config: retire the hidden global fallbacks and the mamba-extra-buffer instance reads (#34080)

python/sglang/srt/arg_groups/overrides.py

@@ -290,13 +290,25 @@ def attention_backends_of(cfg: Any) -> tuple:
 
 def mamba_extra_buffer_of(cfg: Any) -> bool:
     """Mid-resolution equivalent of runtime_context.mamba_extra_buffer_enabled:
-    reads the (possibly overlaid) strategy from a config-shaped object."""
+    reads the (possibly overlaid) strategy from a config-shaped object.
+
+    This is the one definition of the predicate: ``ServerArgs`` delegates its
+    member to it, and the runtime_context accessor is its post-publish sibling
+    (which cannot reuse it, because the two leaves land in different bags)."""
     return cfg.disable_radix_cache is False and cfg.mamba_radix_cache_strategy in (
         "extra_buffer",
         "extra_buffer_lazy",
     )
 
 
+def mamba_extra_buffer_lazy_of(cfg: Any) -> bool:
+    """The lazy variant of :func:`mamba_extra_buffer_of`."""
+    return (
+        cfg.disable_radix_cache is False
+        and cfg.mamba_radix_cache_strategy == "extra_buffer_lazy"
+    )
+
+
 def collect_model_override_declarations(
     architecture: str, server_args: Any, hf_config: Any
 ) -> List[Tuple[str, Dict[str, Any]]]:

python/sglang/srt/managers/schedule_batch.py

@@ -2,7 +2,14 @@ from __future__ import annotations
 
 from sglang.srt.dllm.config import DllmConfig
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.runtime_context import get_exec, get_schedule, get_serving, get_spec
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_schedule,
+    get_serving,
+    get_spec,
+    mamba_extra_buffer_enabled,
+    mamba_extra_buffer_lazy_enabled,
+)
 from sglang.srt.utils.common import (
     Range,
     ceil_align,
@@ -2342,7 +2349,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
     def prepare_for_extend(self):
         self.forward_mode = ForwardMode.EXTEND
-        server_args = get_server_args()
 
         if self.is_dllm():
             # For DLLM, we use a separate forward mode
@@ -2471,7 +2477,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 req.already_computed = seq_len
             req.is_retracted = False
 
-            if server_args.enable_mamba_extra_buffer():
+            if mamba_extra_buffer_enabled():
                 track_entry = self._mamba_radix_cache_v2_req_prepare_for_extend(req)
                 mamba_track_mask_cpu.append(track_entry.track_mask)
                 mamba_track_indices_cpu.append(track_entry.track_index)
@@ -2576,7 +2582,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         self.extend_logprob_start_lens = extend_logprob_start_lens
         self.extend_input_logprob_token_ids = extend_input_logprob_token_ids
 
-        if server_args.enable_mamba_extra_buffer():
+        if mamba_extra_buffer_enabled():
             self.mamba_track_indices = torch.tensor(
                 mamba_track_indices_cpu,
                 dtype=torch.int64,
@@ -2662,7 +2668,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
             # In lazy mode, skip the swap — the second ping-pong slot is not
             # allocated yet; it will be allocated on demand at the track boundary
             # in mamba_lazy_prealloc_at_boundary during prepare_for_decode.
-            if not server_args.enable_mamba_extra_buffer_lazy():
+            if not mamba_extra_buffer_lazy_enabled():
                 req.mamba_next_track_idx = (
                     self.req_to_token_pool.get_mamba_ping_pong_other_idx(
                         req.mamba_next_track_idx
@@ -3003,7 +3009,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
 
     def prepare_for_decode(self):
         self.forward_mode = ForwardMode.DECODE
-        server_args = get_server_args()
         # Decode embeds the last output token via embed_tokens; clear the stale
         # prefill-time tensor so it doesn't leak into ForwardBatch.
         self.input_embeds = None
@@ -3057,7 +3062,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                 self.req_pool_indices_cpu,
             )
 
-        if server_args.enable_mamba_extra_buffer():
+        if mamba_extra_buffer_enabled():
             mamba_track_interval = get_exec().mamba.mamba_track_interval
 
             if len(self.reqs) == 0:
@@ -3065,7 +3070,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
                     (0,), dtype=torch.int64, device=self.device
                 )
             else:
-                if server_args.enable_mamba_extra_buffer_lazy():
+                if mamba_extra_buffer_lazy_enabled():
                     self.mamba_lazy_prealloc_at_boundary(mamba_track_interval)
                 set_mamba_track_indices_from_reqs(self)
 

python/sglang/srt/managers/scheduler_components/batch_result_processor.py

@@ -38,6 +38,7 @@ from sglang.srt.runtime_context import (
     get_memory,
     get_observability,
     get_server_args,
+    mamba_extra_buffer_lazy_enabled,
 )
 from sglang.srt.speculative.base_spec_worker import BaseSpecWorker
 from sglang.srt.state_capturer.indexer_topk import get_global_indexer_capturer
@@ -1073,7 +1074,7 @@ class SchedulerBatchResultProcessor:
                     prepare_release(req)
                 is_insert = (
                     req.mamba_lazy_is_insert
-                    if get_server_args().enable_mamba_extra_buffer_lazy()
+                    if mamba_extra_buffer_lazy_enabled()
                     else True
                 )
                 release_kv_cache(req, self.tree_cache, is_insert=is_insert)
@@ -1109,7 +1110,7 @@ class SchedulerBatchResultProcessor:
         if req.mamba_ping_pong_track_buffer is None:
             return
 
-        lazy = get_server_args().enable_mamba_extra_buffer_lazy()
+        lazy = mamba_extra_buffer_lazy_enabled()
         if known_boundary:
             self._mamba_assert_committed_len_lookahead(req)
             track_seqlen = req.kv_committed_len

python/sglang/srt/mem_cache/allocation_sizing.py

@@ -6,10 +6,7 @@ from sglang.srt.runtime_context import get_server_args
 from sglang.srt.server_args import ServerArgs
 
 
-def get_alloc_len_per_decode(server_args: Optional[ServerArgs] = None) -> int:
-    if server_args is None:
-        server_args = get_server_args()
-
+def get_alloc_len_per_decode(server_args: ServerArgs) -> int:
     if server_args.speculative_algorithm is None:
         return 1
 
@@ -39,7 +36,13 @@ def get_alloc_reserve_per_decode(server_args: Optional[ServerArgs] = None) -> in
 
     The 2x is a double-buffer that absorbs the kv_committed_len lag in overlap
     mode; see eagle_utils.eagle_prepare_for_decode.
+
+    Callers on a request path have no config in hand, so this is the module's
+    single "which config" decision point: everything below it takes the
+    instance explicitly.
     """
+    if server_args is None:
+        server_args = get_server_args()
     return 2 * get_alloc_len_per_decode(server_args)
 
 

python/sglang/srt/models/deepseek_v2.py

@@ -858,11 +858,7 @@ class DeepseekV2MoE(nn.Module):
             )
         ]
 
-    def _can_dual_stream_graph(
-        self, hidden_states: torch.Tensor, server_args=None
-    ) -> bool:
-        if server_args is None:
-            server_args = get_server_args()
+    def _can_dual_stream_graph(self, hidden_states: torch.Tensor) -> bool:
         return (
             _enable_pcg_dsv2_dual_stream
             and (is_in_tc_piecewise_cuda_graph() or is_in_breakable_cuda_graph())
@@ -875,7 +871,7 @@ class DeepseekV2MoE(nn.Module):
             and not self._enable_a2a_moe
             and not self._fuse_shared_experts_inside_sbo
             and not getattr(self, "is_hash", False)
-            and not server_args.enable_eplb
+            and not get_exec().moe.enable_eplb
         )
 
     def forward(
@@ -898,8 +894,7 @@ class DeepseekV2MoE(nn.Module):
             )
 
         if not self._enable_a2a_moe:
-            server_args = get_server_args()
-            if self._can_dual_stream_graph(hidden_states, server_args):
+            if self._can_dual_stream_graph(hidden_states):
                 fwd = get_forward()
                 return dsv2_flashinfer_moe_dual_stream_graph(
                     hidden_states,

python/sglang/srt/models/inkling.py

@@ -80,7 +80,7 @@ from sglang.srt.runtime_context import (
     get_model,
     get_parallel,
     get_schedule,
-    get_server_args,
+    mamba_extra_buffer_enabled,
 )
 from sglang.srt.utils import add_prefix, is_cuda, make_layers
 
@@ -1017,12 +1017,11 @@ class InklingForConditionalGeneration(nn.Module):
         self.config = config
         self.text_config = config.text_config
 
-        server_args = get_server_args()
         assert envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.get()
         if get_disagg().disaggregation_mode != "decode":
             assert not get_memory().disable_radix_cache
             assert not get_schedule().disable_hybrid_swa_memory
-            assert server_args.enable_mamba_extra_buffer()
+            assert mamba_extra_buffer_enabled()
 
         from types import SimpleNamespace
 

python/sglang/srt/models/inkling_common/sconv.py

@@ -17,7 +17,11 @@ from sglang.srt.models.inkling_common.kernels.sconv import (
     save_intermediate_conv_windows,
     update_sconv_cache,
 )
-from sglang.srt.runtime_context import get_exec, get_parallel, get_server_args
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_parallel,
+    mamba_extra_buffer_enabled,
+)
 from sglang.srt.utils import is_cuda, set_weight_attrs
 
 
@@ -267,10 +271,7 @@ class ShortConvolution(nn.Module):
             draft_token_num = hidden_states.shape[1]
 
         mamba_track_indices = getattr(forward_batch, "mamba_track_indices", None)
-        do_tracking = (
-            mamba_track_indices is not None
-            and get_server_args().enable_mamba_extra_buffer()
-        )
+        do_tracking = mamba_track_indices is not None and mamba_extra_buffer_enabled()
 
         crossed = track_step = None
         if do_tracking:

python/sglang/srt/runtime_context.py

@@ -1383,3 +1383,26 @@ def reset_context() -> None:
     _CONTEXT.resources = Resources()
     _CONTEXT.forward = ForwardFlags()
     set_global_dwdp_manager(None)
+
+
+def mamba_extra_buffer_enabled() -> bool:
+    """Whether the mamba radix cache keeps its extra state buffer.
+
+    A predicate over two published leaves (``memory.disable_radix_cache`` and
+    ``exec.mamba.mamba_radix_cache_strategy``), so it reads the bags rather
+    than the startup record — the ``ServerArgs`` member of the same name is the
+    pre-publish equivalent used inside the resolution pipeline.
+    """
+    return (
+        get_memory().disable_radix_cache is False
+        and get_exec().mamba.mamba_radix_cache_strategy
+        in ("extra_buffer", "extra_buffer_lazy")
+    )
+
+
+def mamba_extra_buffer_lazy_enabled() -> bool:
+    """The lazy variant of :func:`mamba_extra_buffer_enabled`."""
+    return (
+        get_memory().disable_radix_cache is False
+        and get_exec().mamba.mamba_radix_cache_strategy == "extra_buffer_lazy"
+    )

python/sglang/srt/speculative/spec_utils.py

@@ -46,7 +46,13 @@ from sglang.srt.mem_cache.allocation import (
 from sglang.srt.mem_cache.allocation import (
     assign_req_to_token_pool_func as assign_req_to_token_pool_func,
 )
-from sglang.srt.runtime_context import get_exec, get_server_args
+from sglang.srt.runtime_context import (
+    get_exec,
+    get_server_args,
+    get_spec,
+    mamba_extra_buffer_enabled,
+    mamba_extra_buffer_lazy_enabled,
+)
 from sglang.srt.utils import (
     is_cpu,
     is_cuda,
@@ -250,16 +256,14 @@ def record_stream_for_v2_verify(batch, verify_input, fwd_stream):
     record_stream_each(candidates, fwd_stream)
 
 
-def spec_need_hidden_states(server_args: Optional[ServerArgs] = None) -> bool:
-    if server_args is None:
-        server_args = get_server_args()
-
+def spec_need_hidden_states() -> bool:
     # STANDALONE drafts don't consume `spec_info.hidden_states` (vanilla LLM).
     # multi_layer_eagle, DFLASH, and DSPARK don't relay hidden_states through FutureMap.
     # TODO(lsyin): also skip when step == 1.
-    if server_args.speculative_algorithm in ("STANDALONE", "DFLASH", "DSPARK"):
+    spec = get_spec()
+    if spec.speculative_algorithm in ("STANDALONE", "DFLASH", "DSPARK"):
         return False
-    return not server_args.enable_multi_layer_eagle
+    return not spec.enable_multi_layer_eagle
 
 
 @torch.compile(dynamic=True, disable=_is_npu or _is_xpu)
@@ -761,11 +765,10 @@ def prepare_mamba_track_for_verify(batch: ScheduleBatch) -> None:
     Lazy: gather the positions planned by mamba_lazy_spec_prepare. Runs
     inside forward isolation, so it must not mutate req/pool state.
     """
-    server_args = get_server_args()
-    if not server_args.enable_mamba_extra_buffer():
+    if not mamba_extra_buffer_enabled():
         return
     track_positions = None
-    if server_args.enable_mamba_extra_buffer_lazy():
+    if mamba_extra_buffer_lazy_enabled():
         track_positions = batch.mamba_lazy_spec_track_positions_cpu
         assert track_positions is not None and len(track_positions) == len(
             batch.reqs
@@ -1028,7 +1031,7 @@ def spec_prepare_for_decode(batch: ScheduleBatch) -> None:
     prep on its draft input -- the dispatcher routes.
     """
     server_args = get_server_args()
-    if server_args.enable_mamba_extra_buffer_lazy():
+    if mamba_extra_buffer_lazy_enabled():
         # Scheduler phase (outside forward isolation).
         batch.mamba_lazy_spec_prepare(
             get_exec().mamba.mamba_track_interval,