config: three cache and pool readers take the bags (#36791)

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

python/sglang/srt/mem_cache/registry.py

@@ -18,7 +18,7 @@ from sglang.srt.environ import envs
 from sglang.srt.hardware_backend.mlx.runtime import use_mlx
 from sglang.srt.mem_cache.base_prefix_cache import BasePrefixCache
 from sglang.srt.mem_cache.cache_init_params import CacheInitParams
-from sglang.srt.runtime_context import get_disagg, get_memory
+from sglang.srt.runtime_context import get_disagg, get_memory, get_serving
 
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
@@ -198,7 +198,7 @@ def _create_unified_radix_cache(
 
 def create_tree_cache(ctx: TreeCacheBuildContext) -> BasePrefixCache:
     """Route to the matching factory to construct Radix Cache."""
-    name = ctx.server_args.radix_cache_backend
+    name = get_memory().radix_cache_backend
     if name:
         factory = get_radix_cache_factory(name)
         if factory is None:
@@ -215,7 +215,7 @@ def create_tree_cache(ctx: TreeCacheBuildContext) -> BasePrefixCache:
 
     if (
         get_memory().enable_hierarchical_cache
-        and ctx.server_args.hicache_host_memory_mode == "buffer_only"
+        and get_memory().hicache_host_memory_mode == "buffer_only"
     ):
         from sglang.srt.mem_cache.unified_radix_cache import UnifiedRadixCache
 
@@ -225,7 +225,7 @@ def create_tree_cache(ctx: TreeCacheBuildContext) -> BasePrefixCache:
                 f"the unified radix tree; this model selected {type(cache).__name__}."
             )
 
-    if ctx.server_args.enable_session_radix_cache and not getattr(
+    if get_memory().enable_session_radix_cache and not getattr(
         cache, "enable_session_radix_cache", False
     ):
         raise ValueError(
@@ -237,7 +237,7 @@ def create_tree_cache(ctx: TreeCacheBuildContext) -> BasePrefixCache:
     hicache_attached = cache.cache_controller is not None
     streaming_wrapped = False
     if (
-        ctx.server_args.enable_streaming_session
+        get_serving().enable_streaming_session
         and not cache.supports_streaming_session()
     ):
         from sglang.srt.session.streaming_session import StreamingSession

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -82,7 +82,7 @@ from sglang.srt.observability.metrics_collector import (
     StorageMetrics,
     StorageMetricsCollector,
 )
-from sglang.srt.runtime_context import get_memory
+from sglang.srt.runtime_context import get_memory, get_observability
 from sglang.srt.session.streaming_session import StreamingSession
 from sglang.srt.utils.common import ceil_align
 
@@ -367,7 +367,7 @@ class UnifiedRadixCache(BasePrefixCache):
 
     def init_hicache(self, server_args: ServerArgs, params: CacheInitParams) -> None:
         """Initialize HiCache infrastructure."""
-        self.host_memory_mode = server_args.hicache_host_memory_mode
+        self.host_memory_mode = get_memory().hicache_host_memory_mode
         if self.host_memory_mode == "buffer_only":
             # FULL and FULL+SWA only: Mamba has no state-handoff channel on
             # the admission-time load-back read path and is not layer-gated.
@@ -387,7 +387,7 @@ class UnifiedRadixCache(BasePrefixCache):
 
         self.load_cache_event = threading.Event()
         self.sidecar_pool_specs.clear()
-        self.extra_metric_labels = server_args.extra_metric_labels
+        self.extra_metric_labels = get_observability().extra_metric_labels
 
         # Parse storage config once, share with assembler and tree
         storage_backend = get_memory().hicache_storage_backend

python/sglang/srt/model_executor/pool_configurator.py

@@ -169,7 +169,7 @@ class DefaultPoolConfigurator(MemoryPoolConfigurator):
         self._zero_kv_max_tokens = (
             torch.iinfo(torch.int64).max
             if has_kv_on_another_pp_stage
-            else kvc.server_args.max_total_tokens or kvc.model_config.context_len
+            else get_schedule().max_total_tokens or kvc.model_config.context_len
         )
 
         # EAGLE/STANDALONE: scale cell_size to account for draft model KV cache.
@@ -792,7 +792,7 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
         self.disaggregation_decode_extra_slots = (
             get_disagg().disaggregation_decode_extra_slots or 0
         )
-        if kvc.server_args.enable_hisparse:
+        if get_memory().enable_hisparse:
             from sglang.srt.mem_cache.sparsity import parse_hisparse_config
 
             self.c4_shrink_factor = parse_hisparse_config(