Report the whole server's world size in the scheduler's internal state (#35929)

python/sglang/srt/managers/scheduler.py

@@ -44,6 +44,7 @@ from sglang.srt.runtime_context import (
     get_observability,
     get_parallel,
     get_schedule,
+    get_server_args,
     get_serving,
     get_spec,
 )
@@ -296,7 +297,7 @@ from sglang.srt.plugins import load_plugins
 from sglang.srt.runtime_context import get_context, publish
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import TOP_K_ALL
-from sglang.srt.server_args import PortArgs, ServerArgs
+from sglang.srt.server_args import PortArgs, ServerArgs, compute_world_size
 from sglang.srt.session.session_controller import SessionController
 from sglang.srt.speculative.base_spec_worker import BaseSpecWorker
 from sglang.srt.speculative.dflash_utils import validate_dflash_request
@@ -4421,6 +4422,7 @@ class Scheduler(
         # Resolved config (pristine server_args + post-publish overrides) so a
         # readback reflects values changed via /set_internal_state, not startup.
         ret = get_context().resolved_server_args_dict()
+        ret["world_size"] = compute_world_size(get_server_args())
         ret["last_gen_throughput"] = self.metrics_reporter.last_gen_throughput
         draft_graph_memory_usage = (
             None if self.draft_worker is None else self.draft_worker.graph_memory_usage

python/sglang/srt/server_args.py

@@ -10450,6 +10450,15 @@ class ServerArgs:
         return self.expert_balancedness_report_mode in ("prometheus", "both")
 
 
+def compute_world_size(server_args: ServerArgs) -> int:
+    """Return the total GPU count across all data-parallel replicas."""
+    return (
+        (1 if server_args.enable_dp_attention else server_args.dp_size)
+        * server_args.tp_size
+        * server_args.pp_size
+    )
+
+
 def m3_fp8_attn_gemm_enabled(args) -> bool:
     """Whether MiniMax-M3 attention GEMMs run in fp8 (no opt-in flag; active
     whenever possible): fp8_e4m3 main + index KV caches, fp8-cast q, fp8