chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Tracing] Remove the deprecated tracing code from mini_lb (#19409)

Browse Source
This commit is contained in:
Feng Su
2026-03-07 11:19:23 +08:00
committed by GitHub
parent f88acf8780
commit 8a411a9a2a
2 changed files with 5 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/observability/req_time_stats.py
View File

@@ -184,17 +184,6 @@ class RequestStage:
metrics_is_observed=True,
)
# mini lb
MINI_LB_LAUNCH = RequestStageConfig(
"mini_lb_launch",
level=1,
)
WAIT_PD_FINISH = RequestStageConfig(
"wait_pd_finish",
level=2,
)
# other
ANONYMOUS = RequestStageConfig("")

102
sgl-model-gateway/bindings/python/src/sglang_router/mini_lb.py
View File

@@ -19,21 +19,6 @@ from fastapi import FastAPI, HTTPException
from fastapi.responses import ORJSONResponse, Response, StreamingResponse
from sglang_router.router_args import RouterArgs
try:
from sglang.srt.tracing.trace import (
process_tracing_init,
trace_get_remote_propagate_context,
trace_req_finish,
trace_req_start,
trace_set_thread_info,
trace_slice_end,
trace_slice_start,
)
trace_package_imported = True
except ImportError:
trace_package_imported = False
logger = logging.getLogger(__name__)
AIOHTTP_STREAM_READ_CHUNK_SIZE = (
@@ -62,14 +47,6 @@ class MiniLoadBalancer:
self.prefill_urls = [url[0] for url in router_args.prefill_urls]
self.prefill_bootstrap_ports = [url[1] for url in router_args.prefill_urls]
self.decode_urls = router_args.decode_urls
self.otlp_traces_endpoint = router_args.otlp_traces_endpoint
self.enable_trace = router_args.enable_trace
if self.enable_trace and not trace_package_imported:
logger.warning(
"Tracing is not supported in this environment. Please install sglang."
)
self.enable_trace = False
self.test_external_dp_routing = router_args.test_external_dp_routing
self.prefill_dp_size = None
self.decode_dp_size = None
@@ -95,9 +72,6 @@ class MiniLoadBalancer:
def start(self):
global lb
lb = self
if self.enable_trace:
process_tracing_init(self.otlp_traces_endpoint, "sglang")
trace_set_thread_info("Mini lb")
uvicorn.run(app, host=self.host, port=self.port)
async def _ensure_dp_sizes(self):
@@ -157,33 +131,12 @@ class MiniLoadBalancer:
total=self.timeout
) # Add timeout for request reliability
) as session:
headers = {}
bootstrap_room_list = []
if self.enable_trace:
bootstrap_room_list = (
modified_request["bootstrap_room"]
if isinstance(modified_request["bootstrap_room"], list)
else [modified_request["bootstrap_room"]]
)
trace_context = trace_get_remote_propagate_context(bootstrap_room_list)
headers = {"trace_context": trace_context}
tasks = [
session.post(
f"{prefill_server}/{endpoint}",
json=prefill_req,
headers=headers,
),
session.post(
f"{decode_server}/{endpoint}",
json=decode_req,
headers=headers,
),
session.post(f"{prefill_server}/{endpoint}", json=prefill_req),
session.post(f"{decode_server}/{endpoint}", json=decode_req),
]
for bootstrap_room in bootstrap_room_list:
trace_slice_end("mini_lb_launch", bootstrap_room, auto_next_anon=True)
# Wait for both responses to complete. Prefill should end first.
prefill_response, decode_response = await asyncio.gather(*tasks)
@@ -202,14 +155,6 @@ class MiniLoadBalancer:
else:
ret_json = await decode_response.json()
for bootstrap_room in bootstrap_room_list:
trace_slice_end(
"wait_PD_finish",
bootstrap_room,
thread_finish_flag=True,
)
trace_req_finish(bootstrap_room)
if expected_decode_dp_rank is not None:
actual = ret_json.get("meta_info", {}).get("dp_rank")
if actual != expected_decode_dp_rank:
@@ -241,36 +186,11 @@ class MiniLoadBalancer:
) # Add timeout for request reliability
) as session:
# Create the tasks for both prefill and decode requests
headers = {}
bootstrap_room_list = []
if self.enable_trace:
bootstrap_room_list = (
modified_request["bootstrap_room"]
if isinstance(modified_request["bootstrap_room"], list)
else [modified_request["bootstrap_room"]]
)
trace_context = trace_get_remote_propagate_context(
bootstrap_room_list
)
headers = {"trace_context": trace_context}
tasks = [
session.post(
f"{prefill_server}/{endpoint}",
json=modified_request,
headers=headers,
),
session.post(
f"{decode_server}/{endpoint}",
json=modified_request,
headers=headers,
),
session.post(f"{prefill_server}/{endpoint}", json=modified_request),
session.post(f"{decode_server}/{endpoint}", json=modified_request),
]
for bootstrap_room in bootstrap_room_list:
trace_slice_end(
"mini_lb_launch", bootstrap_room, auto_next_anon=True
)
# Wait for both responses to complete. Since this is streaming, they return immediately.
prefill_response, decode_response = await asyncio.gather(*tasks)
@@ -310,14 +230,6 @@ class MiniLoadBalancer:
):
yield chunk
for bootstrap_room in bootstrap_room_list:
trace_slice_end(
"wait_PD_finish",
bootstrap_room,
thread_finish_flag=True,
)
trace_req_finish(bootstrap_room)
return StreamingResponse(
stream_results(),
media_type="text/event-stream",
@@ -520,11 +432,7 @@ async def handle_completion_request(request_data: dict):
def _generate_bootstrap_room():
bootstrap_room = random.randint(0, 2**63 - 1)
if lb.enable_trace:
trace_req_start(bootstrap_room, bootstrap_room, role="router")
trace_slice_start("mini_lb_launch", bootstrap_room)
return bootstrap_room
return random.randint(0, 2**63 - 1)
# We may utilize `GenerateReqInput`'s logic later