config: the runtime readers take the published bags (#36254)

python/sglang/srt/model_executor/cpu_graph_runner.py

@@ -609,7 +609,7 @@ class CPUGraphRunner:
         self.enable_profile_cuda_graph = (
             model_runner.server_args.enable_profile_cuda_graph
         )
-        self.tp_size = model_runner.server_args.tp_size
+        self.tp_size = get_parallel().config.tp_size
         self.dp_size = get_parallel().config.dp_size
         self.pp_size = get_parallel().config.pp_size
 

python/sglang/srt/model_executor/mindspore_runner.py

@@ -14,6 +14,7 @@ from mindspore._c_expression import GroupOptions
 from mindspore.communication import create_group
 
 from sglang.srt.distributed.parallel_state import _groups
+from sglang.srt.runtime_context import get_serving
 
 logger = logging.getLogger(__name__)
 
@@ -109,7 +110,7 @@ def init_ms_distributed(world_size, rank, local_rank, server_args, port):
     if server_args.dist_init_addr:
         dist_init_method = f"tcp://{server_args.dist_init_addr}"
     else:
-        dist_init_method = f"tcp://{server_args.host}:{port}"
+        dist_init_method = f"tcp://{get_serving().host}:{port}"
     set_ms_parallel_env(rank, local_rank, world_size, dist_init_method)
 
     ms.set_context(infer_boost="on", jit_level="O0")

python/sglang/srt/model_executor/model_runner.py

@@ -1202,7 +1202,6 @@ class ModelRunner:
         # Pre-expand RoPE cache before CUDA Graph capture
         reserve_rope_cache_for_long_sequences(
             self.model,
-            self.server_args,
             self.model_config,
             logger,
         )

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

@@ -13,6 +13,7 @@ from sglang.srt.eplb.lplb_solver import (
 )
 from sglang.srt.layers.moe.hash_topk import HashTopK
 from sglang.srt.layers.moe.topk import TopK
+from sglang.srt.runtime_context import get_exec
 from sglang.srt.utils import get_bool_env_var, is_hip, log_info_on_rank0
 
 if TYPE_CHECKING:
@@ -54,7 +55,7 @@ def prepare_moe_topk(
         # Redundant experts therefore need to be included in the per-rank
         # expert count used for Waterfill's shared-expert slot remapping.
         num_physical_routed_experts = (
-            num_routed_experts + server_args.ep_num_redundant_experts
+            num_routed_experts + get_exec().moe.ep_num_redundant_experts
         )
         if isinstance(module, TopK):
             routed_scaling_factor = module.topk_config.routed_scaling_factor

python/sglang/srt/model_executor/runner/base_runner.py

@@ -216,7 +216,7 @@ class BaseRunner(ABC):
         self.model_runner = model_runner
         self.device = model_runner.device
         self.device_module = torch.get_device_module(self.device)
-        self.tp_size = model_runner.server_args.tp_size
+        self.tp_size = get_parallel().config.tp_size
         # elastic-EP scale-up rewrites dp_size on the published config
         self.dp_size = get_parallel().config.dp_size
         self.pp_size = get_parallel().config.pp_size

python/sglang/srt/model_executor/runner_backend/utils.py

@@ -64,7 +64,7 @@ def resolve_decode_backend(
     cfg = get_exec().graph.cuda_graph_config
     backend_name = cfg.decode.backend if cfg is not None else Backend.FULL
 
-    enable_memory_saver = model_runner.server_args.enable_memory_saver
+    enable_memory_saver = get_exec().features.enable_memory_saver
 
     if model_runner.device == "npu":
         from sglang.srt.hardware_backend.npu.graph_runner.npu_cudagraph_backend import (
@@ -115,13 +115,13 @@ def resolve_prefill_backend(
     if backend_name == Backend.BREAKABLE:
         return BreakableCudaGraphBackend(
             cuda_graph_runner,
-            enable_memory_saver=model_runner.server_args.enable_memory_saver,
+            enable_memory_saver=get_exec().features.enable_memory_saver,
             debug_eager=get_exec().graph.debug_cuda_graph,
         )
     if backend_name == Backend.FULL:
         return FullCudaGraphBackend(
             cuda_graph_runner,
-            enable_memory_saver=model_runner.server_args.enable_memory_saver,
+            enable_memory_saver=get_exec().features.enable_memory_saver,
         )
     # Default: tc_piecewise.
     return TcPiecewiseCudaGraphBackend(cuda_graph_runner)