[refactor] Move ServerArgs ownership into the runtime context (stack 2/15) (#30064)

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

python/sglang/multimodal_gen/test/unit/test_disagg_trace.py

@@ -33,6 +33,7 @@ from sglang.multimodal_gen.runtime.pipelines_core import Req
 from sglang.srt import server_args as srt_server_args_module
 from sglang.srt.observability import trace as srt_trace
 from sglang.srt.observability.trace import TraceNullContext, TraceReqContext
+from sglang.srt.runtime_context import reset_context
 from sglang.srt.server_args import set_global_server_args_for_scheduler
 
 try:
@@ -62,12 +63,18 @@ def _enable_minimal_otel() -> None:
 
 @contextmanager
 def _srt_trace_server_args():
-    prev_server_args = srt_server_args_module._global_server_args
+    try:
+        prev_server_args = srt_server_args_module.get_global_server_args()
+    except ValueError:  # nothing published yet
+        prev_server_args = None
     set_global_server_args_for_scheduler(SimpleNamespace(trace_modules="request"))
     try:
         yield
     finally:
-        srt_server_args_module._global_server_args = prev_server_args
+        if prev_server_args is None:
+            reset_context()
+        else:
+            set_global_server_args_for_scheduler(prev_server_args)
 
 
 def _traceparent_from(ctx) -> str | None:

python/sglang/srt/runtime_context.py

@@ -21,10 +21,12 @@ wrapper, not a cache. It gives call-sites one import and one naming scheme in
 place of a dozen free functions, plus a test-only ``override()`` hook to force a
 topology without monkeypatching the underlying getters.
 
-``get_server_args()`` returns the process-wide ``ServerArgs`` (the config tier).
-It is a read-through to ``server_args.get_global_server_args()`` — same object,
-same pre-publish error — so new code can adopt the context accessor while the
-legacy getter remains canonical.
+``get_server_args()`` returns the process-wide ``ServerArgs`` (the config
+tier). The context owns the storage: publishing goes through
+``RuntimeContext.set_server_args`` (the legacy
+``set_global_server_args_for_scheduler`` / ``get_global_server_args`` in
+``server_args.py`` are thin shims over this slot), and the object is returned
+by reference — the same live instance everywhere, never a copy.
 """
 
 from __future__ import annotations
@@ -50,12 +52,6 @@ def _dp():
     return dp_attention
 
 
-def _sa():
-    from sglang.srt import server_args
-
-    return server_args
-
-
 _PARALLEL_FIELDS = frozenset(
     {
         "world_size",
@@ -223,15 +219,29 @@ class RuntimeContext:
     """Container for the structured runtime accessors; exposes ``parallel`` and
     ``server_args``."""
 
-    __slots__ = ("parallel",)
+    __slots__ = ("parallel", "_server_args")
 
     def __init__(self, parallel: ParallelContext):
         self.parallel = parallel
+        self._server_args: ServerArgs | None = None
 
     @property
     def server_args(self) -> ServerArgs:
-        """The process-wide ``ServerArgs``, read through the global getter."""
-        return _sa().get_global_server_args()
+        """The process-wide ``ServerArgs`` (context-owned slot)."""
+        server_args = self._server_args
+        if server_args is None:
+            # Verbatim legacy message: tests and user scripts may match on it.
+            raise ValueError("Global server args is not set yet!")
+        return server_args
+
+    def set_server_args(self, server_args: ServerArgs) -> None:
+        """Publish the process-wide ``ServerArgs`` into the context-owned slot.
+
+        Overwrite-allowed: a re-publish replaces the slot (test kits re-publish
+        per test; production ordering discipline lives at the call-sites, e.g.
+        the draft-worker guard in ``ModelRunner.__init__``).
+        """
+        self._server_args = server_args
 
 
 _PARALLEL = ParallelContext()
@@ -248,3 +258,11 @@ def get_parallel() -> ParallelContext:
 
 def get_server_args() -> ServerArgs:
     return _CONTEXT.server_args
+
+
+def reset_context() -> None:
+    """Clear the context-owned store (unit-test teardown).
+
+    Wrapper subsystems (``parallel``) hold no state and are unaffected.
+    """
+    _CONTEXT._server_args = None

python/sglang/srt/server_args.py

@@ -7594,23 +7594,23 @@ class ServerArgs:
         }
 
 
-# NOTE: This is a global variable to hold the server args for scheduler.
-_global_server_args: Optional[ServerArgs] = None
-
-
+# NOTE: The process-wide ServerArgs is owned by the runtime context
+# (sglang.srt.runtime_context). The two functions below are thin shims kept for
+# the existing call-sites; they publish/read the same live object by reference.
+# Imports are in-function so the two modules stay cycle-free at import time.
 def set_global_server_args_for_scheduler(server_args: ServerArgs):
-    global _global_server_args
-    _global_server_args = server_args
+    from sglang.srt.runtime_context import get_context
+
+    get_context().set_server_args(server_args)
 
 
 set_global_server_args_for_tokenizer = set_global_server_args_for_scheduler
 
 
 def get_global_server_args() -> ServerArgs:
-    if _global_server_args is None:
-        raise ValueError("Global server args is not set yet!")
+    from sglang.srt.runtime_context import get_context
 
-    return _global_server_args
+    return get_context().server_args
 
 
 def prepare_server_args(argv: List[str]) -> ServerArgs: