minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] perf_logger: SYNC_STAGE_PROFILING must drain the GPU queue for stage records too (fixes 2-3x inflated DecodingStage readings) (#34124)

Browse Source 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Xiaoyu Zhang
2026-08-09 09:49:52 +08:00
committed by GitHub
co-authored by Claude Fable 5
parent cea16bf229
commit 33ed5d4413
2 changed files with 69 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/utils/perf_logger.py
+19 -12
View File
@@ -212,6 +212,23 @@ class StageProfiler:
def _should_record_as_step(self) -> bool:
return self.record_as_step or self.stage_name.startswith("denoising_step_")
def _maybe_sync_device(self):
"""Drain the device queue when SGLANG_DIFFUSION_SYNC_STAGE_PROFILING=1.
Called at BOTH the timing start and end, for stage records as well as
step records. Historically only step records synced, so a stage that
merely launches kernels (e.g. DenoisingStage's tail) leaked its queued
GPU work into whichever later stage blocked first — DecodingStage
readings came out 2-3x too high. The entry sync attributes queued work
to the stage that launched it; the exit sync includes this stage's own
queued work. Opt-in diagnostics only: the flag defaults off.
"""
if (
os.environ.get("SGLANG_DIFFUSION_SYNC_STAGE_PROFILING", "0") == "1"
and torch.get_device_module().is_available()
):
torch.get_device_module().synchronize()
def __enter__(self):
if self.log_stage_start_end:
msg = f"[{self.stage_name}] started..."
@@ -226,12 +243,7 @@ class StageProfiler:
self.logger.info(msg)
if (self.log_timing and self.metrics) or self.log_stage_start_end:
if (
os.environ.get("SGLANG_DIFFUSION_SYNC_STAGE_PROFILING", "0") == "1"
and self._should_record_as_step()
and torch.get_device_module().is_available()
):
torch.get_device_module().synchronize()
self._maybe_sync_device()
self.start_time = time.perf_counter()
return self
@@ -240,12 +252,7 @@ class StageProfiler:
if not ((self.log_timing and self.metrics) or self.log_stage_start_end):
return False
if (
os.environ.get("SGLANG_DIFFUSION_SYNC_STAGE_PROFILING", "0") == "1"
and self._should_record_as_step()
and torch.get_device_module().is_available()
):
torch.get_device_module().synchronize()
self._maybe_sync_device()
execution_time_s = time.perf_counter() - self.start_time
if exc_type: