config: project the config bags from the resolution result (#35906)

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

python/sglang/srt/arg_groups/overrides.py

@@ -292,9 +292,61 @@ def declare_late_resolution(server_args: Any, source: str, **fields: Any) -> Non
         log = []
         object.__setattr__(server_args, "_runtime_mutations", log)
     log.append((source, dict(fields)))
+    stash = getattr(server_args, "_resolved_overrides", None)
+    if stash is None:
+        stash = []
+        object.__setattr__(server_args, "_resolved_overrides", stash)
+    stash.append((source, dict(fields)))
     _apply_fields(server_args, fields)
 
 
+def declare_direct_writes(
+    server_args: Any, source: str, resolve: Callable[[Any], Any]
+) -> Any:
+    """Run a resolver that writes the fields directly, and declare what it moved.
+
+    Returns whatever the resolver returned, so a provider with a return value
+    can go through the same capture.
+
+    Out-of-tree platform plugins are handed the record and set fields on it.
+    Their implementations live outside this tree, so they cannot be converted
+    by editing the resolver; and the raw snapshot is taken before the pipeline
+    starts, so a plugin's default is neither declared nor raw.
+
+    Rebinding is what the diff sees, and rebinding is all it needs to see: a
+    plugin that mutates a value in place reaches the projection anyway, because
+    the raw snapshot and the stash entries hold the same object it mutated.
+
+    A stand-in record (tests drive the hooks with a plain namespace) has no
+    fields to diff and no projection to feed, so the resolver runs uncaptured.
+    """
+    if not dataclasses.is_dataclass(server_args):
+        return resolve(server_args)
+    before = {
+        field.name: getattr(server_args, field.name)
+        for field in dataclasses.fields(server_args)
+    }
+    already = len(getattr(server_args, "_resolved_overrides", None) or ())
+    result = resolve(server_args)
+    stash = getattr(server_args, "_resolved_overrides", None)
+    if stash is None:
+        stash = []
+        object.__setattr__(server_args, "_resolved_overrides", stash)
+    # A resolver reached this way can also declare properly -- the in-tree
+    # implementations of these hooks do. Those fields are already explained, and
+    # recording them again would attribute them to the wrapper and bury an
+    # actual direct write among the echoes.
+    declared = {name for _source, fields in stash[already:] for name in fields}
+    changed = {
+        name: getattr(server_args, name)
+        for name, previous in before.items()
+        if name not in declared and getattr(server_args, name) is not previous
+    }
+    if changed:
+        stash.append((source, changed))
+    return result
+
+
 def materialize_declarations(server_args: Any) -> None:
     """Apply the accumulated declarations onto ``server_args`` once, at the
     end of ``__post_init__`` (gate order: last writer wins). After this the
@@ -307,6 +359,27 @@ def materialize_declarations(server_args: Any) -> None:
     server_args._declarations_materialized = True
 
 
+def resolution_result(server_args: Any, field: str, default: Any = None) -> Any:
+    """What resolution decided for ``field``: the declaration if there is one,
+    otherwise what the caller supplied.
+
+    This is what the config projection reads. Reading the field instead would
+    work only for as long as declarations materialize onto the record -- and
+    the point of declaring is that they will not, so the projection must not
+    depend on it. A config that never ran the pipeline (a mock, a partial
+    fixture) carries no raw snapshot; its fields are all it has.
+    """
+    for _source, declared in reversed(
+        getattr(server_args, "_resolved_overrides", None) or ()
+    ):
+        if field in declared:
+            return declared[field]
+    raw = getattr(server_args, "_raw_input", None)
+    if raw is not None and field in raw:
+        return raw[field]
+    return getattr(server_args, field, default)
+
+
 def resolved_view(server_args: Any) -> ResolvedView:
     """Read-only view of the resolving configuration for mid-resolution code
     that is not a pass (``__post_init__`` handlers and hooks). Internal to

python/sglang/srt/arg_groups/speculative_hook.py

@@ -5,7 +5,10 @@ import logging
 import os
 from typing import TYPE_CHECKING, Optional
 
-from sglang.srt.arg_groups.overrides import declare_resolution
+from sglang.srt.arg_groups.overrides import (
+    declare_direct_writes,
+    declare_resolution,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -149,7 +152,11 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
 
         # TODO: move the per-algorithm validation below into spec module hooks.
         if isinstance(algo, CustomSpecAlgo) and algo.validate_server_args is not None:
-            algo.validate_server_args(server_args)
+            declare_direct_writes(
+                server_args,
+                "handle_speculative_decoding.custom_validate",
+                algo.validate_server_args,
+            )
 
     if server_args.speculative_skip_dp_mlp_sync:
         assert server_args.speculative_algorithm == "EAGLE", (
@@ -163,7 +170,13 @@ def handle_speculative_decoding(server_args: ServerArgs) -> None:
             _init_adaptive_speculative_params(server_args)
 
     if algo is not None:
-        algo.handle_server_args(server_args)
+        # A registered algorithm's callback lives outside this tree and sets
+        # fields on the record, so the writes are captured around the call.
+        declare_direct_writes(
+            server_args,
+            "handle_speculative_decoding.custom_algo",
+            algo.handle_server_args,
+        )
 
 
 def _handle_dflash(server_args: ServerArgs) -> None:

python/sglang/srt/disaggregation/common/conn.py

@@ -36,7 +36,12 @@ from sglang.srt.layers.dp_attention import (
     get_attention_dp_rank,
     get_attention_dp_size,
 )
-from sglang.srt.runtime_context import get_parallel, get_serving
+from sglang.srt.runtime_context import (
+    configured_pp_size,
+    get_disagg,
+    get_parallel,
+    get_serving,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils.network import (
     NetworkAddress,
@@ -164,9 +169,9 @@ class CommonKVManager(BaseKVManager):
             envs.SGLANG_DISAGGREGATION_DEFERRED_DECODE_KV_RELEASE.get()
         )
         # for p/d multi node infer
-        self.bootstrap_host = server_args.host
-        self.bootstrap_port = server_args.disaggregation_bootstrap_port
-        self.dist_init_addr = server_args.dist_init_addr
+        self.bootstrap_host = get_serving().host
+        self.bootstrap_port = get_disagg().disaggregation_bootstrap_port
+        self.dist_init_addr = get_parallel().dist_init_addr
         parallel = get_parallel()
         self.attn_tp_size = parallel.attn_tp_size
         self.attn_tp_rank = parallel.attn_tp_rank
@@ -182,7 +187,7 @@ class CommonKVManager(BaseKVManager):
         self.system_dp_rank = (
             self.kv_args.system_dp_rank if self.kv_args.system_dp_rank else 0
         )
-        self.pp_size = server_args.pp_size
+        self.pp_size = configured_pp_size()
         self.pp_rank = self.kv_args.pp_rank
         self.local_ip = get_local_ip_auto()
         cp_sharded_prefill = self.attn_cp_size > 1 and (

python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py

@@ -21,7 +21,11 @@ from sglang.srt.mem_cache.memory_pool import (
     MLATokenToKVPool,
     ReqToTokenPool,
 )
-from sglang.srt.runtime_context import get_schedule
+from sglang.srt.runtime_context import (
+    get_memory,
+    get_schedule,
+    get_serving,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils.common import ceil_align
 
@@ -68,10 +72,10 @@ class DecodeKVCacheOffloadManager:
         self.tp_world_size = torch.distributed.get_world_size(group=self.tp_group)
 
         hicache_storage_backend_extra_config = {}
-        if server_args.hicache_storage_backend_extra_config:
+        if get_memory().hicache_storage_backend_extra_config:
             try:
                 hicache_storage_backend_extra_config = json.loads(
-                    server_args.hicache_storage_backend_extra_config
+                    get_memory().hicache_storage_backend_extra_config
                 )
             except json.JSONDecodeError as e:
                 raise ValueError(
@@ -83,10 +87,10 @@ class DecodeKVCacheOffloadManager:
             mem_pool_host=self.decode_host_mem_pool,
             page_size=self.page_size,
             tp_group=tp_group,
-            io_backend=server_args.hicache_io_backend,
+            io_backend=get_memory().hicache_io_backend,
             load_cache_event=threading.Event(),
-            storage_backend=server_args.hicache_storage_backend,
-            model_name=server_args.served_model_name,
+            storage_backend=get_memory().hicache_storage_backend,
+            model_name=get_serving().served_model_name,
             storage_backend_extra_config=hicache_storage_backend_extra_config,
         )
 

python/sglang/srt/distributed/device_communicators/mooncake_transfer_engine.py

@@ -6,6 +6,11 @@ import os
 from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
 from sglang.srt.environ import envs
+from sglang.srt.runtime_context import (
+    get_disagg,
+    get_exec,
+    get_memory,
+)
 from sglang.srt.utils.network import NetworkAddress, get_free_port, get_local_ip_auto
 
 if TYPE_CHECKING:
@@ -313,27 +318,27 @@ def maybe_init_shared_mooncake_transfer_engine(
     """
     use_mooncake_te = (
         (
-            server_args.disaggregation_mode != "null"
-            and server_args.disaggregation_transfer_backend == "mooncake"
+            get_disagg().disaggregation_mode != "null"
+            and get_disagg().disaggregation_transfer_backend == "mooncake"
         )
         or (
-            server_args.enable_hierarchical_cache
-            and server_args.hicache_storage_backend == "mooncake"
+            get_memory().enable_hierarchical_cache
+            and get_memory().hicache_storage_backend == "mooncake"
             and envs.SGLANG_HICACHE_MOONCAKE_REUSE_TE.get()
         )
         or (
-            server_args.encoder_only
-            and server_args.encoder_transfer_backend == "mooncake"
+            get_disagg().encoder_only
+            and get_disagg().encoder_transfer_backend == "mooncake"
         )
         or (
-            server_args.language_only
-            and server_args.encoder_transfer_backend == "mooncake"
+            get_disagg().language_only
+            and get_disagg().encoder_transfer_backend == "mooncake"
         )
         or (
-            server_args.enable_elastic_expert_backup
-            and server_args.elastic_ep_backend is not None
+            get_exec().moe.enable_elastic_expert_backup
+            and get_exec().moe.elastic_ep_backend is not None
         )
-        or server_args.elastic_ep_backend == "mooncake"
+        or get_exec().moe.elastic_ep_backend == "mooncake"
     )
 
     if use_mooncake_te:
@@ -341,11 +346,12 @@ def maybe_init_shared_mooncake_transfer_engine(
             hostname=get_local_ip_auto(),
             gpu_id=gpu_id,
             ib_device=(
-                server_args.disaggregation_ib_device or server_args.mooncake_ib_device
+                get_disagg().disaggregation_ib_device
+                or get_exec().moe.mooncake_ib_device
             ),
         )
 
-        if server_args.elastic_ep_backend == "mooncake":
+        if get_exec().moe.elastic_ep_backend == "mooncake":
             try:
                 from mooncake.pg import set_transfer_engine
             except ImportError as e:

python/sglang/srt/elastic_ep/elastic_ep.py

@@ -11,7 +11,11 @@ from sglang.srt.distributed import get_world_group, parallel_state
 from sglang.srt.distributed.utils import get_global_tcp_store
 from sglang.srt.eplb.expert_location import broadcast_global_expert_location_metadata
 from sglang.srt.managers.schedule_batch import ServerArgs
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import (
+    configured_tp_size,
+    get_exec,
+    get_parallel,
+)
 from sglang.srt.utils import is_cpu, is_cuda
 
 if TYPE_CHECKING:
@@ -87,9 +91,9 @@ class ElasticEPStateManager:
         if cls._instance is not None:
             return cls._instance
 
-        if server_args.elastic_ep_backend is not None:
+        if get_exec().moe.elastic_ep_backend is not None:
             world_size = torch.distributed.get_world_size()
-            active_rank_capacity = server_args.max_ep_size or world_size
+            active_rank_capacity = get_parallel().max_ep_size or world_size
             assert active_rank_capacity >= world_size, (
                 f"--max-ep-size ({active_rank_capacity}) must be >= "
                 f"world_size ({world_size})."
@@ -103,10 +107,10 @@ class ElasticEPStateManager:
                 inst.snapshot_active_to_last()
                 inst.sync_active_to_cpu()
 
-            if server_args.moe_a2a_backend == "nixl":
+            if get_exec().moe.moe_a2a_backend == "nixl":
                 cls._on_scale = cls._on_scale_nixl
 
-            inst.ep_join_rank_offset = server_args.ep_join_rank_offset
+            inst.ep_join_rank_offset = get_parallel().ep_join_rank_offset
             if server_args.is_ep_joiner:
                 cls._init_joiner_state(inst, server_args)
 
@@ -122,12 +126,13 @@ class ElasticEPStateManager:
         inst.snapshot_active_to_last()
         inst.sync_active_to_cpu()
 
-        if server_args.ep_join_mode == "scale":
+        if get_exec().moe.ep_join_mode == "scale":
             inst.effective_ep_size = (
-                server_args.ep_join_rank_offset + server_args.tp_size
+                get_parallel().ep_join_rank_offset + configured_tp_size()
             )
             inst.original_ep_size = (
-                server_args.elastic_ep_initial_size or server_args.ep_join_rank_offset
+                get_parallel().elastic_ep_initial_size
+                or get_parallel().ep_join_rank_offset
             )
             inst.has_scaled = True
         else:

python/sglang/srt/elastic_ep/expert_backup_manager.py

@@ -17,7 +17,14 @@ from sglang.srt.managers.io_struct import (
 )
 from sglang.srt.model_loader.loader import DefaultModelLoader, get_model_loader
 from sglang.srt.model_loader.utils import set_default_torch_dtype
-from sglang.srt.runtime_context import publish
+from sglang.srt.runtime_context import (
+    configured_tp_size,
+    get_disagg,
+    get_exec,
+    get_model,
+    get_parallel,
+    publish,
+)
 from sglang.srt.server_args import (
     PortArgs,
     ServerArgs,
@@ -38,14 +45,14 @@ def extract_expert_id(param_name):
 
 class ExpertBackupManager:
     def __init__(self, server_args: ServerArgs, port_args: PortArgs):
-        self.load_format = server_args.load_format
+        self.load_format = get_model().load_format
         self.model_config = ModelConfig.from_server_args(server_args)
         self.continuous_buffer = None
         self.weight_pointer_map = {}
         self.transfer_engine = None
         self.session_id = None
-        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.expert_num = self.model_config.hf_config.n_routed_experts
         self.idmn = (self.expert_num // self.engine_num) * self.engine_rank
         self.idmx = (self.expert_num // self.engine_num) * (self.engine_rank + 1)
@@ -53,11 +60,11 @@ class ExpertBackupManager:
         # Synchronization socket to avoid PUB/SUB slow joiner issues.
         self.recv_from_expert_backup_client = context.socket(zmq.PULL)
         self.recv_from_expert_backup_client.bind(
-            f"tcp://{get_local_ip_auto()}:{PORT_BASE + server_args.node_rank * 2}"
+            f"tcp://{get_local_ip_auto()}:{PORT_BASE + get_parallel().node_rank * 2}"
         )
         self.send_to_expert_backup_client = context.socket(zmq.PUB)
         self.send_to_expert_backup_client.bind(
-            f"tcp://{get_local_ip_auto()}:{PORT_BASE + server_args.node_rank * 2 + 1}"
+            f"tcp://{get_local_ip_auto()}:{PORT_BASE + get_parallel().node_rank * 2 + 1}"
         )
         self.backup_weights_from_disk()
         self.start_transfer_server()
@@ -66,7 +73,7 @@ class ExpertBackupManager:
         # losing the initial PUB message due to slow joiners.
         num_ready_clients = 0
 
-        while num_ready_clients < server_args.tp_size:
+        while num_ready_clients < configured_tp_size():
             sock_recv(self.recv_from_expert_backup_client)
             num_ready_clients += 1
 
@@ -168,7 +175,7 @@ def run_expert_backup_manager_process(
         hostname=get_local_ip_auto(),
         gpu_id=0,
         ib_device=(
-            server_args.disaggregation_ib_device or server_args.mooncake_ib_device
+            get_disagg().disaggregation_ib_device or get_exec().moe.mooncake_ib_device
         ),
     )
     manager = ExpertBackupManager(server_args, port_args)

python/sglang/srt/eplb/expert_distribution.py

@@ -45,7 +45,11 @@ from sglang.srt.observability.metrics_collector import (
     ExpertDispatchCollector,
     resolve_collector_class,
 )
-from sglang.srt.runtime_context import get_schedule
+from sglang.srt.runtime_context import get_device as get_device_namespace
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_schedule,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import Withable, get_device, get_int_env_var
 
@@ -84,7 +88,7 @@ class ExpertDistributionRecorder(ABC):
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
-        if server_args.expert_distribution_recorder_mode is not None:
+        if get_exec().moe.expert_distribution_recorder_mode is not None:
             assert (
                 expert_location_metadata is not None
             ), "ExpertLocationMetadata is required for expert distribution recording. One possible"
@@ -178,7 +182,7 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
         if server_args.should_report_expert_balancedness():
             logger.info(
                 "ExpertDistributionRecorder auto start record since "
-                f"expert_balancedness_report_mode={server_args.expert_balancedness_report_mode}"
+                f"expert_balancedness_report_mode={get_exec().moe.expert_balancedness_report_mode}"
             )
             self.start_record()
 
@@ -328,30 +332,30 @@ class _SinglePassGatherer(ABC):
         expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ) -> _SinglePassGatherer:
-        if server_args.expert_distribution_recorder_mode == "per_token":
+        if get_exec().moe.expert_distribution_recorder_mode == "per_token":
             return _DetailSinglePassGatherer(
                 server_args, expert_location_metadata, rank
             )
 
-        if server_args.moe_a2a_backend == "mori":
+        if get_exec().moe.moe_a2a_backend == "mori":
             return _DeepepLowLatencySinglePassGatherer(expert_location_metadata, rank)
 
-        if server_args.expert_distribution_recorder_mode == "stat_approx":
-            if server_args.moe_a2a_backend != "none" and (
-                server_args.deepep_mode == "normal"
+        if get_exec().moe.expert_distribution_recorder_mode == "stat_approx":
+            if get_exec().moe.moe_a2a_backend != "none" and (
+                get_exec().moe.deepep_mode == "normal"
             ):
                 return _DeepepNormalSinglePassGatherer(expert_location_metadata, rank)
             else:
                 raise NotImplementedError
 
-        if server_args.moe_a2a_backend == "deepep":
-            if server_args.deepep_mode == "normal":
+        if get_exec().moe.moe_a2a_backend == "deepep":
+            if get_exec().moe.deepep_mode == "normal":
                 return _SelectExpertsSinglePassGatherer(expert_location_metadata, rank)
-            elif server_args.deepep_mode == "low_latency":
+            elif get_exec().moe.deepep_mode == "low_latency":
                 return _DeepepLowLatencySinglePassGatherer(
                     expert_location_metadata,
                     rank,
-                    elastic_ep_enabled=server_args.elastic_ep_backend is not None,
+                    elastic_ep_enabled=get_exec().moe.elastic_ep_backend is not None,
                 )
             else:
                 raise NotImplementedError
@@ -412,11 +416,11 @@ class _DetailSinglePassGatherer(_SinglePassGatherer):
                 self._TOP_K_NUM,
             ),
             dtype=torch.int32,
-            device=server_args.device,
+            device=get_device_namespace().device,
         )
         self._misc_objects: List[Dict[str, Any]] = []
         assert (
-            not server_args.enable_two_batch_overlap
+            not get_exec().overlap.enable_two_batch_overlap
         ), "DetailSinglePassGatherer does not support TBO yet"
         # TODO assert shared experts fusion is disabled, o/w data is wrong
 
@@ -678,7 +682,7 @@ class _Accumulator(ABC):
             "stat_approx": _StatAccumulator,
             "per_pass": _DetailAccumulator,
             "per_token": _DetailAccumulator,
-        }[server_args.expert_distribution_recorder_mode]
+        }[get_exec().moe.expert_distribution_recorder_mode]
 
     def __init__(
         self,

python/sglang/srt/eplb/expert_location.py

@@ -25,6 +25,12 @@ import torch
 import torch.distributed
 import torch.nn.functional as F
 
+from sglang.srt.runtime_context import (
+    configured_tp_size,
+    get_device,
+    get_exec,
+)
+
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
     from sglang.srt.server_args import ServerArgs
@@ -141,7 +147,7 @@ class ExpertLocationMetadata:
     ):
         if not isinstance(physical_to_logical_map, torch.Tensor):
             physical_to_logical_map = torch.tensor(physical_to_logical_map)
-        physical_to_logical_map = physical_to_logical_map.to(server_args.device)
+        physical_to_logical_map = physical_to_logical_map.to(get_device().device)
 
         common = ExpertLocationMetadata._init_common(server_args, model_config)
 
@@ -184,7 +190,7 @@ class ExpertLocationMetadata:
             logical_count = torch.tensor(logical_count)
         if len(logical_count.shape) == 2:
             logical_count = logical_count.unsqueeze(0)
-        logical_count = logical_count.to(server_args.device)
+        logical_count = logical_count.to(get_device().device)
 
         from sglang.srt.runtime_context import get_parallel
 
@@ -208,7 +214,7 @@ class ExpertLocationMetadata:
                 num_groups=num_groups,
                 num_nodes=num_nodes,
                 algorithm=eplb_algorithms.compute_algorithm(
-                    raw_algorithm=server_args.eplb_algorithm,
+                    raw_algorithm=get_exec().moe.eplb_algorithm,
                     num_groups=num_groups,
                     num_nodes=num_nodes,
                 ),
@@ -217,9 +223,9 @@ class ExpertLocationMetadata:
 
         return ExpertLocationMetadata._init_raw(
             ep_size=common["ep_size"],
-            physical_to_logical_map=physical_to_logical_map.to(server_args.device),
+            physical_to_logical_map=physical_to_logical_map.to(get_device().device),
             logical_to_all_physical_map=logical_to_all_physical_map.to(
-                server_args.device
+                get_device().device
             ),
         )
 
@@ -246,7 +252,7 @@ class ExpertLocationMetadata:
             if get_exec().moe.ep_join_mode == "scale":
                 ep_size = max(
                     ep_size,
-                    get_parallel().ep_join_rank_offset + server_args.tp_size,
+                    get_parallel().ep_join_rank_offset + configured_tp_size(),
                 )
             num_physical_experts, num_local_physical_experts = (
                 _compute_elastic_expert_layout(
@@ -781,7 +787,7 @@ def compute_initial_expert_location_metadata(
     model_config: ModelConfig,
     moe_ep_rank: int,
 ) -> Optional[ExpertLocationMetadata]:
-    data = server_args.init_expert_location
+    data = get_exec().moe.init_expert_location
     if data == "trivial":
         return ExpertLocationMetadata.init_trivial(
             server_args, model_config, moe_ep_rank

python/sglang/srt/kv_canary/api.py

@@ -18,6 +18,10 @@ from sglang.srt.model_executor.cuda_graph_config import (
     check_cuda_graph_backend,
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.runtime_context import (
+    get_disagg,
+    get_spec,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.kv_canary.token_oracle.oracle_manager import TokenOracleManager
@@ -59,13 +63,12 @@ def install_canary(
         allocator if isinstance(allocator, SWATokenToKVPoolAllocator) else None
     )
     launch_capacities = CanaryLaunchCapacities.from_args(
-        server_args=model_runner.server_args,
         req_to_token_pool_size=model_runner.req_to_token_pool.size,
         max_seq_len_per_req=model_runner.req_to_token_pool.req_to_token.shape[1],
         pool_slot_count=model_runner.max_total_num_tokens,
     )
     swa_window_size = model_runner.sliding_window_size or 0
-    speculative_num_steps = int(server_args.speculative_num_steps or 1)
+    speculative_num_steps = int(get_spec().speculative_num_steps or 1)
     manager = CanaryManager(
         config=config,
         perturb_config=perturb_config,
@@ -88,7 +91,7 @@ def install_canary(
         "install_canary: disaggregation_mode=%s config=%s perturb_config=%s "
         "launch_capacities=%s n_buffer_groups=%d buffer_group_kinds=%s "
         "swa_window_size=%d speculative_num_steps=%d",
-        server_args.disaggregation_mode,
+        get_disagg().disaggregation_mode,
         config,
         perturb_config,
         launch_capacities,

python/sglang/srt/kv_canary/capacities.py

@@ -2,10 +2,12 @@ from __future__ import annotations
 
 import math
 from dataclasses import dataclass
-from typing import TYPE_CHECKING
 
-if TYPE_CHECKING:
-    from sglang.srt.server_args import ServerArgs
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_schedule,
+    get_spec,
+)
 
 
 @dataclass(frozen=True, slots=True, kw_only=True)
@@ -43,7 +45,6 @@ class CanaryLaunchCapacities:
     def from_args(
         cls,
         *,
-        server_args: ServerArgs,
         req_to_token_pool_size: int,
         max_seq_len_per_req: int,
         pool_slot_count: int,
@@ -63,7 +64,7 @@ class CanaryLaunchCapacities:
                 f"kv-canary: pool_slot_count must be positive, got {pool_slot_count}"
             )
 
-        cuda_graph_config = server_args.cuda_graph_config
+        cuda_graph_config = get_exec().graph.cuda_graph_config
         cuda_graph_max_bs = (
             cuda_graph_config.decode.max_bs if cuda_graph_config is not None else 0
         ) or 0
@@ -72,7 +73,7 @@ class CanaryLaunchCapacities:
                 f"kv-canary: cuda_graph_max_bs must be non-negative, got {cuda_graph_max_bs}"
             )
 
-        spec_num_draft_tokens = server_args.speculative_num_draft_tokens
+        spec_num_draft_tokens = get_spec().speculative_num_draft_tokens
         if spec_num_draft_tokens is None:
             spec_num_draft_tokens = 0
         if spec_num_draft_tokens < 0:
@@ -81,7 +82,7 @@ class CanaryLaunchCapacities:
                 f"got {spec_num_draft_tokens}"
             )
 
-        max_prefill_tokens = server_args.max_prefill_tokens
+        max_prefill_tokens = get_schedule().max_prefill_tokens
         if max_prefill_tokens <= 0:
             raise ValueError(
                 f"kv-canary: max_prefill_tokens must be positive, got {max_prefill_tokens}"
@@ -93,7 +94,7 @@ class CanaryLaunchCapacities:
 
         max_bs = max(cuda_graph_max_bs, req_to_token_pool_size)
 
-        chunked_prefill_size = server_args.chunked_prefill_size
+        chunked_prefill_size = get_schedule().chunked_prefill_size
         chunked_limit = (
             chunked_prefill_size
             if chunked_prefill_size is not None and chunked_prefill_size >= 0

python/sglang/srt/managers/disagg_service.py

@@ -8,6 +8,11 @@ from sglang.srt.disaggregation.utils import (
     TransferBackend,
     get_kv_class,
 )
+from sglang.srt.runtime_context import (
+    get_disagg,
+    get_parallel,
+    get_serving,
+)
 from sglang.srt.server_args import ServerArgs
 
 
@@ -15,8 +20,8 @@ def start_disagg_service(
     server_args: ServerArgs,
 ):
     # Start kv bootstrap server on prefill
-    disagg_mode = DisaggregationMode(server_args.disaggregation_mode)
-    transfer_backend = TransferBackend(server_args.disaggregation_transfer_backend)
+    disagg_mode = DisaggregationMode(get_disagg().disaggregation_mode)
+    transfer_backend = TransferBackend(get_disagg().disaggregation_transfer_backend)
 
     if disagg_mode == DisaggregationMode.PREFILL:
         # only start bootstrap server on prefill tm
@@ -24,8 +29,8 @@ def start_disagg_service(
             transfer_backend, KVClassType.BOOTSTRAP_SERVER
         )
         bootstrap_server = kv_bootstrap_server_class(
-            host=server_args.host,
-            port=server_args.disaggregation_bootstrap_port,
+            host=get_serving().host,
+            port=get_disagg().disaggregation_bootstrap_port,
         )
         maybe_create_ascend_config_store(
             server_args=server_args, transfer_backend=transfer_backend
@@ -43,7 +48,9 @@ def maybe_create_ascend_config_store(
     bootstrap-server subclasses are all plain ``CommonKVBootstrapServer``,
     which the rust registry ports verbatim), leaving this store as the only
     ``start_disagg_service`` duty left to perform."""
-    if not (server_args.node_rank == 0 and transfer_backend == TransferBackend.ASCEND):
+    if not (
+        get_parallel().node_rank == 0 and transfer_backend == TransferBackend.ASCEND
+    ):
         return
     try:
         from memfabric_hybrid import create_config_store

python/sglang/srt/managers/overlap_utils.py

@@ -8,6 +8,10 @@ import torch
 
 from sglang.kernels.ops.speculative.gather_spec_extras import gather_spec_extras
 from sglang.srt.environ import envs
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_spec,
+)
 from sglang.srt.utils import is_cuda, is_hip, is_npu
 
 if TYPE_CHECKING:
@@ -34,10 +38,10 @@ def decide_needs_cpu_seq_lens(
     # importable everywhere; spec_info pulls in the spec/schedule_batch graph.
     from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 
-    if server_args.enable_two_batch_overlap:
+    if get_exec().overlap.enable_two_batch_overlap:
         # FIXME: support TBO without seq lens cpu value
         return True
-    algo = SpeculativeAlgorithm.from_string(server_args.speculative_algorithm)
+    algo = SpeculativeAlgorithm.from_string(get_spec().speculative_algorithm)
     if algo.is_ngram():
         # ngram's USE_FULL_MASK verify path reads seq_lens_cpu per req to size
         # the tree mask, regardless of the attn backend (e.g. Triton opts out).
@@ -56,7 +60,7 @@ def decide_needs_confidence_relay(server_args: ServerArgs) -> bool:
     )
     from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 
-    algo = SpeculativeAlgorithm.from_string(server_args.speculative_algorithm)
+    algo = SpeculativeAlgorithm.from_string(get_spec().speculative_algorithm)
     if not algo.is_dspark():
         return False
     return read_ragged_verify_mode() is not RaggedVerifyMode.STATIC

python/sglang/srt/managers/schedule_batch.py

@@ -1930,7 +1930,7 @@ def release_req(
     # Callers that will recompute the KV instead (PD true-retraction rebootstrap)
     # pass offload_kv=False to skip the wasteful device->host copy.
     backup_saved = True
-    if server_args.disaggregation_mode == "decode" and offload_kv:
+    if get_disagg().disaggregation_mode == "decode" and offload_kv:
         backup_saved = retraction_backup(
             req,
             tree_cache,
@@ -2826,7 +2826,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self, server_args: ServerArgs
     ) -> Tuple[List[Req], float, List[Req]]:
         """Retract the decoding requests when there is not enough memory."""
-        sorted_indices = self._get_decode_retraction_order(self.reqs, server_args)
+        sorted_indices = self._get_decode_retraction_order(self.reqs)
 
         retracted_reqs = []
         reqs_to_abort: List[Req] = []
@@ -2886,9 +2886,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         return retracted_reqs, new_estimate_ratio, reqs_to_abort
 
     @staticmethod
-    def _get_decode_retraction_order(
-        reqs: List[Req], server_args: ServerArgs
-    ) -> List[int]:
+    def _get_decode_retraction_order(reqs: List[Req]) -> List[int]:
         """Return indices ordered from most-preferred to least-preferred to keep.
 
         The retraction loop pops from the end of this list, so the least-preferred
@@ -2901,15 +2899,17 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         def length_key(req: Req) -> Tuple[int, int]:
             return (len(req.output_ids), -len(req.origin_input_ids))
 
-        if server_args.retraction_policy == "priority":
-            priority_sign = 1 if server_args.schedule_low_priority_values_first else -1
+        if get_schedule().retraction_policy == "priority":
+            priority_sign = (
+                1 if get_schedule().schedule_low_priority_values_first else -1
+            )
 
             def retraction_key(req: Req) -> Tuple[int, int, int]:
                 priority = req.priority
                 if priority is None:
                     priority = (
                         sys.maxsize
-                        if server_args.schedule_low_priority_values_first
+                        if get_schedule().schedule_low_priority_values_first
                         else -sys.maxsize - 1
                     )
                 return (priority * (-priority_sign), *length_key(req))

python/sglang/srt/managers/scheduler.py

@@ -30,6 +30,7 @@ from typing import TYPE_CHECKING, Any, Deque, Dict, List, Optional, Set, Tuple,
 from sglang.srt.runtime_context import (
     attention_backends,
     configured_attn_cp_size,
+    configured_dcp_size,
     configured_moe_dp_size,
     configured_pp_size,
     configured_tp_size,
@@ -417,71 +418,75 @@ class Scheduler(
         # Parse args
         self.server_args = server_args
         self.nccl_port = port_args.nccl_port
-        self.schedule_policy = server_args.schedule_policy
-        self.enable_priority_scheduling = server_args.enable_priority_scheduling
+        self.schedule_policy = get_schedule().schedule_policy
+        self.enable_priority_scheduling = get_schedule().enable_priority_scheduling
         self.abort_on_priority_when_disabled = (
-            server_args.abort_on_priority_when_disabled
+            get_schedule().abort_on_priority_when_disabled
         )
         self.schedule_low_priority_values_first = (
-            server_args.schedule_low_priority_values_first
+            get_schedule().schedule_low_priority_values_first
         )
         self.priority_scheduling_preemption_threshold = (
-            server_args.priority_scheduling_preemption_threshold
+            get_schedule().priority_scheduling_preemption_threshold
         )
-        self.enable_lora = server_args.enable_lora
-        self.enable_lora_overlap_loading = server_args.enable_lora_overlap_loading
-        self.max_loras_per_batch = server_args.max_loras_per_batch
-        self.enable_overlap = not server_args.disable_overlap_schedule and not use_mlx()
-        self.enable_overlap_mlx = not server_args.disable_overlap_schedule and use_mlx()
-        self.enable_pdmux = server_args.enable_pdmux
+        self.enable_lora = get_lora().enable_lora
+        self.enable_lora_overlap_loading = get_lora().enable_lora_overlap_loading
+        self.max_loras_per_batch = get_lora().max_loras_per_batch
+        self.enable_overlap = (
+            not get_schedule().disable_overlap_schedule and not use_mlx()
+        )
+        self.enable_overlap_mlx = (
+            not get_schedule().disable_overlap_schedule and use_mlx()
+        )
+        self.enable_pdmux = get_disagg().enable_pdmux
         self.skip_tokenizer_init = get_serving().skip_tokenizer_init
-        self.stream_interval = server_args.stream_interval
+        self.stream_interval = get_serving().stream_interval
         self.spec_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
         self.page_size = get_schedule().page_size
-        self.enable_hierarchical_cache = server_args.enable_hierarchical_cache
-        self.enable_session_radix_cache = server_args.enable_session_radix_cache
-        self.enable_hicache_storage = server_args.hicache_storage_backend is not None
+        self.enable_hierarchical_cache = get_memory().enable_hierarchical_cache
+        self.enable_session_radix_cache = get_memory().enable_session_radix_cache
+        self.enable_hicache_storage = get_memory().hicache_storage_backend is not None
         self.enable_decode_hicache = (
-            server_args.disaggregation_decode_enable_radix_cache
+            get_disagg().disaggregation_decode_enable_radix_cache
             and self.enable_hierarchical_cache
         )
         self.max_recv_per_poll = envs.SGLANG_SCHEDULER_MAX_RECV_PER_POLL.get()
         self.max_new_tokens_limit = envs.SGLANG_MAX_NEW_TOKENS_LIMIT.get()
-        self.enable_hisparse = server_args.enable_hisparse
+        self.enable_hisparse = get_memory().enable_hisparse
         self.enable_dp_attention = get_parallel().enable_dp_attention
-        self.enable_unified_memory = server_args.enable_unified_memory
+        self.enable_unified_memory = get_memory().enable_unified_memory
 
         # Distributed rank info
         attn_tp_rank, attn_tp_size, attn_dp_rank, attn_dp_size = (
             compute_dp_attention_world_info(
                 get_parallel().enable_dp_attention,
                 tp_rank,
-                server_args.tp_size,
+                configured_tp_size(),
                 get_parallel().dp_size,
-                server_args.attn_cp_size,
+                configured_attn_cp_size(),
             )
         )
         self.ps = ParallelState(
             tp_rank=tp_rank,
-            tp_size=server_args.tp_size,
+            tp_size=configured_tp_size(),
             pp_rank=pp_rank,
-            pp_size=server_args.pp_size,
+            pp_size=configured_pp_size(),
             dp_rank=dp_rank,
             dp_size=get_parallel().dp_size,
             attn_tp_rank=attn_tp_rank,
             attn_tp_size=attn_tp_size,
             attn_cp_rank=attn_cp_rank,
-            attn_cp_size=server_args.attn_cp_size,
-            attn_dcp_rank=tp_rank % server_args.dcp_size,
-            attn_dcp_size=server_args.dcp_size,
+            attn_cp_size=configured_attn_cp_size(),
+            attn_dcp_rank=tp_rank % configured_dcp_size(),
+            attn_dcp_size=configured_dcp_size(),
             attn_dp_rank=attn_dp_rank,
             attn_dp_size=attn_dp_size,
             moe_ep_rank=moe_ep_rank,
             moe_ep_size=get_parallel().ep_size,
             moe_dp_rank=moe_dp_rank,
-            moe_dp_size=server_args.moe_dp_size,
+            moe_dp_size=configured_moe_dp_size(),
             gpu_id=gpu_id,
         )
 
@@ -839,7 +844,7 @@ class Scheduler(
         if (
             self.model_config.is_multimodal
             and self.processor is not None
-            and not server_args.language_model_only
+            and not get_disagg().language_model_only
         ):
             try:
                 import_processors("sglang.srt.multimodal.processors")
@@ -1282,7 +1287,7 @@ class Scheduler(
         self.new_token_ratio_tracker = NewTokenRatioTracker.from_config()
 
     def init_soft_watchdog(self, server_args: ServerArgs):
-        if (x := server_args.soft_watchdog_timeout) is not None:
+        if (x := get_device().soft_watchdog_timeout) is not None:
             self.soft_watchdog = create_scheduler_watchdog(
                 self, watchdog_timeout=x, soft=True
             )
@@ -5172,16 +5177,16 @@ def run_scheduler_process(
     parent_process = psutil.Process().parent()
 
     # Set up tracing
-    if server_args.enable_trace:
+    if get_observability().enable_trace:
         process_tracing_init(
-            server_args.otlp_traces_endpoint,
+            get_observability().otlp_traces_endpoint,
             "sglang",
-            trace_modules=server_args.trace_modules,
+            trace_modules=get_observability().trace_modules,
         )
         thread_label = "Scheduler"
-        if server_args.disaggregation_mode == "prefill":
+        if get_disagg().disaggregation_mode == "prefill":
             thread_label = "Prefill Scheduler"
-        elif server_args.disaggregation_mode == "decode":
+        elif get_disagg().disaggregation_mode == "decode":
             thread_label = "Decode Scheduler"
         trace_set_thread_info(thread_label, tp_rank, dp_rank, pp_rank)
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -131,7 +131,9 @@ from sglang.srt.runtime_context import (
     get_memory,
     get_mm,
     get_model,
+    get_observability,
     get_parallel,
+    get_schedule,
     get_serving,
     get_spec,
 )
@@ -411,14 +413,14 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         self.elastic_pending_ep_size = None
         self.elastic_scale_phase = "idle"
         self.elastic_last_error = None
-        self.enable_metrics = server_args.enable_metrics
-        self.incremental_streaming_output = server_args.incremental_streaming_output
+        self.enable_metrics = get_observability().enable_metrics
+        self.incremental_streaming_output = get_serving().incremental_streaming_output
         self.enable_lora = get_lora().enable_lora
-        self.enable_trace = server_args.enable_trace
-        self.allow_auto_truncate = server_args.allow_auto_truncate
-        self.skip_tokenizer_init = server_args.skip_tokenizer_init
+        self.enable_trace = get_observability().enable_trace
+        self.allow_auto_truncate = get_serving().allow_auto_truncate
+        self.skip_tokenizer_init = get_serving().skip_tokenizer_init
         self.preferred_sampling_params = get_serving().preferred_sampling_params
-        self.crash_dump_folder = server_args.crash_dump_folder
+        self.crash_dump_folder = get_observability().crash_dump_folder
 
         # Init model config
         self.init_model_config()
@@ -462,15 +464,15 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         model_config_class = getattr(self, "model_config_class", ModelConfig)
 
         # Read model args
-        self.model_path = server_args.model_path
-        self.served_model_name = server_args.served_model_name
+        self.model_path = get_model().model_path
+        self.served_model_name = get_serving().served_model_name
         self.model_config = model_config_class.from_server_args(server_args)
         self.is_generation = self.model_config.is_generation
         self.context_len = self.model_config.context_len
         self.image_token_id = self.model_config.image_token_id
         self.max_req_input_len = None  # Will be set later in engine.py
-        self.enable_priority_scheduling = server_args.enable_priority_scheduling
-        self.default_priority_value = server_args.default_priority_value
+        self.enable_priority_scheduling = get_schedule().enable_priority_scheduling
+        self.default_priority_value = get_schedule().default_priority_value
         self.num_reserved_tokens = compute_num_reserved_tokens()
         self.validate_total_tokens = True
 
@@ -478,7 +480,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         server_args = self.server_args
 
         # Initialize tokenizer and processor
-        if self.model_config.is_multimodal and not server_args.language_model_only:
+        if self.model_config.is_multimodal and not get_disagg().language_model_only:
             import_processors("sglang.srt.multimodal.processors")
             if mm_process_pkg := envs.SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE.get():
                 import_processors(mm_process_pkg, overwrite=True)
@@ -496,7 +498,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 model_config=self.model_config,
             )
 
-            if server_args.skip_tokenizer_init:
+            if get_serving().skip_tokenizer_init:
                 self.tokenizer = self.processor = None
             else:
                 self.processor = _processor
@@ -505,26 +507,26 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         else:
             self.mm_processor = self.processor = None
 
-            if server_args.skip_tokenizer_init:
+            if get_serving().skip_tokenizer_init:
                 self.tokenizer = None
             else:
                 self.tokenizer = get_tokenizer(
                     get_serving().tokenizer_path,
-                    tokenizer_mode=server_args.tokenizer_mode,
-                    trust_remote_code=server_args.trust_remote_code,
-                    revision=server_args.revision,
-                    tokenizer_backend=server_args.tokenizer_backend,
+                    tokenizer_mode=get_serving().tokenizer_mode,
+                    trust_remote_code=get_model().trust_remote_code,
+                    revision=get_model().revision,
+                    tokenizer_backend=get_serving().tokenizer_backend,
                 )
 
         # Initialize async dynamic batch tokenizer if enabled (common for both multimodal and non-multimodal)
         if (
-            server_args.enable_dynamic_batch_tokenizer
-            and not server_args.skip_tokenizer_init
+            get_serving().enable_dynamic_batch_tokenizer
+            and not get_serving().skip_tokenizer_init
         ):
             self.async_dynamic_batch_tokenizer = AsyncDynamicbatchTokenizer(
                 self.tokenizer,
-                max_batch_size=server_args.dynamic_batch_tokenizer_batch_size,
-                batch_wait_timeout_s=server_args.dynamic_batch_tokenizer_batch_timeout,
+                max_batch_size=get_serving().dynamic_batch_tokenizer_batch_size,
+                batch_wait_timeout_s=get_serving().dynamic_batch_tokenizer_batch_timeout,
             )
         else:
             self.async_dynamic_batch_tokenizer = None
@@ -547,7 +549,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         self.recv_from_detokenizer = get_zmq_socket(
             context, zmq.PULL, port_args.tokenizer_ipc_name, True
         )
-        if self.server_args.tokenizer_worker_num == 1:
+        if get_serving().tokenizer_worker_num == 1:
             self.send_to_scheduler = get_zmq_socket(
                 context, zmq.PUSH, port_args.scheduler_input_ipc_name, True
             )
@@ -595,10 +597,10 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         # TODO: Refactor and organize the log export code.
         # Request logging
         self.request_logger = RequestLogger(
-            log_requests=self.server_args.log_requests,
-            log_requests_level=self.server_args.log_requests_level,
-            log_requests_format=self.server_args.log_requests_format,
-            log_requests_target=self.server_args.log_requests_target,
+            log_requests=get_observability().log_requests,
+            log_requests_level=get_observability().log_requests_level,
+            log_requests_format=get_observability().log_requests_format,
+            log_requests_target=get_observability().log_requests_target,
         )
 
         # Dumping
@@ -621,7 +623,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
     def init_weight_update(self):
         # Initial weights status
         self.initial_weights_loaded = True
-        if self.server_args.checkpoint_engine_wait_weights_before_ready:
+        if get_model().checkpoint_engine_wait_weights_before_ready:
             self.initial_weights_loaded = False
 
         # Weight updates
@@ -668,7 +670,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
 
         # Encoder Disaggregation
         self.encoder_bootstrap_server = None
-        if self.server_args.language_only:
+        if get_disagg().language_only:
             from sglang.srt.disaggregation.encoder.receiver import (
                 EncoderBootstrapServer,
             )
@@ -677,10 +679,10 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             # entries as encoders register, the receiver reads from the same
             # list.  Pre-populated with static --encoder-urls so the legacy
             # CLI flag still works (alongside dynamic registrations).
-            self.encoder_urls: List[str] = list(self.server_args.encoder_urls)
+            self.encoder_urls: List[str] = list(get_disagg().encoder_urls)
             self.encoder_bootstrap_server = EncoderBootstrapServer(
-                host=self.server_args.host,
-                port=self.server_args.encoder_bootstrap_port,
+                host=get_serving().host,
+                port=get_disagg().encoder_bootstrap_port,
                 urls=self.encoder_urls,
             )
             self.mm_receiver = create_mm_receiver(
@@ -698,16 +700,18 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             )
 
             labels = {
-                "model_name": self.server_args.served_model_name,
+                "model_name": get_serving().served_model_name,
                 "engine_type": engine_type,
             }
             if self.enable_priority_scheduling:
                 labels["priority"] = ""
-            if self.server_args.tokenizer_metrics_allowed_custom_labels:
-                for label in self.server_args.tokenizer_metrics_allowed_custom_labels:
+            if get_observability().tokenizer_metrics_allowed_custom_labels:
+                for (
+                    label
+                ) in get_observability().tokenizer_metrics_allowed_custom_labels:
                     labels[label] = ""
-            if self.server_args.extra_metric_labels:
-                labels.update(self.server_args.extra_metric_labels)
+            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,
@@ -716,15 +720,15 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             self.metrics_collector = tokenizer_collector_cls(
                 server_args=self.server_args,
                 labels=labels,
-                bucket_time_to_first_token=self.server_args.bucket_time_to_first_token,
-                bucket_e2e_request_latency=self.server_args.bucket_e2e_request_latency,
-                bucket_inter_token_latency=self.server_args.bucket_inter_token_latency,
+                bucket_time_to_first_token=get_observability().bucket_time_to_first_token,
+                bucket_e2e_request_latency=get_observability().bucket_e2e_request_latency,
+                bucket_inter_token_latency=get_observability().bucket_inter_token_latency,
             )
 
             start_cpu_monitor_thread("tokenizer")
 
-        if self.server_args.gc_warning_threshold_secs > 0.0:
-            configure_gc_warning(self.server_args.gc_warning_threshold_secs)
+        if get_observability().gc_warning_threshold_secs > 0.0:
+            configure_gc_warning(get_observability().gc_warning_threshold_secs)
         self.soft_watchdog = Watchdog.create(
             debug_name="TokenizerManager",
             watchdog_timeout=get_device().soft_watchdog_timeout,
@@ -773,7 +777,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         if (
             isinstance(obj, GenerateReqInput)
             and obj.max_thinking_tokens is not None
-            and not self.server_args.enable_strict_thinking
+            and not get_serving().enable_strict_thinking
         ):
             raise ValueError(
                 "max_thinking_tokens requires the server to be launched with "
@@ -793,7 +797,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
 
         self._init_req_state(obj, request)
         try:
-            if self.server_args.language_only:
+            if get_disagg().language_only:
                 self._handle_epd_disaggregation_encode_request(obj)
 
             # Log the request
@@ -993,7 +997,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 )
 
         contains_mm_input = obj.contains_mm_input()
-        if contains_mm_input and self.server_args.language_model_only:
+        if contains_mm_input and get_disagg().language_model_only:
             raise ValueError(
                 "Multimodal inputs are not supported when --language-model-only "
                 "is set; the encoder is not loaded. Restart without the flag."
@@ -1028,10 +1032,10 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             )
 
             if (
-                not self.server_args.language_only
+                not get_disagg().language_only
                 or get_disagg().encoder_transfer_backend == "zmq_to_tokenizer"
             ):
-                if self.server_args.language_only:
+                if get_disagg().language_only:
                     mm_inputs = await self.mm_receiver.recv_mm_data(
                         request_obj=obj,
                         mm_processor=self.mm_processor,
@@ -1053,7 +1057,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                         max_req_input_len=self.max_req_input_len,
                     )
             elif (
-                self.server_args.language_only
+                get_disagg().language_only
                 and get_disagg().encoder_transfer_backend
                 in ["zmq_to_scheduler", "mooncake"]
                 and not obj.need_wait_for_mm_inputs
@@ -1238,7 +1242,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 )
             if (
                 obj.custom_logit_processor
-                and not self.server_args.enable_custom_logit_processor
+                and not get_exec().features.enable_custom_logit_processor
             ):
                 raise ValueError(
                     "The server is not configured to enable custom logit processor. "
@@ -1945,7 +1949,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             return
         if (
             not abort_all
-            and self.server_args.tokenizer_worker_num == 1
+            and get_serving().tokenizer_worker_num == 1
             and rid not in self.rid_to_state
         ):
             return
@@ -2175,7 +2179,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         else:
             customized_info = None
         pending_notify: dict[str, ReqState] = {}
-        batch_notify_size = self.server_args.batch_notify_size
+        batch_notify_size = get_serving().batch_notify_size
         for i, rid in enumerate(recv_obj.rids):
             state = self.rid_to_state.get(rid, None)
             if state is None:
@@ -3183,7 +3187,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 state.obj.top_logprobs_num,
                 state.obj.token_ids_logprob,
                 state.obj.return_text_in_logprobs
-                and not self.server_args.skip_tokenizer_init,
+                and not get_serving().skip_tokenizer_init,
             )
 
         output_ids = state.output_ids
@@ -3304,12 +3308,12 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             unique_lora_paths = set(obj.lora_path)
 
         if (
-            self.server_args.max_loaded_loras is not None
-            and len(unique_lora_paths) > self.server_args.max_loaded_loras
+            get_lora().max_loaded_loras is not None
+            and len(unique_lora_paths) > get_lora().max_loaded_loras
         ):
             raise ValueError(
                 f"Received request with {len(unique_lora_paths)} unique loras requested "
-                f"but max loaded loras is {self.server_args.max_loaded_loras}"
+                f"but max loaded loras is {get_lora().max_loaded_loras}"
             )
 
         # Reload all existing LoRA adapters that have been dynamically unloaded
@@ -3442,7 +3446,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         if isinstance(obj, GenerateReqInput) and obj.contains_mm_input():
             # dispatch to encoder by default
             should_dispatch = True
-            if self.server_args.enable_adaptive_dispatch_to_encoder:
+            if get_disagg().enable_adaptive_dispatch_to_encoder:
                 should_dispatch = self._should_dispatch_to_encoder(obj)
 
             # Set need_wait_for_mm_inputs flag based on whether we dispatch to encoder
@@ -3579,7 +3583,7 @@ def get_processor_wrapper(server_args):
 
 
 def determine_tensor_transport_mode(server_args: ServerArgs) -> TensorTransportMode:
-    is_cross_node = server_args.dist_init_addr
+    is_cross_node = get_parallel().dist_init_addr
 
     if is_cross_node:
         # Fallback to default CPU transport for multi-node

python/sglang/srt/managers/tp_worker.py

@@ -52,7 +52,14 @@ from sglang.srt.model_executor.graph_memory_usage import (
     merge_graph_time_usage,
 )
 from sglang.srt.model_executor.pool_configurator import MemoryPoolConfig
-from sglang.srt.runtime_context import get_exec, get_model, get_schedule, get_spec
+from sglang.srt.runtime_context import (
+    get_device,
+    get_exec,
+    get_model,
+    get_schedule,
+    get_serving,
+    get_spec,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import MultiprocessingSerializer, broadcast_pyobj, set_random_seed
 from sglang.srt.utils.hf_transformers_utils import (
@@ -341,28 +348,28 @@ class TpModelWorker(BaseTpWorker):
 
         self._init_dllm_algorithm()
 
-        if server_args.skip_tokenizer_init or self.is_draft_worker:
+        if get_serving().skip_tokenizer_init or self.is_draft_worker:
             # A draft worker's tokenizer would only duplicate the target's:
             # tokenizer_path always points at the target model.
             self.tokenizer = self.processor = None
         else:
             if self.model_config.is_multimodal:
                 self.processor = get_processor(
-                    server_args.tokenizer_path,
-                    tokenizer_mode=server_args.tokenizer_mode,
-                    trust_remote_code=server_args.trust_remote_code,
-                    revision=server_args.revision,
-                    tokenizer_backend=server_args.tokenizer_backend,
-                    model_name=server_args.model_path,
+                    get_serving().tokenizer_path,
+                    tokenizer_mode=get_serving().tokenizer_mode,
+                    trust_remote_code=get_model().trust_remote_code,
+                    revision=get_model().revision,
+                    tokenizer_backend=get_serving().tokenizer_backend,
+                    model_name=get_model().model_path,
                 )
                 self.tokenizer = get_tokenizer_from_processor(self.processor)
             else:
                 self.tokenizer = get_tokenizer(
-                    server_args.tokenizer_path,
-                    tokenizer_mode=server_args.tokenizer_mode,
-                    trust_remote_code=server_args.trust_remote_code,
-                    revision=server_args.revision,
-                    tokenizer_backend=server_args.tokenizer_backend,
+                    get_serving().tokenizer_path,
+                    tokenizer_mode=get_serving().tokenizer_mode,
+                    trust_remote_code=get_model().trust_remote_code,
+                    revision=get_model().revision,
+                    tokenizer_backend=get_serving().tokenizer_backend,
                 )
         self.device = self.model_runner.device
 
@@ -373,18 +380,18 @@ class TpModelWorker(BaseTpWorker):
         # Sync random seed across TP workers.
         # Elastic joiners cannot enter the launch-time WORLD broadcast.
         if server_args.is_ep_joiner:
-            self.random_seed = server_args.random_seed
+            self.random_seed = get_device().random_seed
         else:
             self.random_seed = broadcast_pyobj(
-                [server_args.random_seed],
+                [get_device().random_seed],
                 self.ps.tp_size * self.ps.pp_rank + self.ps.tp_rank,
                 self.world_group.cpu_group,
                 src=self.world_group.ranks[0],
             )[0]
         set_random_seed(self.random_seed)
 
-        self.enable_overlap = not server_args.disable_overlap_schedule
-        self.enable_spec = server_args.speculative_algorithm is not None
+        self.enable_overlap = not get_schedule().disable_overlap_schedule
+        self.enable_spec = get_spec().speculative_algorithm is not None
         self.hicache_layer_transfer_counter = None
 
     def alloc_memory_pool(

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -65,7 +65,11 @@ from sglang.srt.observability.metrics_collector import (
     StorageMetricsCollector,
     resolve_collector_class,
 )
-from sglang.srt.runtime_context import get_memory
+from sglang.srt.runtime_context import (
+    get_memory,
+    get_observability,
+    get_serving,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.mem_cache.cache_init_params import CacheInitParams
@@ -88,9 +92,9 @@ class HiRadixCache(RadixCache):
             self.token_to_kv_pool_host = get_mha_host_pool_cls(self.kv_cache)(
                 self.kv_cache,
                 get_memory().hicache_ratio,
-                server_args.hicache_size,
+                get_memory().hicache_size,
                 self.page_size,
-                server_args.hicache_mem_layout,
+                get_memory().hicache_mem_layout,
                 allocator_type=allocator_type,
             )
         elif isinstance(self.kv_cache, DSATokenToKVPool):
@@ -106,9 +110,9 @@ class HiRadixCache(RadixCache):
             self.token_to_kv_pool_host = MLATokenToKVPoolHost(
                 self.kv_cache,
                 get_memory().hicache_ratio,
-                server_args.hicache_size,
+                get_memory().hicache_size,
                 self.page_size,
-                server_args.hicache_mem_layout,
+                get_memory().hicache_mem_layout,
                 allocator_type=allocator_type,
                 dcp_size=_parallel.attn_dcp_size,
                 dcp_rank=_parallel.attn_dcp_rank,
@@ -123,9 +127,9 @@ class HiRadixCache(RadixCache):
         self.tp_world_size = torch.distributed.get_world_size(group=self.tp_group)
         self.pp_rank = params.pp_rank
         self.pp_size = params.pp_size
-        self.enable_storage = server_args.hicache_storage_backend is not None
+        self.enable_storage = get_memory().hicache_storage_backend is not None
         self.enable_storage_metrics = self.enable_storage and params.enable_metrics
-        self.extra_metric_labels = server_args.extra_metric_labels
+        self.extra_metric_labels = get_observability().extra_metric_labels
 
         (
             extra_config,
@@ -133,11 +137,11 @@ class HiRadixCache(RadixCache):
             prefetch_timeout_config,
             hicache_storage_pass_prefix_keys,
         ) = self._parse_storage_backend_extra_config(
-            server_args.hicache_storage_backend_extra_config
+            get_memory().hicache_storage_backend_extra_config
         )
         # TODO: support more timeout check functions
         self.is_prefetch_timeout = self._prefetch_timeout_check_linear_func
-        self.prefetch_stop_policy = server_args.hicache_storage_prefetch_policy
+        self.prefetch_stop_policy = get_memory().hicache_storage_prefetch_policy
 
         self.load_cache_event = threading.Event()
         if isinstance(self.kv_cache, DSATokenToKVPool):
@@ -174,16 +178,16 @@ class HiRadixCache(RadixCache):
                 attn_cp_group=self.attn_cp_group,
                 attn_tp_group=self.attn_tp_group,
                 pp_group=self.pp_group,
-                write_policy=server_args.hicache_write_policy,
-                io_backend=server_args.hicache_io_backend,
-                storage_backend=server_args.hicache_storage_backend,
+                write_policy=get_memory().hicache_write_policy,
+                io_backend=get_memory().hicache_io_backend,
+                storage_backend=get_memory().hicache_storage_backend,
                 prefetch_threshold=prefetch_threshold,
-                model_name=server_args.served_model_name,
+                model_name=get_serving().served_model_name,
                 storage_backend_extra_config=extra_config,
                 enable_storage_metrics=self.enable_storage_metrics,
             )
         self._apply_storage_runtime_config(
-            storage_backend=server_args.hicache_storage_backend,
+            storage_backend=get_memory().hicache_storage_backend,
             prefetch_threshold=prefetch_threshold,
             prefetch_timeout_config=prefetch_timeout_config,
             hicache_storage_pass_prefix_keys=hicache_storage_pass_prefix_keys,
@@ -205,7 +209,7 @@ class HiRadixCache(RadixCache):
         self.work_list: List[torch.distributed.Work] = []
         # todo: dynamically adjust the threshold
         self.write_through_threshold = (
-            1 if server_args.hicache_write_policy == "write_through" else 2
+            1 if get_memory().hicache_write_policy == "write_through" else 2
         )
         self.load_back_threshold = 10
         # Detach storage backend automatically on process shutdown

python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py

@@ -28,7 +28,11 @@ from sglang.srt.mem_cache.pool_host.mha import (
 )
 from sglang.srt.mem_cache.pool_host.mla import MLATokenToKVPoolHost
 from sglang.srt.mem_cache.unified_cache.component_type import ComponentType
-from sglang.srt.runtime_context import get_memory, get_parallel
+from sglang.srt.runtime_context import (
+    get_memory,
+    get_parallel,
+    get_serving,
+)
 
 if TYPE_CHECKING:
     import torch
@@ -100,9 +104,9 @@ def build_kv_host_pool(
     return kv_host_pool_cls(
         kv_pool,
         get_memory().hicache_ratio,
-        server_args.hicache_size if host_size is None else host_size,
+        get_memory().hicache_size if host_size is None else host_size,
         page_size,
-        server_args.hicache_mem_layout,
+        get_memory().hicache_mem_layout,
         allocator_type=_get_allocator_type(server_args),
         pool_label=pool_label,
         **kwargs,
@@ -299,8 +303,8 @@ def build_kv_only_stack(
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
         pp_group=params.pp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -340,9 +344,9 @@ def build_hybrid_swa_stack(
     )
 
     kv_host_size = swa_host_size = None
-    if server_args.hicache_size > 0:
+    if get_memory().hicache_size > 0:
         kv_host_size, swa_host_size = _split_hicache_size(
-            server_args.hicache_size, (full_kv_pool, swa_kv_pool)
+            get_memory().hicache_size, (full_kv_pool, swa_kv_pool)
         )
 
     host_pool_group = build_hybrid_swa_group(
@@ -370,8 +374,8 @@ def build_hybrid_swa_stack(
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
         pp_group=params.pp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -399,7 +403,7 @@ def _deepseek_v4_num_host_pages(
 
     device_swa_pages = (kvcache.swa_size + swa_page_size - 1) // swa_page_size
 
-    if server_args.hicache_size > 0:
+    if get_memory().hicache_size > 0:
         raise ValueError(
             "DeepSeek V4 HiCache currently does not support --hicache-size; "
             "use --hicache-ratio instead."
@@ -496,7 +500,7 @@ def build_deepseek_v4_hicache_stack(
     )
 
     logical_host_pool = LogicalHostPool(
-        num_host_pages * page_size, page_size, layout=server_args.hicache_mem_layout
+        num_host_pages * page_size, page_size, layout=get_memory().hicache_mem_layout
     )
     entries = [
         build_pool_entry(
@@ -519,7 +523,7 @@ def build_deepseek_v4_hicache_stack(
             item_bytes=kvcache.swa_kv_pool.bytes_per_page_padded,
             num_host_pages=swa_num_host_pages,
             slot_page_size=kvcache.swa_page_size,
-            layout=server_args.hicache_mem_layout,
+            layout=get_memory().hicache_mem_layout,
             allocator_type=_get_allocator_type(server_args),
         )
         swa_attn_allocator = params.token_to_kv_pool_allocator.swa_attn_allocator
@@ -545,7 +549,7 @@ def build_deepseek_v4_hicache_stack(
             item_bytes=c4_item_bytes,
             num_host_pages=num_host_pages,
             slot_page_size=page_size,
-            layout=server_args.hicache_mem_layout,
+            layout=get_memory().hicache_mem_layout,
             allocator_type=_get_allocator_type(server_args),
         )
         c4_indexer_host_pool = DeepSeekV4PagedHostPool(
@@ -557,7 +561,7 @@ def build_deepseek_v4_hicache_stack(
             ),
             num_host_pages=num_host_pages,
             slot_page_size=page_size,
-            layout=server_args.hicache_mem_layout,
+            layout=get_memory().hicache_mem_layout,
             allocator_type=_get_allocator_type(server_args),
         )
         entries.extend(
@@ -588,7 +592,7 @@ def build_deepseek_v4_hicache_stack(
                 ],
                 num_host_pages=swa_num_host_pages,
                 swa_page_size=kvcache.swa_page_size,
-                layout=server_args.hicache_mem_layout,
+                layout=get_memory().hicache_mem_layout,
                 allocator_type=_get_allocator_type(server_args),
             )
             c4_indexer_state_host_pool = DeepSeekV4StateHostPool(
@@ -599,7 +603,7 @@ def build_deepseek_v4_hicache_stack(
                 ],
                 num_host_pages=swa_num_host_pages,
                 swa_page_size=kvcache.swa_page_size,
-                layout=server_args.hicache_mem_layout,
+                layout=get_memory().hicache_mem_layout,
                 allocator_type=_get_allocator_type(server_args),
             )
             entries.extend(
@@ -631,7 +635,7 @@ def build_deepseek_v4_hicache_stack(
             item_bytes=c128_item_bytes,
             num_host_pages=num_host_pages,
             slot_page_size=page_size,
-            layout=server_args.hicache_mem_layout,
+            layout=get_memory().hicache_mem_layout,
             allocator_type=_get_allocator_type(server_args),
         )
         # C128 state pool is intentionally not registered with hicache.
@@ -658,8 +662,8 @@ def build_deepseek_v4_hicache_stack(
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
         pp_group=params.pp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -697,9 +701,9 @@ def build_hybrid_mamba_stack(
         pool.full_kv_pool for pool in params.mtp_draft_device_pools
     )
     kv_host_size, mamba_host_size = None, 0
-    if server_args.hicache_size > 0:
+    if get_memory().hicache_size > 0:
         kv_host_size, mamba_host_size = _split_hicache_size(
-            server_args.hicache_size, (kv_pool, mamba_pool)
+            get_memory().hicache_size, (kv_pool, mamba_pool)
         )
     kv_host_pool = build_kv_host_pool(
         kv_pool=kv_pool,
@@ -721,7 +725,7 @@ def build_hybrid_mamba_stack(
         get_memory().hicache_ratio,
         mamba_host_size,
         allocator_type=_get_allocator_type(server_args),
-        layout=server_args.hicache_mem_layout,
+        layout=get_memory().hicache_mem_layout,
     )
     entries = [
         build_pool_entry(
@@ -754,8 +758,8 @@ def build_hybrid_mamba_stack(
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
         pp_group=params.pp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -801,9 +805,9 @@ def build_hybrid_mamba_swa_stack(
     swa_attn_allocator = params.token_to_kv_pool_allocator.swa_attn_allocator
     mamba_allocator = params.req_to_token_pool.mamba_allocator
     kv_host_size, swa_host_size, mamba_host_size = None, None, 0
-    if server_args.hicache_size > 0:
+    if get_memory().hicache_size > 0:
         kv_host_size, swa_host_size, mamba_host_size = _split_hicache_size(
-            server_args.hicache_size, (full_kv_pool, swa_kv_pool, mamba_pool)
+            get_memory().hicache_size, (full_kv_pool, swa_kv_pool, mamba_pool)
         )
     kv_host_pool = build_kv_host_pool(
         kv_pool=full_kv_pool,
@@ -825,8 +829,8 @@ def build_hybrid_mamba_swa_stack(
         mamba_pool,
         get_memory().hicache_ratio,
         mamba_host_size,
-        allocator_type=server_args.hicache_storage_backend,
-        layout=server_args.hicache_mem_layout,
+        allocator_type=get_memory().hicache_storage_backend,
+        layout=get_memory().hicache_mem_layout,
     )
     entries = [
         build_pool_entry(
@@ -870,8 +874,8 @@ def build_hybrid_mamba_swa_stack(
         attn_cp_group=attn_cp_group,
         attn_tp_group=attn_tp_group,
         pp_group=pp_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -948,8 +952,8 @@ def build_anchor_sidecar_stack(
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
         pp_group=params.pp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -1019,7 +1023,7 @@ def build_full_draft_pools(
         pool=pool,
         host_to_device_ratio=host_pool_group.logical_size / pool.size,
         page_size=controller.page_size,
-        layout=server_args.hicache_mem_layout,
+        layout=get_memory().hicache_mem_layout,
         allocator_type=_get_allocator_type(server_args),
         pool_label="draft",
     )
@@ -1045,7 +1049,7 @@ def build_full_draft_pools(
         indexer_host_pool = DSAIndexerPoolHost(
             pool,
             draft_host_pool,
-            server_args.hicache_mem_layout,
+            get_memory().hicache_mem_layout,
             allocator_type=_get_allocator_type(server_args),
         )
         specs.append(
@@ -1502,7 +1506,7 @@ class _DsaStrategy(StackStrategy):
             sidecar_host_pool_factory=lambda kv_host_pool: DSAIndexerPoolHost(
                 full_kv_pool,
                 kv_host_pool,
-                server_args.hicache_mem_layout,
+                get_memory().hicache_mem_layout,
                 allocator_type=_get_allocator_type(server_args),
             ),
             prefetch_threshold=prefetch_threshold,
@@ -1732,7 +1736,7 @@ def attach_hybrid_pool_to_unified_cache(
             storage_backend=storage_backend,
             storage_backend_extra_config=storage_extra_config,
             prefetch_threshold=storage_prefetch_threshold,
-            model_name=server_args.served_model_name,
+            model_name=get_serving().served_model_name,
             enable_storage_metrics=cache._enable_metrics_flag,
         )
         _apply_stack_result(cache, kvcache, params, result)
@@ -1796,8 +1800,8 @@ def build_minimax_sparse_hicache_stack(
         index_host_pool = MHATokenToKOnlyPoolHost(
             index_k_pool,
             kv_host_pool,
-            server_args.hicache_mem_layout,
-            allocator_type=server_args.hicache_storage_backend,
+            get_memory().hicache_mem_layout,
+            allocator_type=get_memory().hicache_storage_backend,
         )
         entries.append(
             build_pool_entry(
@@ -1821,8 +1825,8 @@ def build_minimax_sparse_hicache_stack(
         load_cache_event=load_cache_event,
         attn_cp_group=params.attn_cp_cache_group,
         attn_tp_group=params.attn_tp_cache_group,
-        write_policy=server_args.hicache_write_policy,
-        io_backend=server_args.hicache_io_backend,
+        write_policy=get_memory().hicache_write_policy,
+        io_backend=get_memory().hicache_io_backend,
         storage_backend=storage_backend,
         prefetch_threshold=prefetch_threshold,
         model_name=model_name,
@@ -1871,10 +1875,10 @@ def attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
                     layer_id: layer_id for layer_id in range(main_pool.layer_num)
                 },
                 load_cache_event=load_cache_event,
-                storage_backend=server_args.hicache_storage_backend,
+                storage_backend=get_memory().hicache_storage_backend,
                 use_mla=False,
                 prefetch_threshold=prefetch_threshold,
-                model_name=server_args.served_model_name,
+                model_name=get_serving().served_model_name,
                 storage_backend_extra_config=extra_config,
                 enable_storage_metrics=enable_storage_metrics,
             )
@@ -1885,9 +1889,9 @@ def attach_hybrid_minimax_sparse_pool_to_hiradix_cache(
                 server_args=server_args,
                 sparse_pool=sparse_pool,
                 load_cache_event=load_cache_event,
-                storage_backend=server_args.hicache_storage_backend,
+                storage_backend=get_memory().hicache_storage_backend,
                 prefetch_threshold=prefetch_threshold,
-                model_name=server_args.served_model_name,
+                model_name=get_serving().served_model_name,
                 storage_backend_extra_config=extra_config,
                 enable_storage_metrics=enable_storage_metrics,
             )
@@ -1933,17 +1937,17 @@ def attach_hybrid_dsa_pool_to_hiradix_cache(
             sidecar_pool_name=PoolName.INDEXER,
             full_layer_mapping=layer_mapping,
             load_cache_event=load_cache_event,
-            storage_backend=server_args.hicache_storage_backend,
+            storage_backend=get_memory().hicache_storage_backend,
             use_mla=True,
             override_kv_cache_dim=kv.kv_cache_dim,
             prefetch_threshold=prefetch_threshold,
             sidecar_host_pool_factory=lambda kv_host_pool: DSAIndexerPoolHost(
                 kv,
                 kv_host_pool,
-                server_args.hicache_mem_layout,
+                get_memory().hicache_mem_layout,
                 allocator_type=_get_allocator_type(server_args),
             ),
-            model_name=server_args.served_model_name,
+            model_name=get_serving().served_model_name,
             storage_backend_extra_config=extra_config,
             enable_storage_metrics=enable_storage_metrics,
         )

python/sglang/srt/mem_cache/radix_cache_cpp.py

@@ -22,6 +22,9 @@ from sglang.srt.mem_cache.cpp_radix_tree.radix_tree import (
     TreeNodeCpp,
 )
 from sglang.srt.mem_cache.radix_cache import RadixKey
+from sglang.srt.runtime_context import (
+    get_memory,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
@@ -59,7 +62,7 @@ class RadixCacheCpp(BasePrefixCache):
         self.ongoing_load_back: Set[IOHandle] = set()
         # todo: dynamically adjust the threshold
         self.write_through_threshold = (
-            1 if server_args.hicache_write_policy == "write_through" else 2
+            1 if get_memory().hicache_write_policy == "write_through" else 2
         )
         self.token_to_kv_pool_allocator = params.token_to_kv_pool_allocator
         self.device = self.token_to_kv_pool_allocator.device
@@ -72,7 +75,7 @@ class RadixCacheCpp(BasePrefixCache):
         if params.enable_metrics:
             self.init_metrics_collector()
 
-        if not server_args.enable_hierarchical_cache:
+        if not get_memory().enable_hierarchical_cache:
             self.tree = RadixTreeCpp(
                 disabled=self.disable,
                 page_size=self.page_size,

python/sglang/srt/model_executor/model_runner.py

@@ -170,10 +170,13 @@ from sglang.srt.platforms import current_platform
 from sglang.srt.runtime_context import (
     ensure_published,
     get_context,
+    get_device,
     get_exec,
     get_global_dwdp_manager,
     get_lora,
+    get_memory,
     get_model,
+    get_observability,
     get_parallel,
     get_schedule,
     get_spec,
@@ -279,7 +282,7 @@ def resolve_draft_attention_backend(
     """
     if not is_draft_worker:
         return None
-    return draft_attention_backend or server_args.speculative_draft_attention_backend
+    return draft_attention_backend or get_spec().speculative_draft_attention_backend
 
 
 class ModelRunner:
@@ -323,7 +326,6 @@ class ModelRunner:
         # workers so they reuse target's resolved sizes (replaces legacy
         # `server_args._draft_pool_config` mutation hack).
         self.memory_pool_config = memory_pool_config
-        self.device = server_args.device
         self.gpu_id = gpu_id
         self.ps = ps
         self.model_config = model_config
@@ -340,6 +342,7 @@ class ModelRunner:
         # Set by maybe_init_lora_manager; stays None when LoRA is off and on
         # draft runners, which serve adapters' target model unadapted.
         self.lora_manager: Optional[LoRAManager] = None
+        self.device = get_device().device
         self.draft_attention_backend = resolve_draft_attention_backend(
             draft_attention_backend=draft_attention_backend,
             server_args=server_args,
@@ -356,7 +359,7 @@ class ModelRunner:
             model_config.is_multimodal_chunked_prefill_supported
         )
         self.spec_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
         self.capture_tail_hooks = []
         self.page_size = get_schedule().page_size
@@ -367,12 +370,12 @@ class ModelRunner:
         self.is_hybrid_swa_compress = model_config.is_hybrid_swa_compress
         self.use_mla_backend = self.model_config.attention_arch == AttentionArch.MLA
         self.attention_chunk_size = model_config.attention_chunk_size
-        self.enable_elastic_ep = server_args.elastic_ep_backend is not None
+        self.enable_elastic_ep = get_exec().moe.elastic_ep_backend is not None
         self.forward_pass_id = 0
         self._pending_elastic_scale_update = None
         self.init_new_workspace = False
         self.draft_model_idx = draft_model_idx
-        self.enable_hisparse = server_args.enable_hisparse
+        self.enable_hisparse = get_memory().enable_hisparse
         self._sampling_observer: Optional[SamplingObserver] = None
 
         self.init_startup_observability()
@@ -385,7 +388,7 @@ class ModelRunner:
         self.init_spec_aux_hidden_state()
 
         # Apply the rank zero filter to logger
-        if server_args.show_time_cost:
+        if get_observability().show_time_cost:
             enable_show_time_cost()
 
         misc_utils.maybe_disable_chunked_prefix_cache(
@@ -1348,7 +1351,7 @@ class ModelRunner:
                     else False
                 ),
                 speculative_draft_attention_backend=self.draft_attention_backend,
-                speculative_draft_kv_cache_dtype=self.server_args.speculative_draft_kv_cache_dtype,
+                speculative_draft_kv_cache_dtype=get_spec().speculative_draft_kv_cache_dtype,
             )
         )
         # This runner's OWN resolved dtype string (target or draft). Attention

python/sglang/srt/model_executor/model_runner_components/load_model_utils.py

@@ -25,6 +25,11 @@ from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
     trigger_init_weights_send_group_for_remote_instance_request,
 )
 from sglang.srt.platforms import current_platform
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_model,
+    get_observability,
+)
 from sglang.srt.utils.common import is_npu
 from sglang.srt.utils.network import NetworkAddress
 
@@ -86,18 +91,16 @@ def maybe_trigger_remote_instance_nccl_send_group(
     ``--speculative-draft-draft-load-format`` needs its own send group, and the
     target's format cannot answer for it."""
     if (
-        (load_format or server_args.load_format) == LoadFormat.REMOTE_INSTANCE
-        and server_args.remote_instance_weight_loader_backend
-        == RemoteInstanceWeightLoaderBackend.NCCL
-    ):
+        load_format or get_model().load_format
+    ) == LoadFormat.REMOTE_INSTANCE and get_model().remote_instance_weight_loader_backend == RemoteInstanceWeightLoaderBackend.NCCL:
         if tp_rank == 0:
             instance_ip = NetworkAddress.resolve_host(socket.gethostname())
             t = threading.Thread(
                 target=trigger_init_weights_send_group_for_remote_instance_request,
                 args=(
-                    server_args.remote_instance_weight_loader_seed_instance_ip,
-                    server_args.remote_instance_weight_loader_seed_instance_service_port,
-                    server_args.remote_instance_weight_loader_send_weights_group_ports,
+                    get_model().remote_instance_weight_loader_seed_instance_ip,
+                    get_model().remote_instance_weight_loader_seed_instance_service_port,
+                    get_model().remote_instance_weight_loader_send_weights_group_ports,
                     instance_ip,
                 ),
             )
@@ -111,12 +114,12 @@ def load_kv_cache_scales(
     defaulted: a fallback to ``server_args`` would be a hidden global read for
     any future caller that forgets to pass one."""
     if kv_cache_dtype == "fp8_e4m3":
-        if server_args.quantization_param_path is not None:
+        if get_model().quantization_param_path is not None:
             if callable(getattr(model, "load_kv_cache_scales", None)):
-                model.load_kv_cache_scales(server_args.quantization_param_path)
+                model.load_kv_cache_scales(get_model().quantization_param_path)
                 logger.info(
                     "Loaded KV cache scaling factors from %s",
-                    server_args.quantization_param_path,
+                    get_model().quantization_param_path,
                 )
             else:
                 raise RuntimeError(
@@ -154,13 +157,13 @@ def report_online_quantization(*, model, server_args: ServerArgs) -> None:
         getattr(model, "quant_config", None), "quantized_layers", None
     )
     if (
-        server_args.quantization is not None
+        get_model().quantization is not None
         and isinstance(quantized_layers, tuple)
         and len(quantized_layers) == 2
     ):
         layer_types, quantized_layers_count = quantized_layers
         logger.info(
-            f"Online {server_args.quantization} quantization: quantized {quantized_layers_count} layers of types: {layer_types}"
+            f"Online {get_model().quantization} quantization: quantized {quantized_layers_count} layers of types: {layer_types}"
         )
 
 
@@ -174,15 +177,15 @@ def maybe_register_debug_tensor_dump_hook(
     tp_rank: int,
     pp_rank: int,
 ) -> None:
-    if server_args.debug_tensor_dump_output_folder is not None:
-        dump_folder = server_args.debug_tensor_dump_output_folder
+    if get_observability().debug_tensor_dump_output_folder is not None:
+        dump_folder = get_observability().debug_tensor_dump_output_folder
         if spec_algorithm.is_eagle():
             role = "draft" if is_draft_worker else "target"
             dump_folder = os.path.join(dump_folder, role)
         register_forward_hook_for_model(
             model,
             dump_folder,
-            server_args.debug_tensor_dump_layers,
+            get_observability().debug_tensor_dump_layers,
             tp_size,
             tp_rank,
             pp_rank,
@@ -203,28 +206,28 @@ def build_load_config(
     from sglang.srt.configs.modelopt_config import ModelOptConfig
 
     modelopt_config = ModelOptConfig(
-        quant=server_args.modelopt_quant,
-        checkpoint_restore_path=server_args.modelopt_checkpoint_restore_path,
-        checkpoint_save_path=server_args.modelopt_checkpoint_save_path,
-        export_path=server_args.modelopt_export_path,
-        quantize_and_serve=server_args.quantize_and_serve,
+        quant=get_model().modelopt_quant,
+        checkpoint_restore_path=get_model().modelopt_checkpoint_restore_path,
+        checkpoint_save_path=get_model().modelopt_checkpoint_save_path,
+        export_path=get_model().modelopt_export_path,
+        quantize_and_serve=get_model().quantize_and_serve,
     )
 
     return LoadConfig(
-        load_format=load_format or server_args.load_format,
-        download_dir=server_args.download_dir,
-        model_loader_extra_config=server_args.model_loader_extra_config,
+        load_format=load_format or get_model().load_format,
+        download_dir=get_model().download_dir,
+        model_loader_extra_config=get_model().model_loader_extra_config,
         tp_rank=tp_rank,
-        remote_instance_weight_loader_seed_instance_ip=server_args.remote_instance_weight_loader_seed_instance_ip,
-        remote_instance_weight_loader_seed_instance_service_port=server_args.remote_instance_weight_loader_seed_instance_service_port,
-        remote_instance_weight_loader_send_weights_group_ports=server_args.remote_instance_weight_loader_send_weights_group_ports,
-        remote_instance_weight_loader_backend=server_args.remote_instance_weight_loader_backend,
+        remote_instance_weight_loader_seed_instance_ip=get_model().remote_instance_weight_loader_seed_instance_ip,
+        remote_instance_weight_loader_seed_instance_service_port=get_model().remote_instance_weight_loader_seed_instance_service_port,
+        remote_instance_weight_loader_send_weights_group_ports=get_model().remote_instance_weight_loader_send_weights_group_ports,
+        remote_instance_weight_loader_backend=get_model().remote_instance_weight_loader_backend,
         remote_instance_weight_loader_transfer_engine=remote_instance_weight_transporter_engine,
         remote_instance_weight_loader_transfer_engine_session_id=remote_instance_weight_transporter_session_id,
         modelexpress_url=server_args.modelexpress_url,
         modelexpress_transport=server_args.modelexpress_transport,
         modelopt_config=modelopt_config,
-        rl_quant_profile=server_args.rl_quant_profile,
+        rl_quant_profile=get_model().rl_quant_profile,
         draft_model_idx=draft_model_idx,
         weight_cache_mode=weight_cache_mode,
         weight_cache_socket=weight_cache_socket,
@@ -246,7 +249,7 @@ def maybe_enable_ipc_weight_cache(
     the format swap is guarded on ``!= IPC_CACHE`` so a second call (e.g. a
     weight reload) can't overwrite the captured fallback format.
     """
-    if server_args.weight_cache_mode == "off":
+    if get_model().weight_cache_mode == "off":
         return
 
     if load_config.load_format != LoadFormat.IPC_CACHE:
@@ -278,13 +281,13 @@ def load_model_with_memory_saver(
     # Remove monkey_patch when linear.py quant remove dependencies with vllm
     monkey_patch_vllm_parallel_state()
 
-    enable_cpu_backup = server_args.enable_weights_cpu_backup or (
-        is_draft_worker and server_args.enable_draft_weights_cpu_backup
+    enable_cpu_backup = get_exec().features.enable_weights_cpu_backup or (
+        is_draft_worker and get_exec().features.enable_draft_weights_cpu_backup
     )
 
     # In zero-copy IPC mode, the weights are shared with the daemon via
     # CUDA IPC and must not be offloaded/reloaded by the memory saver.
-    is_ipc_zero_copy = server_args.weight_cache_mode != "off"
+    is_ipc_zero_copy = get_model().weight_cache_mode != "off"
     if is_ipc_zero_copy and enable_cpu_backup:
         logger.warning(
             "[ModelRunner] Disabling weights CPU backup in zero-copy IPC mode — "

python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py

@@ -6,6 +6,11 @@ from typing import TYPE_CHECKING, Any, Optional
 import msgspec
 
 from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.runtime_context import (
+    configured_tp_size,
+    get_model,
+    get_spec,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.server_args import ServerArgs
@@ -75,13 +80,13 @@ def _resolve_eagle_aux_hidden_state(
     if (
         (spec_algorithm.is_eagle() or spec_algorithm.is_standalone())
         and not is_draft_worker
-        and server_args.speculative_draft_model_path
+        and get_spec().speculative_draft_model_path
     ):
         # Load draft config to get layer count for KV cache sizing
         draft_model_config = ModelConfig.from_server_args(
             server_args,
-            model_path=server_args.speculative_draft_model_path,
-            model_revision=server_args.speculative_draft_model_revision,
+            model_path=get_spec().speculative_draft_model_path,
+            model_revision=get_spec().speculative_draft_model_revision,
             is_draft_model=True,
         )
         num_nextn_predict_layers = draft_model_config.num_nextn_predict_layers
@@ -134,8 +139,8 @@ def _resolve_dflash_aux_hidden_state(
         # Select target layers to capture for building draft context features.
         draft_model_config = ModelConfig.from_server_args(
             server_args,
-            model_path=(server_args.speculative_draft_model_path),
-            model_revision=server_args.speculative_draft_model_revision,
+            model_path=(get_spec().speculative_draft_model_path),
+            model_revision=get_spec().speculative_draft_model_revision,
             is_draft_model=True,
         )
         dflash_draft_config = parse_dflash_draft_config(
@@ -221,23 +226,23 @@ def _resolve_dflash_draft_cell_size(
 
     try:
         _, draft_kv_cache_dtype = configure_kv_cache_dtype(
-            server_args_kv_cache_dtype=server_args.kv_cache_dtype,
+            server_args_kv_cache_dtype=get_model().kv_cache_dtype,
             speculative_draft_kv_cache_dtype=(
-                server_args.speculative_draft_kv_cache_dtype
+                get_spec().speculative_draft_kv_cache_dtype
             ),
             model=None,
             model_dtype=draft_model_config.dtype,
             is_draft_worker=True,
             is_dflash=True,
             speculative_draft_attention_backend=(
-                server_args.speculative_draft_attention_backend
+                get_spec().speculative_draft_attention_backend
             ),
         )
         return dflash_draft_cell_size_per_token(
             draft_model_config=draft_model_config,
             draft_num_layers=draft_num_layers,
             draft_kv_cache_dtype=draft_kv_cache_dtype,
-            tp_size=server_args.tp_size,
+            tp_size=configured_tp_size(),
         )
     except Exception as e:  # noqa: BLE001
         logger.warning(

python/sglang/srt/model_executor/model_runner_components/startup_weight_load.py

@@ -20,7 +20,18 @@ from sglang.srt.model_loader.weight_utils import (
     CheckpointFilePrefetchHandle,
 )
 from sglang.srt.platforms import current_platform
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import (
+    configured_attn_cp_size,
+    configured_dcp_size,
+    configured_pp_size,
+    configured_tp_size,
+    get_device,
+    get_exec,
+    get_lora,
+    get_model,
+    get_parallel,
+    get_spec,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
@@ -97,37 +108,37 @@ class StartupWeightLoadOptions:
         server_args: ServerArgs,
         is_draft_worker: bool,
     ) -> StartupWeightLoadOptions:
-        cuda_graph_config = server_args.cuda_graph_config
+        cuda_graph_config = get_exec().graph.cuda_graph_config
         cuda_graph_enabled = any(
             getattr(cuda_graph_config, phase).backend != Backend.DISABLED
             for phase in Phase.ALL
         )
         return cls(
-            device=server_args.device,
+            device=get_device().device,
             is_cuda_platform=current_platform.is_cuda(),
             cuda_graph_enabled=cuda_graph_enabled,
             prefill_cuda_graph_backend=cuda_graph_config.prefill.backend,
             is_draft_worker=is_draft_worker,
-            speculative_algorithm=server_args.speculative_algorithm,
-            tp_size=server_args.tp_size,
-            attn_cp_size=server_args.attn_cp_size,
-            dcp_size=server_args.dcp_size,
-            pp_size=server_args.pp_size,
+            speculative_algorithm=get_spec().speculative_algorithm,
+            tp_size=configured_tp_size(),
+            attn_cp_size=configured_attn_cp_size(),
+            dcp_size=configured_dcp_size(),
+            pp_size=configured_pp_size(),
             dp_size=get_parallel().dp_size,
             ep_size=get_parallel().ep_size,
-            cpu_offload_gb=server_args.cpu_offload_gb,
-            offload_group_size=server_args.offload_group_size,
-            enable_memory_saver=server_args.enable_memory_saver,
-            enable_weights_cpu_backup=server_args.enable_weights_cpu_backup,
-            enable_lora=server_args.enable_lora,
-            has_lora_paths=bool(server_args.lora_paths),
-            weight_loader_disable_mmap=server_args.weight_loader_disable_mmap,
+            cpu_offload_gb=get_exec().offload.cpu_offload_gb,
+            offload_group_size=get_exec().offload.offload_group_size,
+            enable_memory_saver=get_exec().features.enable_memory_saver,
+            enable_weights_cpu_backup=get_exec().features.enable_weights_cpu_backup,
+            enable_lora=get_lora().enable_lora,
+            has_lora_paths=bool(get_lora().lora_paths),
+            weight_loader_disable_mmap=get_model().weight_loader_disable_mmap,
             weight_loader_drop_cache_after_load=(
-                server_args.weight_loader_drop_cache_after_load
+                get_model().weight_loader_drop_cache_after_load
             ),
-            has_custom_weight_loader=bool(server_args.custom_weight_loader),
-            enable_torch_compile=server_args.enable_torch_compile,
-            prefetch_num_threads=server_args.weight_loader_prefetch_num_threads,
+            has_custom_weight_loader=bool(get_model().custom_weight_loader),
+            enable_torch_compile=get_exec().graph.enable_torch_compile,
+            prefetch_num_threads=get_model().weight_loader_prefetch_num_threads,
         )
 
 

python/sglang/srt/observability/metrics_collector.py

@@ -26,6 +26,12 @@ from typing import TYPE_CHECKING, Any, Dict, List, Mapping, Optional, Set, Union
 from sglang.srt.disaggregation.utils import DisaggregationMode
 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 (
+    get_disagg,
+    get_observability,
+    get_schedule,
+    get_serving,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.utils import get_bool_env_var
 from sglang.srt.utils.gauge_histogram import GaugeHistogram
@@ -957,7 +963,7 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
         # =================================================================
         # Prefill delayer
         # =================================================================
-        max_delay = server_args.prefill_delayer_max_delay_passes
+        max_delay = get_schedule().prefill_delayer_max_delay_passes
         self.prefill_delayer_wait_forward_passes = Histogram(
             name="sglang:prefill_delayer_wait_forward_passes",
             documentation="Histogram of forward passes waited by prefill delayer.",
@@ -966,7 +972,7 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
                 set(
                     x
                     for x in (
-                        server_args.prefill_delayer_forward_passes_buckets
+                        get_schedule().prefill_delayer_forward_passes_buckets
                         or [5, 20, 50, 100, 200]
                     )
                     if x < max_delay
@@ -981,7 +987,7 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
             labelnames=labels.keys(),
             buckets=sorted(
                 set(
-                    server_args.prefill_delayer_wait_seconds_buckets
+                    get_schedule().prefill_delayer_wait_seconds_buckets
                     or [1, 2, 5, 10, 20, 50, 100, 200, 500]
                 )
                 # Need bucket "<=0" for zero-delay cases
@@ -1077,13 +1083,14 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
         enable_lora: bool,
         enable_hierarchical_cache: bool,
     ) -> SchedulerMetricsCollectorContext:
-        enable_metrics = server_args.enable_metrics
+        enable_metrics = get_observability().enable_metrics
         is_stats_logging_rank = ps.attn_tp_rank == 0
         current_scheduler_metrics_enabled = enable_metrics and (
-            is_stats_logging_rank or server_args.enable_metrics_for_all_schedulers
+            is_stats_logging_rank
+            or get_observability().enable_metrics_for_all_schedulers
         )
         enable_kv_cache_events = bool(
-            server_args.kv_events_config
+            get_observability().kv_events_config
             and ps.pp_rank == 0
             and ps.attn_tp_rank == 0
             and ps.attn_cp_rank == 0
@@ -1091,10 +1098,10 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
         collector: Optional[SchedulerMetricsCollector] = None
         if enable_metrics:
             engine_type = DisaggregationMode.to_engine_type(
-                server_args.disaggregation_mode
+                get_disagg().disaggregation_mode
             )
             labels = {
-                "model_name": server_args.served_model_name,
+                "model_name": get_serving().served_model_name,
                 "engine_type": engine_type,
                 "tp_rank": tp_rank,
                 "pp_rank": pp_rank,
@@ -1104,8 +1111,8 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
                 labels["priority"] = ""
             if dp_rank is not None:
                 labels["dp_rank"] = dp_rank
-            if server_args.extra_metric_labels:
-                labels.update(server_args.extra_metric_labels)
+            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
             )
@@ -1113,7 +1120,7 @@ class SchedulerMetricsCollector(_StatLoggerDIMixin):
                 labels=labels,
                 enable_lora=enable_lora,
                 enable_hierarchical_cache=enable_hierarchical_cache,
-                enable_streaming_session=server_args.enable_streaming_session,
+                enable_streaming_session=get_serving().enable_streaming_session,
                 server_args=server_args,
             )
         return SchedulerMetricsCollectorContext(
@@ -1568,7 +1575,7 @@ class TokenizerMetricsCollector(_StatLoggerDIMixin):
             documentation="Histogram of prompt token length.",
             labelnames=labels.keys(),
             buckets=generate_buckets(
-                server_args.prompt_tokens_buckets, default_bucket_prompt_tokens
+                get_observability().prompt_tokens_buckets, default_bucket_prompt_tokens
             ),
         )
         self.uncached_prompt_tokens_histogram = Histogram(
@@ -1576,7 +1583,7 @@ class TokenizerMetricsCollector(_StatLoggerDIMixin):
             documentation="Histogram of uncached (compute) prompt token length.",
             labelnames=labels.keys(),
             buckets=generate_buckets(
-                server_args.prompt_tokens_buckets, default_bucket_prompt_tokens
+                get_observability().prompt_tokens_buckets, default_bucket_prompt_tokens
             ),
         )
         self.generation_tokens_histogram = Histogram(
@@ -1584,7 +1591,7 @@ class TokenizerMetricsCollector(_StatLoggerDIMixin):
             documentation="Histogram of generation token length.",
             labelnames=labels.keys(),
             buckets=generate_buckets(
-                server_args.generation_tokens_buckets,
+                get_observability().generation_tokens_buckets,
                 default_bucket_prompt_tokens,
             ),
         )

python/sglang/srt/runtime_context.py

@@ -678,18 +678,22 @@ class _ConfigBag:
 
 
 def _build_config_bags(server_args: Any) -> dict:
-    """Snapshot resolved ``server_args`` into the namespace bag tree, driven by
-    the ``NS(...)`` metadata on the dataclass fields. Returns
+    """Snapshot the resolution result into the namespace bag tree, driven by
+    the ``NS(...)`` metadata on the dataclass fields. Each leaf comes from
+    ``resolution_result`` -- the declaration if resolution made one, else what
+    the caller supplied -- rather than from the field, which carries the same
+    value only while declarations still materialize. Returns
     ``{top_level_name: _ConfigBag}``, arbitrarily nested (``exec.moe.eplb.…``).
     Only dataclass fields carry ``NS`` markers, so derived properties/methods are
     naturally excluded (they stay on the bag). A name used as both a leaf and a
     subgroup at the same level is a hard error — no silent shadowing."""
     from sglang.srt.arg_groups.arg_utils import namespace_of
+    from sglang.srt.arg_groups.overrides import resolution_result
 
     _MISSING = object()
     tops: dict = {}
     for field, path in namespace_of(type(server_args)).items():
-        value = getattr(server_args, field, _MISSING)
+        value = resolution_result(server_args, field, _MISSING)
         if value is _MISSING:
             # Every NS-declared field is a dataclass field, so a resolved config
             # always carries it; a miss means a malformed/partial config object
@@ -1028,7 +1032,10 @@ class _ServerArgsOverride:
         self._prev_publish_role = ctx._publish_role
         self._prev_parallel_config = ctx.parallel._config
         self._prev_capture = ctx.flags.capture.enable_torch_compile
-        from sglang.srt.arg_groups.overrides import _apply_fields
+        from sglang.srt.arg_groups.overrides import (
+            _apply_fields,
+            declare_resolution,
+        )
 
         server_args = ServerArgs(model_path="dummy")
         # Underscore names seed private property caches (the strict guard
@@ -1040,7 +1047,18 @@ class _ServerArgsOverride:
             raise ValueError(
                 f"override_server_args: unknown ServerArgs field(s): {sorted(unknown)}"
             )
-        _apply_fields(server_args, self._fields)
+        # Declared so the projection sees it.
+        # Underscore names are not fields at all (they seed private property
+        # caches), so they stay a direct write.
+        declared = {
+            name: value for name, value in self._fields.items() if name[0] != "_"
+        }
+        if declared:
+            declare_resolution(server_args, "override_server_args", **declared)
+        _apply_fields(
+            server_args,
+            {name: value for name, value in self._fields.items() if name[0] == "_"},
+        )
         # The dummy boundary skips materialization, which would leave the
         # strict mutation guard unarmed on the published object — mark it
         # materialized so bare post-publish writes raise like they do on a
@@ -1662,6 +1680,10 @@ def configured_attn_cp_size() -> int:
     return _configured_parallel("attn_cp_size")
 
 
+def configured_dcp_size() -> int:
+    return _configured_parallel("dcp_size")
+
+
 def is_ep_joiner() -> bool:
     """True in a process launched as an elastic-EP joiner (scale or recover).
 

python/sglang/srt/server_args.py

@@ -42,6 +42,7 @@ from sglang.srt.arg_groups.argparse_actions import (
 )
 from sglang.srt.arg_groups.overrides import (
     attention_backends_of,
+    declare_direct_writes,
     mamba_extra_buffer_lazy_of,
     mamba_extra_buffer_of,
     remote_instance_transfer_engine_of,
@@ -3677,6 +3678,12 @@ class ServerArgs:
            belong in the helper or signal that the helper should be split.
         """
 
+        # What the caller asked for, before any handler runs; this plus the
+        # stash is the resolution result the projection reads.
+        self._raw_input = {
+            field.name: getattr(self, field.name) for field in dataclasses.fields(self)
+        }
+
         # Declaration stash for the override/post-process passes. Set before any
         # short-circuit (none/dummy model paths) so run_post_process_pass and
         # direct handler invocations can rely on it even when
@@ -3742,8 +3749,13 @@ class ServerArgs:
         self._handle_mps_backends()
         self._handle_xpu_backends()
 
-        # Allow OOT platform plugins to apply server args defaults.
-        current_platform.apply_server_args_defaults(self)
+        # OOT platform plugins set fields directly (an interface this tree
+        # does not own); the diff records what they applied.
+        declare_direct_writes(
+            self,
+            f"platform:{current_platform.device_name}",
+            current_platform.apply_server_args_defaults,
+        )
 
         # Get GPU memory capacity, which is a common dependency for several configuration steps.
         gpu_mem = get_device_memory_capacity(self.device)

python/sglang/srt/speculative/dflash_worker_v2.py

@@ -33,7 +33,12 @@ from sglang.srt.model_executor.forward_batch_info import (
     ForwardMode,
     compute_position,
 )
-from sglang.srt.runtime_context import get_exec, get_schedule, mamba_track_grid
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_schedule,
+    get_spec,
+    mamba_track_grid,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.speculative.base_spec_worker import BaseSpecWorker
 from sglang.srt.speculative.dflash_info import DFlashVerifyInput
@@ -279,9 +284,7 @@ class DFlashWorkerV2(BaseSpecWorker):
         self._need_mamba_verify_commit = False
         self.page_size = get_schedule().page_size
         # Normalized in arg_groups.speculative_hook.handle_speculative_decoding.
-        self.draft_window_size: Optional[int] = (
-            server_args.speculative_draft_window_size
-        )
+        self.draft_window_size: Optional[int] = get_spec().speculative_draft_window_size
         self.use_compact_draft_cache = self.draft_window_size is not None
         self.device = target_worker.device
 
@@ -305,11 +308,11 @@ class DFlashWorkerV2(BaseSpecWorker):
         draft_config = parse_dflash_draft_config(
             draft_hf_config=self.draft_model_runner.model_config.hf_config
         )
-        if server_args.speculative_num_draft_tokens is None:
+        if get_spec().speculative_num_draft_tokens is None:
             # Should not happen (ServerArgs should have inferred it), but keep a fallback.
             self.block_size = int(draft_config.resolve_block_size(default=16))
         else:
-            self.block_size = int(server_args.speculative_num_draft_tokens)
+            self.block_size = int(get_spec().speculative_num_draft_tokens)
             model_block_size = draft_config.block_size
             if model_block_size is None:
                 model_block_size = getattr(self.draft_model, "block_size", None)

python/sglang/srt/speculative/dspark_components/dspark_config.py

@@ -6,6 +6,10 @@ from typing import TYPE_CHECKING, Any, List, Optional
 
 import msgspec
 
+from sglang.srt.runtime_context import (
+    get_model,
+    get_spec,
+)
 from sglang.srt.speculative.dflash_utils import parse_dflash_draft_config
 
 if TYPE_CHECKING:
@@ -25,15 +29,15 @@ def draft_is_deepseek_v4(*, server_args: ServerArgs) -> bool:
     from sglang.srt.configs.model_config import is_deepseek_v4
     from sglang.srt.utils.hf_transformers_utils import get_config
 
-    draft_model_path = server_args.speculative_draft_model_path
+    draft_model_path = get_spec().speculative_draft_model_path
     if not draft_model_path:
         return False
     draft_hf_config = get_config(
         draft_model_path,
-        trust_remote_code=server_args.trust_remote_code,
-        revision=server_args.speculative_draft_model_revision,
-        model_override_args=json.loads(server_args.json_model_override_args),
-        model_config_parser=server_args.model_config_parser,
+        trust_remote_code=get_model().trust_remote_code,
+        revision=get_spec().speculative_draft_model_revision,
+        model_override_args=json.loads(get_model().json_model_override_args),
+        model_config_parser=get_model().model_config_parser,
     )
     return draft_hf_config is not None and is_deepseek_v4(draft_hf_config)
 
@@ -126,14 +130,23 @@ def resolve_runtime_config(
 
 def read_draft_checkpoint_gamma(*, server_args: ServerArgs) -> Optional[int]:
     """Load the draft checkpoint's hf config and read its DSpark gamma
-    (block_size). Raises on config-load failure; callers pick the fallback."""
+    (block_size). Raises on config-load failure; callers pick the fallback.
+
+    Reads the *resolving* configuration, not the bags: the speculative hook
+    calls this from inside resolution, where no bag exists yet -- and the
+    caller swallows exceptions, so a bag read here does not fail loudly, it
+    silently drops the checkpoint's gamma and the cross-check with
+    `--speculative-num-draft-tokens` along with it.
+    """
+    from sglang.srt.arg_groups.overrides import resolved_view
     from sglang.srt.utils.hf_transformers_utils import get_config
 
+    resolving = resolved_view(server_args)
     draft_hf_config = get_config(
-        server_args.speculative_draft_model_path,
-        trust_remote_code=server_args.trust_remote_code,
-        revision=server_args.speculative_draft_model_revision,
-        model_override_args=json.loads(server_args.json_model_override_args),
+        resolving.speculative_draft_model_path,
+        trust_remote_code=resolving.trust_remote_code,
+        revision=resolving.speculative_draft_model_revision,
+        model_override_args=json.loads(resolving.json_model_override_args),
     )
     return parse_dspark_draft_config(draft_hf_config=draft_hf_config).resolve_gamma(
         default=None

python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py

@@ -21,6 +21,7 @@ from sglang.srt.model_executor.forward_batch_info import (
     compute_position,
 )
 from sglang.srt.runtime_context import (
+    get_disagg,
     get_exec,
     get_parallel,
     get_schedule,
@@ -106,7 +107,7 @@ class DSparkWorkerV2(BaseSpecWorker):
         self._draft_dp_context_enabled = (
             get_parallel().enable_dp_attention and not self._draft_is_moe
         )
-        self._is_pd_prefill = server_args.disaggregation_mode == "prefill"
+        self._is_pd_prefill = get_disagg().disaggregation_mode == "prefill"
         self._decode_graph_allowed = (
             not get_exec().graph.disable_cuda_graph and not self._is_pd_prefill
         )
@@ -156,7 +157,7 @@ class DSparkWorkerV2(BaseSpecWorker):
         self._target_is_mambaish = mambaish_config(target_model_config) is not None
         runtime_config = resolve_runtime_config(
             draft_hf_config=self.draft_model_runner.model_config.hf_config,
-            speculative_num_draft_tokens=server_args.speculative_num_draft_tokens,
+            speculative_num_draft_tokens=get_spec().speculative_num_draft_tokens,
             target_vocab_size=int(target_embed_rows),
         )
         self.gamma = runtime_config.gamma

python/sglang/srt/speculative/eagle_worker_v2.py

@@ -47,6 +47,7 @@ from sglang.srt.model_executor.runner import (
 )
 from sglang.srt.runtime_context import (
     get_context,
+    get_device,
     get_exec,
     get_model,
     get_parallel,
@@ -145,14 +146,14 @@ class EagleDraftWorker(EagleDraftWorkerBase):
         self.target_worker = target_worker
 
         # Args for easy access
-        self.device = server_args.device
-        self.topk = server_args.speculative_eagle_topk
+        self.device = get_device().device
+        self.topk = get_spec().speculative_eagle_topk
         if get_spec().speculative_use_rejection_sampling:
             assert self.topk == 1, "Chain speculative sampling supports only topk=1"
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self._rebuild_topk1_chain_buffers()
@@ -1060,16 +1061,16 @@ class EAGLEWorkerV2(BaseSpecWorker):
 
         # Parse arguments
         self.server_args = server_args
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.ps = ps
         self.gpu_id = gpu_id
-        self.device = server_args.device
+        self.device = get_device().device
         self._target_worker = target_worker
         self.page_size = get_schedule().page_size
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self._draft_worker = EagleDraftWorker(
@@ -1082,10 +1083,10 @@ class EAGLEWorkerV2(BaseSpecWorker):
 
         # Adaptive speculative
         self.adaptive_controller: Optional[AdaptiveController] = None
-        if server_args.speculative_adaptive:
+        if get_spec().speculative_adaptive:
             self.adaptive_controller = AdaptiveController(
                 self,
-                config_path=server_args.speculative_adaptive_config,
+                config_path=get_spec().speculative_adaptive_config,
             )
 
         # Some dummy tensors

python/sglang/srt/speculative/frozen_kv_mtp_worker_v2.py

@@ -46,6 +46,7 @@ from sglang.srt.model_executor.forward_context import ForwardContext, forward_co
 from sglang.srt.model_executor.pool_configurator import MemoryPoolConfig
 from sglang.srt.runtime_context import (
     attention_backends,
+    get_device,
     get_parallel,
     get_schedule,
     get_spec,
@@ -107,16 +108,16 @@ class FrozenKVMTPDraftWorker(EagleDraftWorkerBase, TpModelWorker):
         EagleDraftWorkerBase.__init__(self)
 
         self.server_args = server_args
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.ps = ps
         self.gpu_id = gpu_id
-        self.device = server_args.device
+        self.device = get_device().device
         self.target_worker = target_worker
         self.page_size = get_schedule().page_size
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
         assert self.speculative_algorithm.is_frozen_kv_mtp(), (
             "FrozenKVMTPDraftWorker should only be instantiated for "
@@ -695,16 +696,16 @@ class FrozenKVMTPWorkerV2(EAGLEWorkerV2):
         # an EagleDraftWorker (with its own draft KV pool). The frozen draft owns
         # no KV, so we mirror the relevant setup and build a FrozenKVMTPDraftWorker.
         self.server_args = server_args
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.ps = ps
         self.gpu_id = gpu_id
-        self.device = server_args.device
+        self.device = get_device().device
         self._target_worker = target_worker
         self.page_size = get_schedule().page_size
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self.req_to_token_pool, self.token_to_kv_pool_allocator = (
@@ -720,7 +721,7 @@ class FrozenKVMTPWorkerV2(EAGLEWorkerV2):
 
         # Frozen MTP does not wire the adaptive controller yet.
         assert (
-            not server_args.speculative_adaptive
+            not get_spec().speculative_adaptive
         ), "Frozen-KV MTP does not support adaptive speculative decoding yet."
         self.adaptive_controller = None
 

python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py

@@ -43,7 +43,12 @@ from sglang.srt.model_executor.forward_batch_info import (
     CaptureHiddenMode,
     ForwardBatch,
 )
-from sglang.srt.runtime_context import get_parallel, get_schedule
+from sglang.srt.runtime_context import (
+    get_device,
+    get_parallel,
+    get_schedule,
+    get_spec,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.speculative.base_spec_worker import BaseSpecWorker, EagleDraftWorkerBase
 from sglang.srt.speculative.draft_utils import DraftBackendFactory
@@ -128,22 +133,22 @@ class MultiLayerEagleDraftWorker(EagleDraftWorkerBase):
         self.model_config = target_worker.model_config
 
         # Args for easy access
-        self.device = server_args.device
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.device = get_device().device
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         # Leviathan/Chen rejection sampling (temp>0): the draft samples X ~ q and
         # provides q so the verify accepts iff coin*q < p and resamples the residual.
         # Single-CG runner samples in-graph (_sample_draft_proposal); per-step
         # runner samples worker-side between replays.
-        self.use_rejection_sampling = server_args.speculative_use_rejection_sampling
+        self.use_rejection_sampling = get_spec().speculative_use_rejection_sampling
         assert self.speculative_num_draft_tokens == self.speculative_num_steps + 1, (
             "multi-layer EAGLE requires speculative_num_draft_tokens == "
             "speculative_num_steps + 1, "
             f"got {self.speculative_num_draft_tokens} and {self.speculative_num_steps}"
         )
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self._rebuild_topk1_chain_buffers()
@@ -928,15 +933,15 @@ class MultiLayerEagleWorkerV2(BaseSpecWorker):
 
         # Parse arguments
         self.server_args = server_args
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.gpu_id = gpu_id
-        self.device = server_args.device
+        self.device = get_device().device
         self._target_worker = target_worker
         self.page_size = get_schedule().page_size
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self._draft_worker = MultiLayerEagleDraftWorker(

python/sglang/srt/speculative/ngram_worker.py

@@ -15,7 +15,11 @@ from sglang.srt.managers.scheduler import GenerationBatchResult
 from sglang.srt.managers.tp_worker import TpModelWorker
 from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.srt.observability.req_time_stats import set_time_batch
-from sglang.srt.runtime_context import get_schedule
+from sglang.srt.runtime_context import (
+    get_device,
+    get_schedule,
+    get_spec,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.speculative.base_spec_worker import BaseSpecWorker, EagleDraftWorkerBase
 from sglang.srt.speculative.cpp_ngram.ngram_corpus import NgramCorpus
@@ -88,19 +92,19 @@ class NGRAMWorker(BaseSpecWorker):
         super().__init__()
 
         self.server_args = server_args
-        self.enable_overlap = not server_args.disable_overlap_schedule
+        self.enable_overlap = not get_schedule().disable_overlap_schedule
         self._target_worker = target_worker
         self.model_runner = target_worker.model_runner
         self.tp_rank = ps.tp_rank
         self.page_size = get_schedule().page_size
-        self.draft_token_num: int = server_args.speculative_num_draft_tokens
-        self.max_trie_depth: int = server_args.speculative_ngram_max_trie_depth
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
+        self.draft_token_num: int = get_spec().speculative_num_draft_tokens
+        self.max_trie_depth: int = get_spec().speculative_ngram_max_trie_depth
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
         # req_to_token_pool / token_to_kv_pool_allocator are set in
         # alloc_memory_pool(), after the target pools are allocated.
-        self.device = server_args.device
+        self.device = get_device().device
 
         self.adaptive_controller = None
         # rids of the last decode batch; used to erase corpus match state for
@@ -109,26 +113,26 @@ class NGRAMWorker(BaseSpecWorker):
         self.grammar_tree_host: Optional[tuple] = None
 
         self.ngram_corpus = NgramCorpus(
-            min_bfs_breadth=server_args.speculative_ngram_min_bfs_breadth,
-            max_bfs_breadth=server_args.speculative_ngram_max_bfs_breadth,
-            match_type=server_args.speculative_ngram_match_type,
-            capacity=server_args.speculative_ngram_capacity,
-            max_trie_depth=server_args.speculative_ngram_max_trie_depth,
-            draft_token_num=server_args.speculative_num_draft_tokens,
-            external_sam_budget=server_args.speculative_ngram_external_sam_budget,
-            external_corpus_max_tokens=server_args.speculative_ngram_external_corpus_max_tokens,
+            min_bfs_breadth=get_spec().speculative_ngram_min_bfs_breadth,
+            max_bfs_breadth=get_spec().speculative_ngram_max_bfs_breadth,
+            match_type=get_spec().speculative_ngram_match_type,
+            capacity=get_spec().speculative_ngram_capacity,
+            max_trie_depth=get_spec().speculative_ngram_max_trie_depth,
+            draft_token_num=get_spec().speculative_num_draft_tokens,
+            external_sam_budget=get_spec().speculative_ngram_external_sam_budget,
+            external_corpus_max_tokens=get_spec().speculative_ngram_external_corpus_max_tokens,
         )
-        if server_args.speculative_ngram_external_corpus_path is not None:
+        if get_spec().speculative_ngram_external_corpus_path is not None:
             from sglang.srt.speculative.cpp_ngram.external_corpus import (
                 iter_external_corpus_chunks,
             )
 
-            corpus_path = server_args.speculative_ngram_external_corpus_path
+            corpus_path = get_spec().speculative_ngram_external_corpus_path
             chunks = list(
                 iter_external_corpus_chunks(
                     corpus_path,
                     target_worker.tokenizer,
-                    server_args.speculative_ngram_external_corpus_max_tokens,
+                    get_spec().speculative_ngram_external_corpus_max_tokens,
                 )
             )
             loaded = self.add_external_corpus(corpus_path, chunks)

python/sglang/srt/speculative/standalone_worker_v2.py

@@ -10,7 +10,12 @@ from sglang.srt.layers.moe.utils import (
     speculative_moe_backend_context,
 )
 from sglang.srt.managers.tp_worker import TpModelWorker
-from sglang.srt.runtime_context import get_parallel, get_schedule
+from sglang.srt.runtime_context import (
+    get_device,
+    get_parallel,
+    get_schedule,
+    get_spec,
+)
 from sglang.srt.server_args import ServerArgs
 from sglang.srt.speculative.adaptive_runtime_state import (
     AdaptiveController,
@@ -53,12 +58,12 @@ class StandaloneDraftWorker(EagleDraftWorker):
         self.target_worker = target_worker
 
         # Args for easy access
-        self.device = server_args.device
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.device = get_device().device
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         self._rebuild_topk1_chain_buffers()
@@ -158,15 +163,15 @@ class StandaloneWorkerV2(EAGLEWorkerV2):
 
         # Parse arguments
         self.server_args = server_args
-        self.topk = server_args.speculative_eagle_topk
-        self.speculative_num_steps = server_args.speculative_num_steps
-        self.speculative_num_draft_tokens = server_args.speculative_num_draft_tokens
+        self.topk = get_spec().speculative_eagle_topk
+        self.speculative_num_steps = get_spec().speculative_num_steps
+        self.speculative_num_draft_tokens = get_spec().speculative_num_draft_tokens
         self.gpu_id = gpu_id
-        self.device = server_args.device
+        self.device = get_device().device
         self._target_worker = target_worker
         self.page_size = get_schedule().page_size
         self.speculative_algorithm = SpeculativeAlgorithm.from_string(
-            server_args.speculative_algorithm
+            get_spec().speculative_algorithm
         )
 
         # Create our custom draft worker that doesn't share embeddings/lm_head

python/sglang/test/test_utils.py

@@ -2105,6 +2105,20 @@ def server_args_variant(server_args, **fields):
     }
     if unknown:
         raise ValueError(f"unknown ServerArgs field(s): {sorted(unknown)}")
+    # Reach the stash as well as the fields (the bags project from raw input
+    # + declarations); through `object` because the copy keeps its read-only
+    # guard.
+    stash = getattr(variant, "_resolved_overrides", None)
+    if stash is None:
+        stash = []
+        object.__setattr__(variant, "_resolved_overrides", stash)
+    declared = {
+        name: value
+        for name, value in fields.items()
+        if name in cls.__dataclass_fields__
+    }
+    if declared:
+        stash.append(("server_args_variant", dict(declared)))
     for name, value in fields.items():
         object.__setattr__(variant, name, value)
     return variant