRecorder(ABC):
 
     @staticmethod
     def init_new(
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
@@ -95,9 +94,7 @@ def init_new(
             ), "ExpertLocationMetadata is required for expert distribution recording. One possible"
             "reason is that you are using a model that does not support expert distribution"
             "recording. Try setting `get_model_config_for_expert_location` in your model."
-            return _ExpertDistributionRecorderReal(
-                server_args, expert_location_metadata, rank
-            )
+            return _ExpertDistributionRecorderReal(expert_location_metadata, rank)
         else:
             return _ExpertDistributionRecorderNoop()
 
@@ -160,11 +157,9 @@ class _ExpertDistributionRecorderNoop(ExpertDistributionRecorder):
 class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
     def __init__(
         self,
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
-        self._server_args = server_args
         self._expert_location_metadata = expert_location_metadata
 
         self._recording = False
@@ -172,15 +167,13 @@ def __init__(
         self._current_forward_pass_id = Withable()
         self._current_layer_idx = Withable()
         self._current_debug_name = Withable()
-        self._accumulator = _Accumulator.init_new(
-            server_args, expert_location_metadata, rank
-        )
+        self._accumulator = _Accumulator.init_new(expert_location_metadata, rank)
         self._single_pass_gatherers = {
-            k: _SinglePassGatherer.init_new(server_args, expert_location_metadata, rank)
+            k: _SinglePassGatherer.init_new(expert_location_metadata, rank)
             for k in self._accumulator.get_single_pass_gatherer_keys()
         }
 
-        if should_report_expert_balancedness(server_args):
+        if reports_expert_balancedness():
             logger.info(
                 "ExpertDistributionRecorder auto start record since "
                 f"expert_balancedness_report_mode={get_exec().moe.expert_balancedness_report_mode}"
@@ -329,14 +322,11 @@ def set_global_expert_distribution_recorder(value):
 class _SinglePassGatherer(ABC):
     @staticmethod
     def init_new(
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ) -> _SinglePassGatherer:
         if get_exec().moe.expert_distribution_recorder_mode == "per_token":
-            return _DetailSinglePassGatherer(
-                server_args, expert_location_metadata, rank
-            )
+            return _DetailSinglePassGatherer(expert_location_metadata, rank)
 
         if get_exec().moe.moe_a2a_backend == "mori":
             return _DeepepLowLatencySinglePassGatherer(expert_location_metadata, rank)
@@ -403,7 +393,6 @@ class _DetailSinglePassGatherer(_SinglePassGatherer):
 
     def __init__(
         self,
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
@@ -668,11 +657,10 @@ def _convert_local_to_global_physical_count(
 class _Accumulator(ABC):
     @staticmethod
     def init_new(
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ) -> _Accumulator:
-        return _Accumulator.get_class()(server_args, expert_location_metadata, rank)
+        return _Accumulator.get_class()(expert_location_metadata, rank)
 
     @staticmethod
     def get_class() -> Type[_Accumulator]:
@@ -685,11 +673,9 @@ def get_class() -> Type[_Accumulator]:
 
     def __init__(
         self,
-        server_args: ServerArgs,
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
-        self._server_args = server_args
         self._expert_location_metadata = expert_location_metadata
         self._rank = rank
 
@@ -719,7 +705,7 @@ class _UtilizationRateAccumulatorMixin(_Accumulator):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
-        self._enable = should_report_expert_balancedness(self._server_args)
+        self._enable = reports_expert_balancedness()
 
         if self._enable:
             self.window_sizes = EPLB_BALANCEDNESS_WINDOW_SIZES
@@ -727,7 +713,6 @@ def __init__(self, *args, **kwargs):
             self._reset_server_log_history = True
             self._rank = torch.distributed.get_rank()
             expert_dispatch_cls = resolve_collector_class(
-                self._server_args,
                 STAT_LOGGER_ROLE_EXPERT_DISPATCH,
                 ExpertDispatchCollector,
             )
@@ -777,12 +762,10 @@ def _append_utilization_rate(
                 compute_utilization_rate(gpu_physical_count)
             )
             should_track_history = not math.isclose(
-                self._server_args.eplb_min_rebalancing_utilization_threshold, 1.0
+                get_exec().moe.eplb_min_rebalancing_utilization_threshold, 1.0
             )
 
-            should_log = (
-                self._server_args.should_log_expert_balancedness_to_server_log()
-            )
+            should_log = logs_expert_balancedness_to_server_log()
             outputs["metrics"] = ExpertDistributionMetrics(
                 forward_pass_id=forward_pass_id,
                 eplb_balancedness=utilization_rate_gpu,
@@ -954,7 +937,7 @@ def dump(self, output_mode: _OutputMode):
 
     def _get_global_average_utilization_rate(self):
         if not self._enable or math.isclose(
-            self._server_args.eplb_min_rebalancing_utilization_threshold, 1.0
+            get_exec().moe.eplb_min_rebalancing_utilization_threshold, 1.0
         ):
             return None
 
diff --git a/python/sglang/srt/layers/moe/kt_ep_wrapper.py b/python/sglang/srt/layers/moe/kt_ep_wrapper.py
index 4c5494bf50ad..17c55559520a 100644
--- a/python/sglang/srt/layers/moe/kt_ep_wrapper.py
+++ b/python/sglang/srt/layers/moe/kt_ep_wrapper.py
@@ -14,6 +14,7 @@
 
 from sglang.srt.layers.quantization.base_config import FusedMoEMethodBase
 from sglang.srt.runtime_context import (
+    get_exec,
     get_parallel,
     get_schedule,
 )
@@ -73,7 +74,7 @@ def create_kt_config_from_server_args(
     Returns:
         KTConfig if KT is configured, None otherwise
     """
-    if server_args.kt_weight_path is None:
+    if get_exec().moe.kt_weight_path is None:
         return None
 
     from sglang.srt.arg_groups.overrides import model_config_of
@@ -84,13 +85,13 @@ def create_kt_config_from_server_args(
 
     return KTConfig(
         layer_idx=layer_idx,
-        num_gpu_experts=server_args.kt_num_gpu_experts,
-        cpuinfer_threads=server_args.kt_cpuinfer,
-        threadpool_count=server_args.kt_threadpool_count,
-        weight_path=server_args.kt_weight_path,
+        num_gpu_experts=get_exec().moe.kt_num_gpu_experts,
+        cpuinfer_threads=get_exec().moe.kt_cpuinfer,
+        threadpool_count=get_exec().moe.kt_threadpool_count,
+        weight_path=get_exec().moe.kt_weight_path,
         chunked_prefill_size=get_schedule().chunked_prefill_size,
-        method=server_args.kt_method,
-        max_deferred_experts_per_token=server_args.kt_max_deferred_experts_per_token,
+        method=get_exec().moe.kt_method,
+        max_deferred_experts_per_token=get_exec().moe.kt_max_deferred_experts_per_token,
         num_layers=num_layers,
     )
 
diff --git a/python/sglang/srt/managers/prefill_delayer.py b/python/sglang/srt/managers/prefill_delayer.py
index 198be9a0db18..a0bc19303cf0 100644
--- a/python/sglang/srt/managers/prefill_delayer.py
+++ b/python/sglang/srt/managers/prefill_delayer.py
@@ -78,7 +78,6 @@ def __init__(
         dp_size: int,
         attn_tp_size: int,
         cpu_group,
-        server_args,
         max_delay_passes: int,
         token_usage_low_watermark: Optional[float],
         metrics_collector: Optional["SchedulerMetricsCollector"] = None,
@@ -91,14 +90,14 @@ def __init__(
         self._debug_log_enabled = _DEBUG_LOG and debug_log_enabled
         # Queue-based trigger is opt-in: activates only when queue_min_ratio
         # is explicitly set. Additive with the slot-based trigger.
-        self._queue_min_ratio = server_args.prefill_delayer_queue_min_ratio
+        self._queue_min_ratio = get_schedule().prefill_delayer_queue_min_ratio
         # Fall back to 5000ms if unset; this is a local safety cap, not a
         # semantic default, so we don't surface it via ServerArgs.
-        self._max_delay_ms = server_args.prefill_delayer_max_delay_ms
+        self._max_delay_ms = get_schedule().prefill_delayer_max_delay_ms
         if self._max_delay_ms is None:
             self._max_delay_ms = 5000.0
         self._queue_trigger_enabled = self._queue_min_ratio is not None
-        self._prefill_max_requests = server_args.prefill_max_requests
+        self._prefill_max_requests = get_schedule().prefill_max_requests
         logger.info(
             f"PrefillDelayer initialized with "
             f"max_delay_passes={self._max_delay_passes} "
diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index e9a42dabfea9..e83d6e4ac0a3 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -1300,7 +1300,6 @@ def init_schedule_policy(self):
                     attn_tp_size=self.ps.attn_tp_size,
                     cpu_group=self.tp_cpu_group,
                     device_group=self.tp_group.device_group,
-                    server_args=self.server_args,
                     metrics_collector=(
                         self.metrics_collector
                         if self.metrics_reporter.enable_metrics
@@ -2820,7 +2819,7 @@ def _prefetch_kvcache(self, req: Req):
         if self.enable_hicache_storage:
             req.init_next_round_input(self.tree_cache, cow_mamba=False)
             tree_cache = self.tree_cache
-            buffer_mode = self.server_args.hicache_host_memory_mode == "buffer_only"
+            buffer_mode = get_memory().hicache_host_memory_mode == "buffer_only"
             last_host_node = req.last_host_node
             # Buffer mode host-backups nothing, so match_prefix anchors at
             # root; re-anchor on the deepest device node. The anchor is only
@@ -3544,7 +3543,7 @@ def _get_new_batch_prefill_raw(
             req.init_next_round_input(self.tree_cache)
             if (
                 self.enable_hicache_storage
-                and self.server_args.hicache_host_memory_mode == "buffer_only"
+                and get_memory().hicache_host_memory_mode == "buffer_only"
             ):
                 # Buffer mode: surface a staged prefetch as the request's host
                 # hit (consumed through init_load_back) plus its SWA window,
@@ -4452,7 +4451,7 @@ def is_fully_idle(self, for_health_check=False) -> bool:
                 if tc.enable_storage:
                     idle &= len(tc.ongoing_prefetch) == 0
                     idle &= len(tc.ongoing_backup) == 0
-                    if self.server_args.hicache_host_memory_mode == "buffer_only":
+                    if get_memory().hicache_host_memory_mode == "buffer_only":
                         # Queued writes, staged prefetches, and in-flight
                         # storage writes still hold host staging
                         # (buffer-mode unified tree only).
diff --git a/python/sglang/srt/managers/scheduler_components/metrics_reporter.py b/python/sglang/srt/managers/scheduler_components/metrics_reporter.py
index 1d1f0ddf7c36..cffa87d5079f 100644
--- a/python/sglang/srt/managers/scheduler_components/metrics_reporter.py
+++ b/python/sglang/srt/managers/scheduler_components/metrics_reporter.py
@@ -23,11 +23,13 @@
     compute_routing_key_stats,
 )
 from sglang.srt.runtime_context import (
+    exports_expert_balancedness_to_prometheus,
     get_context,
     get_disagg,
     get_observability,
     get_parallel,
     get_spec,
+    logs_expert_balancedness_to_server_log,
 )
 from sglang.srt.utils.device_timer import DeviceTimer
 from sglang.srt.utils.scheduler_status_logger import SchedulerStatusLogger
@@ -1009,9 +1011,7 @@ def log_batch_result_stats(
         if (m := result.expert_distribution_metrics) is not None:
             balancedness = m.eplb_balancedness.item()
 
-            if (
-                self.scheduler.server_args.should_log_expert_balancedness_to_server_log()
-            ):
+            if logs_expert_balancedness_to_server_log():
                 if m.reset_server_log_history:
                     for history in self._eplb_balancedness_history:
                         history.clear()
@@ -1033,10 +1033,7 @@ def log_batch_result_stats(
                     f"gpu_physical_count_sum={gpu_physical_count_sum}"
                 )
 
-            if (
-                self.enable_metrics
-                and self.scheduler.server_args.should_export_expert_balancedness_to_prometheus()
-            ):
+            if self.enable_metrics and exports_expert_balancedness_to_prometheus():
                 assert self.metrics_collector is not None
                 self.metrics_collector.increment_eplb_balancedness(
                     forward_mode=batch.forward_mode.name.lower(),
diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index f826be496bcd..895484239de5 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -717,7 +717,6 @@ def init_metric_collector_watchdog(self):
             if get_observability().extra_metric_labels:
                 labels.update(get_observability().extra_metric_labels)
             tokenizer_collector_cls = resolve_collector_class(
-                self.server_args,
                 STAT_LOGGER_ROLE_TOKENIZER,
                 TokenizerMetricsCollector,
             )
diff --git a/python/sglang/srt/mem_cache/base_prefix_cache.py b/python/sglang/srt/mem_cache/base_prefix_cache.py
index 15aee4c93ea0..b58a7be4a608 100644
--- a/python/sglang/srt/mem_cache/base_prefix_cache.py
+++ b/python/sglang/srt/mem_cache/base_prefix_cache.py
@@ -243,14 +243,10 @@ class BasePrefixCache(ABC, PrefixCacheTrait):
     kv_events: Optional[KVCacheEventRecorder] = None
 
     def init_metrics_collector(self):
-        from sglang.srt.runtime_context import get_server_args
-
-        server_args = get_server_args()
         labels = {"cache_type": self.__class__.__name__}
         if get_observability().extra_metric_labels:
             labels.update(get_observability().extra_metric_labels)
         radix_cache_cls = resolve_collector_class(
-            server_args,
             STAT_LOGGER_ROLE_RADIX_CACHE,
             RadixCacheMetricsCollector,
         )
diff --git a/python/sglang/srt/mem_cache/hiradix_cache.py b/python/sglang/srt/mem_cache/hiradix_cache.py
index 292a980f0e0a..3bd2d6c84953 100644
--- a/python/sglang/srt/mem_cache/hiradix_cache.py
+++ b/python/sglang/srt/mem_cache/hiradix_cache.py
@@ -344,10 +344,8 @@ def _apply_storage_runtime_config(
                 labels.update(extra_metric_labels)
             existing_collector = getattr(self, "storage_metrics_collector", None)
             if existing_collector is None:
-                from sglang.srt.runtime_context import get_server_args
 
                 storage_cls = resolve_collector_class(
-                    get_server_args(),
                     STAT_LOGGER_ROLE_STORAGE,
                     StorageMetricsCollector,
                 )
diff --git a/python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py b/python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py
index 57e68848ce2b..3ed55425e2d8 100644
--- a/python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py
+++ b/python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py
@@ -327,7 +327,7 @@ def build_kv_only_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
@@ -394,7 +394,7 @@ def build_hybrid_swa_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
@@ -679,7 +679,7 @@ def build_deepseek_v4_hicache_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
@@ -773,7 +773,7 @@ def build_hybrid_mamba_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
@@ -885,7 +885,7 @@ def build_hybrid_mamba_swa_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
@@ -964,7 +964,7 @@ def build_anchor_sidecar_stack(
         storage_backend_extra_config=storage_backend_extra_config,
         transfer_layer_num=transfer_layer_num,
         enable_storage_metrics=enable_storage_metrics,
-        host_memory_mode=server_args.hicache_host_memory_mode,
+        host_memory_mode=get_memory().hicache_host_memory_mode,
     )
     return host_pool_group, cache_controller
 
diff --git a/python/sglang/srt/mem_cache/unified_cache/storage_attachment.py b/python/sglang/srt/mem_cache/unified_cache/storage_attachment.py
index 88ac41d49e98..2700c9b92657 100644
--- a/python/sglang/srt/mem_cache/unified_cache/storage_attachment.py
+++ b/python/sglang/srt/mem_cache/unified_cache/storage_attachment.py
@@ -280,10 +280,8 @@ def _resolve_metrics_collector(
 
         existing_collector = cache.storage_metrics_collector
         if existing_collector is None:
-            from sglang.srt.runtime_context import get_server_args
 
             storage_cls = resolve_collector_class(
-                get_server_args(),
                 STAT_LOGGER_ROLE_STORAGE,
                 StorageMetricsCollector,
             )
diff --git a/python/sglang/srt/model_executor/model_runner.py b/python/sglang/srt/model_executor/model_runner.py
index 80a69f5752ce..0288f44db974 100644
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -504,7 +504,6 @@ def _initialize_elastic_ep_joiner(self) -> None:
         )
         set_global_expert_distribution_recorder(
             ExpertDistributionRecorder.init_new(
-                self.server_args,
                 get_global_expert_location_metadata(),
                 rank=global_ep_rank,
             )
@@ -709,7 +708,6 @@ def maybe_init_expert_location_metadata(self):
             )
         set_global_expert_distribution_recorder(
             ExpertDistributionRecorder.init_new(
-                self.server_args,
                 get_global_expert_location_metadata(),
                 rank=expert_rank,
             )
@@ -750,7 +748,6 @@ def init_token_oracle(self):
     def maybe_init_expert_backup_client(self):
         self.expert_backup_client = (
             ExpertBackupClient(
-                server_args=self.server_args,
                 model_config=self.model_config,
                 moe_ep_size=self.ps.moe_ep_size,
                 moe_ep_rank=self.ps.moe_ep_rank,
@@ -971,7 +968,6 @@ def post_capture_elastic_ep_recover(self):
         )
         set_global_expert_distribution_recorder(
             ExpertDistributionRecorder.init_new(
-                self.server_args,
                 get_global_expert_location_metadata(),
                 rank=global_ep_rank,
             )
@@ -1967,7 +1963,6 @@ def _reset_eplb_after_elastic_scale_failure(self) -> None:
             return
         set_global_expert_distribution_recorder(
             ExpertDistributionRecorder.init_new(
-                self.server_args,
                 get_global_expert_location_metadata(),
                 rank=self._elastic_global_rank(),
             )
@@ -2032,7 +2027,6 @@ def _finalize_scale_up(
         ElasticEPStateManager.on_scale(effective_size, target_size)
         set_global_expert_distribution_recorder(
             ExpertDistributionRecorder.init_new(
-                self.server_args,
                 get_global_expert_location_metadata(),
                 rank=self._elastic_global_rank(),
             )
diff --git a/python/sglang/srt/observability/metrics_collector.py b/python/sglang/srt/observability/metrics_collector.py
index c47d5ff3c06d..a50af6e67165 100644
--- a/python/sglang/srt/observability/metrics_collector.py
+++ b/python/sglang/srt/observability/metrics_collector.py
@@ -27,6 +27,7 @@
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.srt.observability.utils import exponential_buckets, generate_buckets
 from sglang.srt.runtime_context import (
+    exports_expert_balancedness_to_prometheus,
     get_disagg,
     get_observability,
     get_schedule,
@@ -203,15 +204,17 @@ def to_labels(self):
 STAT_LOGGER_ROLE_EXPERT_DISPATCH = "expert_dispatch"
 
 
-def resolve_collector_class(
-    server_args: Optional[ServerArgs], role: str, default_cls: type
-) -> type:
-    """Return the subclass registered for `role` on `server_args.stat_loggers`,
-    or `default_cls` if none is registered. Tolerates `server_args=None` and
-    `stat_loggers=None`."""
-    if server_args is None:
+def resolve_collector_class(role: str, default_cls: type) -> type:
+    """Return the subclass registered for `role` in the published
+    ``observability`` bag, or `default_cls` if none is registered.
+
+    An unpublished ``observability`` namespace answers with the default.
+    """
+    from sglang.srt.runtime_context import get_context, get_observability
+
+    if not get_context().is_config_namespace_published("observability"):
         return default_cls
-    stat_loggers = getattr(server_args, "stat_loggers", None)
+    stat_loggers = get_observability().stat_loggers
     if not stat_loggers:
         return default_cls
     return stat_loggers.get(role, default_cls)
@@ -877,10 +880,7 @@ def __init__(
         # =================================================================
         # Execution
         # =================================================================
-        if (
-            labels["moe_ep_rank"] == 0
-            and server_args.should_export_expert_balancedness_to_prometheus()
-        ):
+        if labels["moe_ep_rank"] == 0 and exports_expert_balancedness_to_prometheus():
             self.eplb_balancedness = Summary(
                 name="sglang:eplb_balancedness",
                 documentation="Balancedness of MoE in expert parallelism.",
@@ -1114,7 +1114,7 @@ def init_new(
             if get_observability().extra_metric_labels:
                 labels.update(get_observability().extra_metric_labels)
             scheduler_collector_cls = resolve_collector_class(
-                server_args, STAT_LOGGER_ROLE_SCHEDULER, cls
+                STAT_LOGGER_ROLE_SCHEDULER, cls
             )
             collector = scheduler_collector_cls(
                 labels=labels,
diff --git a/python/sglang/srt/runtime_context.py b/python/sglang/srt/runtime_context.py
index f62a9c49f743..02e331178b43 100644
--- a/python/sglang/srt/runtime_context.py
+++ b/python/sglang/srt/runtime_context.py
@@ -1819,6 +1819,24 @@ def process_model_config():
     return model_config_of(get_server_args())
 
 
+def reports_expert_balancedness() -> bool:
+    """Whether the expert-balancedness report is on at all.
+
+    `overrides.should_report_expert_balancedness` is the pre-publish equivalent.
+    """
+    return get_exec().moe.expert_balancedness_report_mode != "off"
+
+
+def logs_expert_balancedness_to_server_log() -> bool:
+    """Whether the balancedness report goes to the server log."""
+    return get_exec().moe.expert_balancedness_report_mode in ("server_log", "both")
+
+
+def exports_expert_balancedness_to_prometheus() -> bool:
+    """Whether the balancedness report goes to Prometheus."""
+    return get_exec().moe.expert_balancedness_report_mode in ("prometheus", "both")
+
+
 def cutedsl_moe_max_num_tokens() -> int:
     """The CuteDSL A2A per-rank token budget.
 
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index f90e9067957a..009184acbf2d 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -4341,16 +4341,6 @@ def describe_kv_events_publisher(self) -> Optional[dict]:
             descriptor["load_topic"] = LOAD_TOPIC
         return descriptor
 
-    def should_log_expert_balancedness_to_server_log(self) -> bool:
-        cfg = resolving_view(self)
-
-        return cfg.expert_balancedness_report_mode in ("server_log", "both")
-
-    def should_export_expert_balancedness_to_prometheus(self) -> bool:
-        cfg = resolving_view(self)
-
-        return cfg.expert_balancedness_report_mode in ("prometheus", "both")
-
 
 # --------------------------------------------------------------------------
 # Module-level ServerArgs helpers and runtime shims.
diff --git a/python/sglang/srt/speculative/dspark_components/dspark_planner.py b/python/sglang/srt/speculative/dspark_components/dspark_planner.py
index edde1bbf1b1b..c3a04111bccd 100644
--- a/python/sglang/srt/speculative/dspark_components/dspark_planner.py
+++ b/python/sglang/srt/speculative/dspark_components/dspark_planner.py
@@ -21,7 +21,6 @@
 )
 from sglang.srt.managers.schedule_batch import ScheduleBatch
 from sglang.srt.runtime_context import get_disagg, get_parallel, get_schedule, get_spec
-from sglang.srt.server_args import ServerArgs
 from sglang.srt.speculative.dflash_info_v2 import DFlashDraftInputV2
 from sglang.srt.speculative.dflash_utils import apply_dflash_verify_logits_adjustments
 from sglang.srt.speculative.dspark_components.dspark_sps import (
@@ -76,22 +75,20 @@ def __init__(
         model_runner,
         device,
         tp_rank: int,
-        server_args: ServerArgs,
         verify_num_draft_tokens: int,
     ) -> None:
         self.draft_model = draft_model
         self.gamma = gamma
         self.model_runner = model_runner
         self.device = device
-        self.server_args = server_args
         self.verify_num_draft_tokens = verify_num_draft_tokens
         self._align_verify_tokens_to_graph_tier = (
-            server_args.speculative_dspark_align_verify_tokens_to_graph_tier
+            get_spec().speculative_dspark_align_verify_tokens_to_graph_tier
         )
 
         self._confidence_head = getattr(self.draft_model, "confidence_head", None)
 
-        sts_path = server_args.speculative_dspark_confidence_sts_path
+        sts_path = get_spec().speculative_dspark_confidence_sts_path
         if sts_path and self._confidence_head is not None:
             calibration = load_sts_calibration_from_path(sts_path)
             sts_temperatures = torch.tensor(
@@ -148,7 +145,6 @@ def __init__(
                     f"SGLANG_RAGGED_VERIFY_MODE=static."
                 )
             sps_table = build_sps_cost_table(
-                server_args=self.server_args,
                 verify_num_draft_tokens=self.verify_num_draft_tokens,
             )
             self._is_verify_all = (
@@ -1120,10 +1116,9 @@ def _ensure_carry(self, *, gamma: int) -> None:
 
 def build_sps_cost_table(
     *,
-    server_args: ServerArgs,
     verify_num_draft_tokens: int,
 ) -> Union[SpsCostTable, SpsAdditiveCostTable]:
-    sps_table_path = server_args.speculative_dspark_sps_table_path
+    sps_table_path = get_spec().speculative_dspark_sps_table_path
     if sps_table_path:
         return load_sps_table_from_path(sps_table_path)
     max_batch_tokens = max(
diff --git a/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py b/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py
index 54b7d30bf141..1edb14f97903 100644
--- a/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py
+++ b/python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py
@@ -219,7 +219,6 @@ def __init__(
             model_runner=self.model_runner,
             device=self.device,
             tp_rank=self.ps.tp_rank,
-            server_args=self.server_args,
             verify_num_draft_tokens=self.verify_num_draft_tokens,
         )
         if (
diff --git a/test/registered/scheduler/test_prefill_delayer.py b/test/registered/scheduler/test_prefill_delayer.py
index 08add3fb8c57..b5f89c3d8588 100644
--- a/test/registered/scheduler/test_prefill_delayer.py
+++ b/test/registered/scheduler/test_prefill_delayer.py
@@ -79,20 +79,17 @@ def _run_negotiate_test(rank, test_cases):
 
     for case in test_cases:
         # The DP-attention gate is a published config leaf.
-        override = get_context().override_server_args(enable_dp_attention=True)
+        override = get_context().override_server_args(
+            enable_dp_attention=True,
+            prefill_delayer_queue_min_ratio=case.queue_min_ratio,
+            prefill_delayer_max_delay_ms=case.max_delay_ms,
+            prefill_max_requests=case.prefill_max_requests,
+        )
         override.install()
         delayer = PrefillDelayer(
             dp_size=world_size,
             attn_tp_size=1,
             cpu_group=cpu_group,
-            server_args=SimpleNamespace(
-                enable_dp_attention=True,
-                disaggregation_mode="null",
-                disable_overlap_schedule=False,
-                prefill_delayer_queue_min_ratio=case.queue_min_ratio,
-                prefill_delayer_max_delay_ms=case.max_delay_ms,
-                prefill_max_requests=case.prefill_max_requests,
-            ),
             max_delay_passes=case.max_delay_passes,
             token_usage_low_watermark=case.token_usage_low_watermark,
         )
diff --git a/test/registered/spec/dspark/test_dspark_sps_table.py b/test/registered/spec/dspark/test_dspark_sps_table.py
index d0b23fe27bd3..ba78ea9388d7 100644
--- a/test/registered/spec/dspark/test_dspark_sps_table.py
+++ b/test/registered/spec/dspark/test_dspark_sps_table.py
@@ -143,23 +143,20 @@ def test_profile_honors_explicit_max_batch_tokens(self):
 
 
 def _build_sps_cost_table_for(testcase, *, sps_table_path):
-    from sglang.srt.runtime_context import get_context, get_server_args
+    from sglang.srt.runtime_context import get_context
     from sglang.srt.speculative.dspark_components.dspark_planner import (
         build_sps_cost_table,
     )
 
-    # The table bound reads `max_running_requests` from the published bags, so
-    # the case publishes it; the table path stays on the handed record, which is
-    # what `build_sps_cost_table` takes.
+    # Both the table path and the bound come from the published bags, so the
+    # case publishes them.
     override = get_context().override_server_args(
         speculative_dspark_sps_table_path=sps_table_path,
         max_running_requests=4,
     )
     override.install()
     testcase.addCleanup(override.restore)
-    return build_sps_cost_table(
-        server_args=get_server_args(), verify_num_draft_tokens=5
-    )
+    return build_sps_cost_table(verify_num_draft_tokens=5)
 
 
 class TestBuildSpsCostTableContract(CustomTestCase):
diff --git a/test/registered/unit/observability/test_stat_loggers_di.py b/test/registered/unit/observability/test_stat_loggers_di.py
index ba00b16f5ff7..99a087470729 100644
--- a/test/registered/unit/observability/test_stat_loggers_di.py
+++ b/test/registered/unit/observability/test_stat_loggers_di.py
@@ -39,17 +39,7 @@
     TokenizerMetricsCollector,
     resolve_collector_class,
 )
-
-
-class _StubArgs:
-    """Minimal ServerArgs stand-in.
-
-    Avoids triggering the heavy real ServerArgs import chain for unit-level
-    ``resolve_collector_class`` cases.
-    """
-
-    def __init__(self, stat_loggers=None):
-        self.stat_loggers = stat_loggers
+from sglang.srt.runtime_context import get_context, reset_context
 
 
 class TestCollectorClassAttrs(unittest.TestCase):
@@ -79,30 +69,37 @@ def test_radix_cache_collector_attrs_default_none(self):
 
 
 class TestResolveCollectorClass(unittest.TestCase):
-    def test_returns_default_when_server_args_none(self):
-        cls = resolve_collector_class(None, "scheduler", SchedulerMetricsCollector)
-        self.assertIs(cls, SchedulerMetricsCollector)
+    """The role table is read from the published `observability` bag."""
+
+    def _resolve(self, role, default_cls, **fields):
+        if not fields:
+            return resolve_collector_class(role, default_cls)
+        with get_context().override_server_args(**fields):
+            return resolve_collector_class(role, default_cls)
+
+    def test_returns_default_when_nothing_is_published(self):
+        reset_context()
+        self.assertIs(
+            resolve_collector_class("scheduler", SchedulerMetricsCollector),
+            SchedulerMetricsCollector,
+        )
 
     def test_returns_default_when_stat_loggers_none(self):
-        cls = resolve_collector_class(
-            _StubArgs(stat_loggers=None), "scheduler", SchedulerMetricsCollector
-        )
+        cls = self._resolve("scheduler", SchedulerMetricsCollector, stat_loggers=None)
         self.assertIs(cls, SchedulerMetricsCollector)
 
     def test_returns_default_when_stat_loggers_empty(self):
-        cls = resolve_collector_class(
-            _StubArgs(stat_loggers={}), "scheduler", SchedulerMetricsCollector
-        )
+        cls = self._resolve("scheduler", SchedulerMetricsCollector, stat_loggers={})
         self.assertIs(cls, SchedulerMetricsCollector)
 
     def test_returns_default_when_role_missing(self):
         class MyTokenizer(TokenizerMetricsCollector):
             pass
 
-        cls = resolve_collector_class(
-            _StubArgs(stat_loggers={"tokenizer": MyTokenizer}),
+        cls = self._resolve(
             "scheduler",
             SchedulerMetricsCollector,
+            stat_loggers={"tokenizer": MyTokenizer},
         )
         self.assertIs(cls, SchedulerMetricsCollector)
 
@@ -110,10 +107,10 @@ def test_returns_subclass_when_role_registered(self):
         class MyScheduler(SchedulerMetricsCollector):
             pass
 
-        cls = resolve_collector_class(
-            _StubArgs(stat_loggers={"scheduler": MyScheduler}),
+        cls = self._resolve(
             "scheduler",
             SchedulerMetricsCollector,
+            stat_loggers={"scheduler": MyScheduler},
         )
         self.assertIs(cls, MyScheduler)