[Config] Round 5.1: the published-side readers ask the bags, and a platform fact gets one address (#37086)

python/sglang/srt/model_executor/cuda_graph_config.py

@@ -26,6 +26,8 @@ import json
 from dataclasses import dataclass, field, replace
 from typing import Any, Dict, List, Optional
 
+from sglang.srt.runtime_context import get_exec
+
 
 class Phase:
     """The two phases of model forward."""
@@ -201,7 +203,6 @@ def check_cuda_graph_backend(phase: str, backend: str) -> bool:
     """True if cuda_graph_config[phase].backend == backend on the
     published config. Returns False if the config has not been published
     yet (e.g. unit tests, early startup)."""
-    from sglang.srt.runtime_context import get_exec
 
     try:
         cfg = get_exec().graph.cuda_graph_config

python/sglang/srt/model_executor/model_runner.py

@@ -444,7 +444,7 @@ class ModelRunner:
 
         # Update deep gemm configure
         if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
-            deep_gemm_wrapper.update_deep_gemm_config(gpu_id, server_args)
+            deep_gemm_wrapper.update_deep_gemm_config(gpu_id)
 
         # For hisparse (must be set before initialize() so CUDA graph capture can see it)
         self.hisparse_coordinator = None
@@ -542,7 +542,7 @@ class ModelRunner:
         self._rearm_eplb_after_elastic_scale()
 
     def init_msprobe(self):
-        self.msprobe_debugger = misc_utils.create_msprobe_debugger(self.server_args)
+        self.msprobe_debugger = misc_utils.create_msprobe_debugger()
 
     def init_weight_updater(self):
         self.weight_updater = WeightUpdater(
@@ -910,7 +910,7 @@ class ModelRunner:
         )
         from sglang.srt.mem_cache.sparsity import parse_hisparse_config
 
-        hisparse_cfg = parse_hisparse_config(self.server_args)
+        hisparse_cfg = parse_hisparse_config()
         hisparse_top_k = getattr(
             self.model_config.hf_text_config, "index_topk", hisparse_cfg.top_k
         )

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

@@ -26,6 +26,7 @@ from sglang.srt.model_loader.remote_instance_weight_loader_utils import (
 )
 from sglang.srt.platforms import current_platform
 from sglang.srt.runtime_context import (
+    get_context,
     get_exec,
     get_model,
     get_observability,
@@ -72,7 +73,6 @@ def maybe_downgrade_dtype_for_legacy_gpu(*, model_config: ModelConfig) -> None:
         logger.info(
             "Compute capability below sm80. Use float16 due to lack of bfloat16 support."
         )
-        from sglang.srt.runtime_context import get_context
 
         # Device-driven, so every runner in the process resolves the same way;
         # the per-runner truth is model_config.dtype, this is the record.

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

@@ -11,13 +11,13 @@ from sglang.srt.configs.model_config import (
 from sglang.srt.runtime_context import (
     attention_backends,
     get_context,
+    get_observability,
     get_schedule,
 )
 from sglang.srt.server_args import CHUNKED_PREFIX_CACHE_SUPPORTED_ATTENTION_BACKENDS
 
 if TYPE_CHECKING:
     from sglang.srt.configs.model_config import ModelConfig
-    from sglang.srt.server_args import ServerArgs
 
 logger = logging.getLogger(__name__)
 
@@ -48,8 +48,8 @@ def maybe_disable_chunked_prefix_cache(
         logger.info("Chunked prefix cache is turned on.")
 
 
-def create_msprobe_debugger(server_args: ServerArgs) -> Optional[Any]:
-    if server_args.msprobe_dump_config is None:
+def create_msprobe_debugger() -> Optional[Any]:
+    if get_observability().msprobe_dump_config is None:
         return None
 
     try:
@@ -62,7 +62,7 @@ def create_msprobe_debugger(server_args: ServerArgs) -> Optional[Any]:
         return None
 
     seed_all(mode=True)
-    return PrecisionDebugger(config_path=server_args.msprobe_dump_config)
+    return PrecisionDebugger(config_path=get_observability().msprobe_dump_config)
 
 
 def resolve_pp_proxy_topk_size(

python/sglang/srt/model_executor/pool_configurator.py

@@ -373,7 +373,7 @@ class DefaultPoolConfigurator(MemoryPoolConfigurator):
         if memory_config.enable_hisparse:
             from sglang.srt.mem_cache.sparsity import parse_hisparse_config
 
-            indexer_ratio = parse_hisparse_config(kvc.server_args).host_to_device_ratio
+            indexer_ratio = parse_hisparse_config().host_to_device_ratio
 
         from sglang.srt.mem_cache.kv_cache_configurator import (
             _should_elide_dsa_index_k,
@@ -794,9 +794,7 @@ class DSV4PoolConfigurator(MemoryPoolConfigurator):
         if get_memory().enable_hisparse:
             from sglang.srt.mem_cache.sparsity import parse_hisparse_config
 
-            self.c4_shrink_factor = parse_hisparse_config(
-                kvc.server_args
-            ).host_to_device_ratio
+            self.c4_shrink_factor = parse_hisparse_config().host_to_device_ratio
         else:
             self.c4_shrink_factor = 1
         assert self.c4_shrink_factor >= 1

python/sglang/srt/model_executor/runner_utils/capture_mode.py

@@ -27,6 +27,7 @@ import torch
 from sglang.srt.model_executor.runner_backend_utils.breakable_cuda_graph.context import (
     is_in_breakable_cuda_graph,
 )
+from sglang.srt.runtime_context import get_flags
 
 # Detect whether the current forward pass is in capture mode.
 is_capture_mode = False
@@ -84,7 +85,6 @@ def _set_capture_dsa_variant(variant: Optional[str]) -> None:
 @contextmanager
 def model_capture_mode():
     global is_capture_mode
-    from sglang.srt.runtime_context import get_flags
 
     # Disable dispose_tensor() during capture: freeing mid-capture records data_ptr()==0 into the graph.
     is_capture_mode = True