diff --git a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
index bddcdffe5..f6905d137 100644
--- a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
+++ b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
@@ -26,10 +26,7 @@ from sglang.multimodal_gen.runtime.utils.logging_utils import (
     globally_suppress_loggers,
     init_logger,
 )
-from sglang.multimodal_gen.runtime.utils.perf_logger import (
-    PerformanceLogger,
-    RequestTimings,
-)
+from sglang.multimodal_gen.runtime.utils.perf_logger import PerformanceLogger
 
 logger = init_logger(__name__)
 
@@ -101,19 +98,18 @@ class GPUWorker:
                 torch.cuda.reset_peak_memory_stats()
 
             start_time = time.monotonic()
-            timings = RequestTimings(request_id=req.request_id)
-            req.timings = timings
 
             output_batch = self.pipeline.forward(req, self.server_args)
-            duration_ms = (time.monotonic() - start_time) * 1000
 
             if self.rank == 0:
                 peak_memory_bytes = torch.cuda.max_memory_allocated()
                 output_batch.peak_memory_mb = peak_memory_bytes / (1024**2)
 
             if output_batch.timings:
+                duration_ms = (time.monotonic() - start_time) * 1000
                 output_batch.timings.total_duration_ms = duration_ms
-                PerformanceLogger.log_request_summary(timings=output_batch.timings)
+                if req.perf_dump_path is not None:
+                    PerformanceLogger.log_request_summary(timings=output_batch.timings)
         except Exception as e:
             logger.error(
                 f"Error executing request {req.request_id}: {e}", exc_info=True
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py b/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
index ea4a62de8..ceabdd476 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
@@ -14,7 +14,7 @@ from __future__ import annotations
 import os
 import pprint
 from dataclasses import asdict, dataclass, field
-from typing import TYPE_CHECKING, Any, Optional
+from typing import Any, Optional
 
 import PIL.Image
 import torch
@@ -26,13 +26,9 @@ from sglang.multimodal_gen.configs.sample.teacache import (
 )
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+from sglang.multimodal_gen.runtime.utils.perf_logger import RequestTimings
 from sglang.multimodal_gen.utils import align_to
 
-if TYPE_CHECKING:
-
-    from sglang.multimodal_gen.runtime.utils.perf_logger import RequestTimings
-
-
 logger = init_logger(__name__)
 
 
@@ -229,6 +225,8 @@ class Req:
         if self.guidance_scale_2 is None:
             self.guidance_scale_2 = self.guidance_scale
 
+        self.timings = RequestTimings(request_id=self.request_id)
+
     def adjust_size(self, server_args: ServerArgs):
         pass
 
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/timestep_preparation.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/timestep_preparation.py
index ca0e65f61..fea00c766 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/timestep_preparation.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/timestep_preparation.py
@@ -127,7 +127,7 @@ class TimestepPreparationStage(PipelineStage):
 
         # Update batch with prepared timesteps
         batch.timesteps = timesteps
-        self.log_debug(f"timesteps: {timesteps}")
+        self.log_debug("timesteps: %s", timesteps)
         return batch
 
     def verify_input(self, batch: Req, server_args: ServerArgs) -> VerificationResult:
diff --git a/python/sglang/multimodal_gen/runtime/utils/perf_logger.py b/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
index d76233696..20a1671e1 100644
--- a/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
+++ b/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
@@ -138,20 +138,24 @@ class StageProfiler:
         self.simple_log = simple_log
         self.start_time = 0.0
 
+        self._metrics_enabled = StageProfiler.metrics_enabled()
+
+    @staticmethod
+    def metrics_enabled():
         # Check env var at runtime to ensure we pick up changes (e.g. from CLI args)
-        self.metrics_enabled = envs.SGLANG_DIFFUSION_STAGE_LOGGING
+        return envs.SGLANG_DIFFUSION_STAGE_LOGGING
 
     def __enter__(self):
         if self.simple_log:
             self.logger.info(f"[{self.stage_name}] started...")
 
-        if (self.metrics_enabled and self.timings) or self.simple_log:
+        if (self._metrics_enabled and self.timings) or self.simple_log:
             self.start_time = time.perf_counter()
 
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        if not ((self.metrics_enabled and self.timings) or self.simple_log):
+        if not ((self._metrics_enabled and self.timings) or self.simple_log):
             return False
 
         execution_time_s = time.perf_counter() - self.start_time
@@ -171,7 +175,7 @@ class StageProfiler:
                 f"[{self.stage_name}] finished in {execution_time_s:.4f} seconds",
             )
 
-        if self.metrics_enabled and self.timings:
+        if self._metrics_enabled and self.timings:
             if "denoising_step_" in self.stage_name:
                 index = int(self.stage_name[len("denoising_step_") :])
                 self.timings.record_steps(index, execution_time_s)
@@ -240,6 +244,8 @@ class PerformanceLogger:
     ):
         """logs the stage metrics and total duration for a completed request
         to the performance_log file.
+
+        Note that this accords to the time spent internally in server, postprocess is not included
         """
         formatted_stages = [
             {"name": name, "execution_time_ms": duration_ms}