Fix trace_modules gate disabling default trace contexts (#27173)

python/sglang/multimodal_gen/runtime/utils/trace_wrapper.py

@@ -32,32 +32,17 @@ def init_diffusion_tracing(server_args, thread_label: str):
     if not server_args.enable_trace:
         return
 
-    from types import SimpleNamespace
-
-    from sglang.srt import server_args as srt_server_args_module
     from sglang.srt.observability.trace import (
         process_tracing_init,
         trace_set_thread_info,
     )
-    from sglang.srt.server_args import set_global_server_args_for_scheduler
 
-    # srt owns TraceReqContext and filters spans through its global trace_modules
-    try:
-        srt_server_args = srt_server_args_module.get_global_server_args()
-    except ValueError:
-        srt_server_args = SimpleNamespace(trace_modules=DIFFUSION_TRACE_MODULE)
-        set_global_server_args_for_scheduler(srt_server_args)
-
-    trace_modules = [
-        module.strip()
-        for module in getattr(srt_server_args, "trace_modules", "").split(",")
-        if module.strip()
-    ]
-    if DIFFUSION_TRACE_MODULE not in trace_modules:
-        trace_modules.append(DIFFUSION_TRACE_MODULE)
-        srt_server_args.trace_modules = ",".join(trace_modules)
-
-    process_tracing_init(server_args.otlp_traces_endpoint, "sglang-diffusion")
+    # srt owns TraceReqContext and filters spans through its trace_modules list
+    process_tracing_init(
+        server_args.otlp_traces_endpoint,
+        "sglang-diffusion",
+        trace_modules=DIFFUSION_TRACE_MODULE,
+    )
     trace_set_thread_info(thread_label)
 
 

python/sglang/srt/entrypoints/engine.py

@@ -257,7 +257,11 @@ class Engine(EngineScoreMixin, EngineBase):
 
         # Enable tracing
         if server_args.enable_trace:
-            process_tracing_init(server_args.otlp_traces_endpoint, "sglang")
+            process_tracing_init(
+                server_args.otlp_traces_endpoint,
+                "sglang",
+                trace_modules=server_args.trace_modules,
+            )
             thread_label = "Tokenizer"
             if server_args.disaggregation_mode == "prefill":
                 thread_label = "Prefill Tokenizer"

python/sglang/srt/entrypoints/http_server.py

@@ -306,7 +306,11 @@ async def lifespan(fast_api_app: FastAPI):
 
     # Init tracing
     if server_args.enable_trace:
-        process_tracing_init(server_args.otlp_traces_endpoint, "sglang")
+        process_tracing_init(
+            server_args.otlp_traces_endpoint,
+            "sglang",
+            trace_modules=server_args.trace_modules,
+        )
         if server_args.disaggregation_mode == "prefill":
             thread_label = "Prefill" + thread_label
         elif server_args.disaggregation_mode == "decode":

python/sglang/srt/managers/data_parallel_controller.py

@@ -667,7 +667,11 @@ def run_data_parallel_controller_process(
 
     configure_logger(server_args)
     if server_args.enable_trace:
-        process_tracing_init(server_args.otlp_traces_endpoint, "sglang")
+        process_tracing_init(
+            server_args.otlp_traces_endpoint,
+            "sglang",
+            trace_modules=server_args.trace_modules,
+        )
         thread_label = "DP Controller"
         if server_args.disaggregation_mode == "prefill":
             thread_label = "Prefill DP Controller"

python/sglang/srt/managers/scheduler.py

@@ -3977,7 +3977,11 @@ def run_scheduler_process(
 
     # Set up tracing
     if server_args.enable_trace:
-        process_tracing_init(server_args.otlp_traces_endpoint, "sglang")
+        process_tracing_init(
+            server_args.otlp_traces_endpoint,
+            "sglang",
+            trace_modules=server_args.trace_modules,
+        )
         thread_label = "Scheduler"
         if server_args.disaggregation_mode == "prefill":
             thread_label = "Prefill Scheduler"

python/sglang/srt/observability/trace.py

@@ -22,14 +22,10 @@ import threading
 import time
 import uuid
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Dict, List, Mapping, Optional
+from typing import Any, Dict, List, Mapping, Optional
 
-from sglang.srt.server_args import get_global_server_args
 from sglang.srt.utils import get_int_env_var
 
-if TYPE_CHECKING:
-    from sglang.srt.server_args import ServerArgs
-
 logger = logging.getLogger(__name__)
 opentelemetry_imported = False
 opentelemetry_initialized = False
@@ -38,6 +34,9 @@ tracer: Optional[trace.Tracer] = None
 
 global_trace_level = get_int_env_var("SGLANG_TRACE_LEVEL", 3)
 
+# Modules allowed to emit spans (from --trace-modules); None means no filtering.
+global_trace_modules: Optional[List[str]] = None
+
 TRACE_HEADERS = ["traceparent", "tracestate"]
 
 try:
@@ -162,10 +161,19 @@ def _get_host_id() -> str:
 
 
 # Should be called by each tracked process.
-def process_tracing_init(otlp_endpoint, server_name):
+def process_tracing_init(
+    otlp_endpoint, server_name, trace_modules: Optional[str] = None
+):
     global opentelemetry_initialized
     global get_cur_time_ns
     global tracer
+    global global_trace_modules
+
+    if trace_modules is not None:
+        global_trace_modules = [
+            module.strip() for module in trace_modules.split(",") if module.strip()
+        ]
+
     if not opentelemetry_imported:
         opentelemetry_initialized = False
         raise RuntimeError(
@@ -263,8 +271,13 @@ class TraceReqContext:
         self.trace_level = global_trace_level
         self.tracing_enable: bool = opentelemetry_initialized and self.trace_level > 0
 
-        server_args: ServerArgs = get_global_server_args()
-        if module_name not in server_args.trace_modules.split(","):
+        # Filter by --trace-modules only for explicitly named modules; contexts
+        # created with the default empty module_name are always traced.
+        if (
+            module_name
+            and global_trace_modules is not None
+            and module_name not in global_trace_modules
+        ):
             self.tracing_enable = False
 
         if not self.tracing_enable: