[Feature] Sglang Tracing: Fine-Grained Tracking for Request Latency - Part 1 (#9962)

Signed-off-by: Feng Su <sufeng@linux.alibaba.com>
Signed-off-by: Huaixin Chang <changhuaixin@linux.alibaba.com>
Signed-off-by: Peng Wang <rocking@linux.alibaba.com>

python/sglang/srt/entrypoints/engine.py

@@ -33,6 +33,8 @@ import zmq
 import zmq.asyncio
 from PIL.Image import Image
 
+from sglang.srt.tracing.trace import process_tracing_init, trace_set_thread_info
+
 # Fix a bug of Python threading
 setattr(threading, "_register_atexit", lambda *args, **kwargs: None)
 
@@ -138,6 +140,12 @@ class Engine(EngineBase):
             context, zmq.DEALER, self.port_args.rpc_ipc_name, True
         )
 
+        if server_args.enable_trace:
+            process_tracing_init(server_args.oltp_traces_endpoint, "sglang")
+            if server_args.disaggregation_mode == "null":
+                thread_label = "Tokenizer"
+                trace_set_thread_info(thread_label)
+
     def generate(
         self,
         # The input prompt. It can be a single prompt or a batch of prompts.

python/sglang/srt/entrypoints/http_server.py

@@ -31,6 +31,8 @@ from typing import Any, AsyncIterator, Callable, Dict, List, Optional
 
 import setproctitle
 
+from sglang.srt.tracing.trace import process_tracing_init, trace_set_thread_info
+
 # Fix a bug of Python threading
 setattr(threading, "_register_atexit", lambda *args, **kwargs: None)
 
@@ -179,6 +181,13 @@ async def init_multi_tokenizer() -> ServerArgs:
             scheduler_info=scheduler_info,
         )
     )
+
+    if server_args.enable_trace:
+        process_tracing_init(server_args.oltp_traces_endpoint, "sglang")
+        if server_args.disaggregation_mode == "null":
+            thread_label = f"MultiTokenizer-{tokenizer_manager.worker_id}"
+            trace_set_thread_info(thread_label)
+
     return server_args
 
 
@@ -1203,6 +1212,12 @@ def launch_server(
             server_args=server_args,
         )
 
+        if server_args.enable_trace:
+            process_tracing_init(server_args.oltp_traces_endpoint, "sglang")
+            if server_args.disaggregation_mode == "null":
+                thread_label = "Tokenizer"
+                trace_set_thread_info(thread_label)
+
     set_global_state(
         _GlobalState(
             tokenizer_manager=tokenizer_manager,