diff --git a/python/sglang/multimodal_gen/runtime/entrypoints/cli/generate.py b/python/sglang/multimodal_gen/runtime/entrypoints/cli/generate.py
index f977afa47..8db2de9fa 100644
--- a/python/sglang/multimodal_gen/runtime/entrypoints/cli/generate.py
+++ b/python/sglang/multimodal_gen/runtime/entrypoints/cli/generate.py
@@ -18,9 +18,11 @@ from sglang.multimodal_gen.runtime.entrypoints.cli.cli_types import CLISubcomman
 from sglang.multimodal_gen.runtime.entrypoints.cli.utils import (
     RaiseNotImplementedAction,
 )
+from sglang.multimodal_gen.runtime.entrypoints.utils import GenerationResult
 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 (
+    MemorySnapshot,
     PerformanceLogger,
     RequestMetrics,
 )
@@ -58,7 +60,12 @@ def add_multimodal_gen_generate_args(parser: argparse.ArgumentParser):
     return parser
 
 
-def maybe_dump_performance(args: argparse.Namespace, server_args, prompt: str, results):
+def maybe_dump_performance(
+    args: argparse.Namespace,
+    server_args,
+    prompt: str,
+    results: GenerationResult | list[GenerationResult] | None,
+):
     """dump performance if necessary"""
     if not (args.perf_dump_path and results):
         return
@@ -68,20 +75,29 @@ def maybe_dump_performance(args: argparse.Namespace, server_args, prompt: str, r
     else:
         result = results
 
-    timings_dict = getattr(result, "timings", None) or (
-        result.get("timings") if isinstance(result, dict) else None
-    )
-    if not (args.perf_dump_path and timings_dict):
+    metrics_dict = result.metrics
+    if not (args.perf_dump_path and metrics_dict):
         return
 
-    timings = RequestMetrics(request_id=timings_dict.get("request_id"))
-    timings.stages = timings_dict.get("stages", {})
-    timings.steps = timings_dict.get("steps", [])
-    timings.total_duration_ms = timings_dict.get("total_duration_ms", 0)
+    metrics = RequestMetrics(request_id=metrics_dict.get("request_id"))
+    metrics.stages = metrics_dict.get("stages", {})
+    metrics.steps = metrics_dict.get("steps", [])
+    metrics.total_duration_ms = metrics_dict.get("total_duration_ms", 0)
+
+    # restore memory snapshots from serialized dict
+    memory_snapshots_dict = metrics_dict.get("memory_snapshots", {})
+    for checkpoint_name, snapshot_dict in memory_snapshots_dict.items():
+        snapshot = MemorySnapshot(
+            allocated_mb=snapshot_dict.get("allocated_mb", 0.0),
+            reserved_mb=snapshot_dict.get("reserved_mb", 0.0),
+            peak_allocated_mb=snapshot_dict.get("peak_allocated_mb", 0.0),
+            peak_reserved_mb=snapshot_dict.get("peak_reserved_mb", 0.0),
+        )
+        metrics.memory_snapshots[checkpoint_name] = snapshot
 
     PerformanceLogger.dump_benchmark_report(
         file_path=args.perf_dump_path,
-        timings=timings,
+        metrics=metrics,
         meta={
             "prompt": prompt,
             "model": server_args.model_path,
diff --git a/python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py b/python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py
index 3d6668c6a..7e03c3bf4 100644
--- a/python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py
+++ b/python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py
@@ -206,9 +206,9 @@ class DiffGenerator:
                         size=(req.height, req.width, req.num_frames),
                         generation_time=timer.duration,
                         peak_memory_mb=output_batch.peak_memory_mb,
-                        timings=(
-                            output_batch.timings.to_dict()
-                            if output_batch.timings
+                        metrics=(
+                            output_batch.metrics.to_dict()
+                            if output_batch.metrics
                             else {}
                         ),
                         trajectory_latents=output_batch.trajectory_latents,
@@ -297,7 +297,7 @@ class DiffGenerator:
         if not results:
             return
         if self.server_args.warmup:
-            total_duration_ms = results[0].timings.get("total_duration_ms", 0)
+            total_duration_ms = results[0].metrics.get("total_duration_ms", 0)
             logger.info(
                 f"Warmed-up request processed in {GREEN}%.2f{RESET} seconds (with warmup excluded)",
                 total_duration_ms / 1000.0,
diff --git a/python/sglang/multimodal_gen/runtime/entrypoints/openai/utils.py b/python/sglang/multimodal_gen/runtime/entrypoints/openai/utils.py
index b1aaab6fa..0f8ea4397 100644
--- a/python/sglang/multimodal_gen/runtime/entrypoints/openai/utils.py
+++ b/python/sglang/multimodal_gen/runtime/entrypoints/openai/utils.py
@@ -304,8 +304,8 @@ def add_common_data_to_response(
     if result.peak_memory_mb and result.peak_memory_mb > 0:
         response["peak_memory_mb"] = result.peak_memory_mb
 
-    if result.timings and result.timings.total_duration_s > 0:
-        response["inference_time_s"] = result.timings.total_duration_s
+    if result.metrics and result.metrics.total_duration_s > 0:
+        response["inference_time_s"] = result.metrics.total_duration_s
 
     response["id"] = request_id
 
diff --git a/python/sglang/multimodal_gen/runtime/entrypoints/utils.py b/python/sglang/multimodal_gen/runtime/entrypoints/utils.py
index 3c5e5b164..f0c980e44 100644
--- a/python/sglang/multimodal_gen/runtime/entrypoints/utils.py
+++ b/python/sglang/multimodal_gen/runtime/entrypoints/utils.py
@@ -105,7 +105,7 @@ class GenerationResult:
     size: tuple | None = None  # (height, width, num_frames)
     generation_time: float = 0.0
     peak_memory_mb: float = 0.0
-    timings: dict = field(default_factory=dict)
+    metrics: dict = field(default_factory=dict)
     trajectory_latents: Any = None
     trajectory_timesteps: Any = None
     trajectory_decoded: Any = None
diff --git a/python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py b/python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py
index 25efaa515..2f96619d3 100644
--- a/python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py
+++ b/python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py
@@ -19,7 +19,6 @@ from sglang.multimodal_gen.runtime.distributed import get_local_torch_device
 from sglang.multimodal_gen.runtime.loader.utils import (
     _normalize_component_type,
     component_name_to_loader_cls,
-    get_memory_usage_of_component,
 )
 from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
@@ -125,11 +124,9 @@ class ComponentLoader(ABC):
             if isinstance(component, nn.Module):
                 component = component.eval()
             current_gpu_mem = current_platform.get_available_gpu_memory()
-            consumed = get_memory_usage_of_component(component)
-            if consumed is None or consumed == 0.0:
-                consumed = gpu_mem_before_loading - current_gpu_mem
+            consumed = gpu_mem_before_loading - current_gpu_mem
             logger.info(
-                f"Loaded %s: %s ({source} version). model size: %.2f GB, avail mem: %.2f GB",
+                f"Loaded %s: %s ({source} version). consumed: %.2f GB, avail mem: %.2f GB",
                 component_name,
                 component.__class__.__name__,
                 consumed,
diff --git a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
index 9cbf0c211..229de9847 100644
--- a/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
+++ b/python/sglang/multimodal_gen/runtime/managers/gpu_worker.py
@@ -158,8 +158,8 @@ class GPUWorker:
 
     def do_mem_analysis(self, output_batch: OutputBatch):
         final_snapshot = capture_memory_snapshot()
-        if output_batch.timings:
-            output_batch.timings.record_memory_snapshot("mem_analysis", final_snapshot)
+        if output_batch.metrics:
+            output_batch.metrics.record_memory_snapshot("mem_analysis", final_snapshot)
 
         # for details on max_memory_reserved: https://docs.pytorch.org/docs/stable/generated/torch.cuda.memory.max_memory_reserved.html
         peak_reserved_bytes = torch.get_device_module().max_memory_reserved()
@@ -219,9 +219,9 @@ class GPUWorker:
             start_time = time.monotonic()
 
             # capture memory baseline before forward
-            if self.rank == 0 and req.timings:
+            if self.rank == 0 and req.metrics:
                 baseline_snapshot = capture_memory_snapshot()
-                req.timings.record_memory_snapshot("before_forward", baseline_snapshot)
+                req.metrics.record_memory_snapshot("before_forward", baseline_snapshot)
 
             req.log(server_args=self.server_args)
             result = self.pipeline.forward(req, self.server_args)
@@ -231,7 +231,7 @@ class GPUWorker:
                     output=result.output,
                     audio=getattr(result, "audio", None),
                     audio_sample_rate=getattr(result, "audio_sample_rate", None),
-                    timings=result.timings,
+                    metrics=result.metrics,
                     trajectory_timesteps=getattr(result, "trajectory_timesteps", None),
                     trajectory_latents=getattr(result, "trajectory_latents", None),
                     noise_pred=getattr(result, "noise_pred", None),
@@ -241,9 +241,9 @@ class GPUWorker:
                 output_batch = result
 
             # capture memory after forward (peak)
-            if self.rank == 0 and output_batch.timings:
+            if self.rank == 0 and output_batch.metrics:
                 peak_snapshot = capture_memory_snapshot()
-                output_batch.timings.record_memory_snapshot(
+                output_batch.metrics.record_memory_snapshot(
                     "after_forward", peak_snapshot
                 )
 
@@ -251,7 +251,7 @@ class GPUWorker:
                 self.do_mem_analysis(output_batch)
 
             duration_ms = (time.monotonic() - start_time) * 1000
-            output_batch.timings.total_duration_ms = duration_ms
+            output_batch.metrics.total_duration_ms = duration_ms
 
             # Save output to file and return file path only if requested. Avoid the serialization
             # and deserialization overhead between scheduler_client and gpu_worker.
@@ -273,11 +273,13 @@ class GPUWorker:
             if req.perf_dump_path is not None or envs.SGLANG_DIFFUSION_STAGE_LOGGING:
                 # Avoid logging warmup perf records that share the same request_id.
                 if not req.is_warmup:
-                    PerformanceLogger.log_request_summary(timings=output_batch.timings)
+                    PerformanceLogger.log_request_summary(metrics=output_batch.metrics)
         except Exception as e:
             logger.error(
                 f"Error executing request {req.request_id}: {e}", exc_info=True
             )
+            if isinstance(e, _oom_exceptions()):
+                logger.warning(OOM_MSG)
             if output_batch is None:
                 output_batch = OutputBatch()
             output_batch.error = f"Error executing request {req.request_id}: {e}"
@@ -427,8 +429,8 @@ OOM detected. Possible solutions:
   - If the OOM occurs during loading:
     1. Enable CPU offload for memory-intensive components, or use `--dit-layerwise-offload` for DiT
   - If the OOM occurs during runtime:
-    1. Reduce the number of output tokens by lowering resolution or decreasing `--num-frames`
-    2. Enable SP and/or TP
+    1. Enable SP and/or TP (in a multi-GPU setup)
+    2. Reduce the number of output tokens by lowering resolution or decreasing `--num-frames`
     3. Opt for a sparse-attention backend
     4. Enable FSDP by `--use-fsdp-inference` (in a multi-GPU setup)
     5. Enable quantization (e.g. nunchaku)
@@ -436,6 +438,14 @@ OOM detected. Possible solutions:
 """
 
 
+def _oom_exceptions():
+    # torch.OutOfMemoryError exists only in some PyTorch builds
+    types = [torch.cuda.OutOfMemoryError]
+    if hasattr(torch, "OutOfMemoryError"):
+        types.append(torch.OutOfMemoryError)
+    return tuple(types)
+
+
 def run_scheduler_process(
     local_rank: int,
     rank: int,
@@ -483,7 +493,7 @@ def run_scheduler_process(
             }
         )
         scheduler.event_loop()
-    except torch.OutOfMemoryError as _e:
+    except _oom_exceptions() as _e:
         logger.warning(OOM_MSG)
         raise
     finally:
diff --git a/python/sglang/multimodal_gen/runtime/managers/scheduler.py b/python/sglang/multimodal_gen/runtime/managers/scheduler.py
index e1fef1df2..add363d4b 100644
--- a/python/sglang/multimodal_gen/runtime/managers/scheduler.py
+++ b/python/sglang/multimodal_gen/runtime/managers/scheduler.py
@@ -395,12 +395,12 @@ class Scheduler:
                         if self._warmup_total > 0:
                             logger.info(
                                 f"Warmup req ({self._warmup_processed}/{self._warmup_total}) processed in {GREEN}%.2f{RESET} seconds",
-                                output_batch.timings.total_duration_s,
+                                output_batch.metrics.total_duration_s,
                             )
                         else:
                             logger.info(
                                 f"Warmup req processed in {GREEN}%.2f{RESET} seconds",
-                                output_batch.timings.total_duration_s,
+                                output_batch.metrics.total_duration_s,
                             )
                     else:
                         if self._warmup_total > 0:
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py b/python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py
index 82e01de79..62d69b0ce 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/composed_pipeline_base.py
@@ -22,6 +22,7 @@ from sglang.multimodal_gen.runtime.pipelines_core.executors.pipeline_executor im
 )
 from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import OutputBatch, Req
 from sglang.multimodal_gen.runtime.pipelines_core.stages import PipelineStage
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import (
     maybe_download_model,
@@ -311,7 +312,11 @@ class ComposedPipelineBase(ABC):
                     f"Required module: {module_name} was not found in loaded modules: {list(loaded_components.keys())}"
                 )
 
-        logger.debug("Memory usage of loaded modules: %s", self.memory_usages)
+        logger.debug(
+            "Memory usage of loaded modules (GiB): %s. Available memory: %s",
+            self.memory_usages,
+            round(current_platform.get_available_gpu_memory(), 2),
+        )
 
         return loaded_components
 
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/executors/pipeline_executor.py b/python/sglang/multimodal_gen/runtime/pipelines_core/executors/pipeline_executor.py
index a0f19834e..9775bfe05 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/executors/pipeline_executor.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/executors/pipeline_executor.py
@@ -31,7 +31,7 @@ class Timer(StageProfiler):
 
     def __init__(self, name="Stage"):
         super().__init__(
-            stage_name=name, timings=None, log_stage_start_end=True, logger=logger
+            stage_name=name, logger=logger, metrics=None, log_stage_start_end=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 34c41a6fa..592151637 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/schedule_batch.py
@@ -152,7 +152,7 @@ class Req:
     VSA_sparsity: float = 0.0
 
     # stage logging
-    timings: Optional["RequestMetrics"] = None
+    metrics: Optional["RequestMetrics"] = None
 
     # results
     output: torch.Tensor | None = None
@@ -267,7 +267,7 @@ class Req:
         if self.guidance_scale_2 is None:
             self.guidance_scale_2 = self.guidance_scale
 
-        self.timings = RequestMetrics(request_id=self.request_id)
+        self.metrics = RequestMetrics(request_id=self.request_id)
 
         if self.is_warmup:
             self.set_as_warmup()
@@ -330,7 +330,7 @@ class OutputBatch:
     output_file_paths: list[str] | None = None
 
     # logged metrics info, directly from Req.timings
-    timings: Optional["RequestMetrics"] = None
+    metrics: Optional["RequestMetrics"] = None
 
     # For ComfyUI integration: noise prediction from denoising stage
     noise_pred: torch.Tensor | None = None
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py
index a0eb92bae..e55243294 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/base.py
@@ -195,10 +195,10 @@ class PipelineStage(ABC):
         with StageProfiler(
             stage_name,
             logger=logger,
-            timings=batch.timings,
-            perf_dump_path_provided=batch.perf_dump_path is not None,
+            metrics=batch.metrics,
             log_stage_start_end=not batch.is_warmup
             and not (self.server_args and self.server_args.comfyui_mode),
+            perf_dump_path_provided=batch.perf_dump_path is not None,
         ):
             result = self.forward(batch, server_args)
 
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py
index ed676d79b..980d15210 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding.py
@@ -232,7 +232,7 @@ class DecodingStage(PipelineStage):
             trajectory_timesteps=batch.trajectory_timesteps,
             trajectory_latents=batch.trajectory_latents,
             trajectory_decoded=trajectory_decoded,
-            timings=batch.timings,
+            metrics=batch.metrics,
         )
 
         self.offload_model()
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding_av.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding_av.py
index 744688ca5..f330367aa 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding_av.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/decoding_av.py
@@ -69,7 +69,7 @@ class LTX2AVDecodingStage(DecodingStage):
             trajectory_timesteps=batch.trajectory_timesteps,
             trajectory_latents=batch.trajectory_latents,
             trajectory_decoded=None,
-            timings=batch.timings,
+            metrics=batch.metrics,
         )
 
         # 2. Decode Audio
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
index 83324b4cd..844999554 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
@@ -1017,7 +1017,7 @@ class DenoisingStage(PipelineStage):
                     with StageProfiler(
                         f"denoising_step_{i}",
                         logger=logger,
-                        timings=batch.timings,
+                        metrics=batch.metrics,
                         perf_dump_path_provided=batch.perf_dump_path is not None,
                     ):
                         t_int = int(t_host.item())
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_av.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_av.py
index c18e6a4d4..3c0921c6a 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_av.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_av.py
@@ -346,7 +346,7 @@ class LTX2AVDenoisingStage(DenoisingStage):
                     with StageProfiler(
                         f"denoising_step_{i}",
                         logger=logger,
-                        timings=batch.timings,
+                        metrics=batch.metrics,
                         perf_dump_path_provided=batch.perf_dump_path is not None,
                     ):
                         t_int = int(t_host.item())
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_dmd.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_dmd.py
index a47477388..504fc429e 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_dmd.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising_dmd.py
@@ -102,7 +102,7 @@ class DmdDenoisingStage(DenoisingStage):
                 with StageProfiler(
                     f"denoising_step_{i}",
                     logger=logger,
-                    timings=batch.timings,
+                    metrics=batch.metrics,
                     perf_dump_path_provided=batch.perf_dump_path is not None,
                 ):
                     t_int = int(t.item())
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/mova.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/mova.py
index da3e8bc3a..1329aa925 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/mova.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/mova.py
@@ -399,7 +399,7 @@ class MOVADenoisingStage(PipelineStage):
             getattr(batch, "extra_step_kwargs", None) or {},
         )
 
-        timings = getattr(batch, "timings", None)
+        metrics = getattr(batch, "metrics", None)
         perf_dump_path_provided = getattr(batch, "perf_dump_path", None) is not None
 
         with self.progress_bar(total=total_steps) as progress_bar:
@@ -407,7 +407,7 @@ class MOVADenoisingStage(PipelineStage):
                 with StageProfiler(
                     f"denoising_step_{idx_step}",
                     logger=logger,
-                    timings=timings,
+                    metrics=metrics,
                     perf_dump_path_provided=perf_dump_path_provided,
                 ):
                     pair_t = paired_timesteps[idx_step]
@@ -908,6 +908,6 @@ class MOVADecodingStage(PipelineStage):
             output=video,
             audio=audio,
             audio_sample_rate=getattr(self.audio_vae, "sample_rate", None),
-            timings=batch.timings,
+            metrics=batch.metrics,
         )
         return output_batch
diff --git a/python/sglang/multimodal_gen/runtime/utils/perf_logger.py b/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
index badfe1420..3f1c0429b 100644
--- a/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
+++ b/python/sglang/multimodal_gen/runtime/utils/perf_logger.py
@@ -1,6 +1,7 @@
 # Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
 import dataclasses
 import json
+import logging
 import os
 import subprocess
 import sys
@@ -15,7 +16,10 @@ from dateutil.tz import UTC
 
 import sglang
 import sglang.multimodal_gen.envs as envs
+from sglang.multimodal_gen.runtime.platforms import current_platform
 from sglang.multimodal_gen.runtime.utils.logging_utils import (
+    CYAN,
+    RESET,
     _SGLDiffusionLogger,
     get_is_main_process,
     init_logger,
@@ -184,13 +188,13 @@ class StageProfiler:
         self,
         stage_name: str,
         logger: _SGLDiffusionLogger,
-        timings: Optional["RequestMetrics"],
+        metrics: Optional["RequestMetrics"],
         log_stage_start_end: bool = False,
         perf_dump_path_provided: bool = False,
         capture_memory: bool = False,
     ):
         self.stage_name = stage_name
-        self.timings = timings
+        self.metrics = metrics
         self.logger = logger
         self.start_time = 0.0
         self.log_timing = perf_dump_path_provided or envs.SGLANG_DIFFUSION_STAGE_LOGGING
@@ -199,9 +203,12 @@ class StageProfiler:
 
     def __enter__(self):
         if self.log_stage_start_end:
-            self.logger.info(f"[{self.stage_name}] started...")
+            msg = f"[{self.stage_name}] started..."
+            if self.logger.isEnabledFor(logging.DEBUG):
+                msg += f" ({round(current_platform.get_available_gpu_memory(), 2)} GB left)"
+            self.logger.info(msg)
 
-        if (self.log_timing and self.timings) or self.log_stage_start_end:
+        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.stage_name.startswith("denoising_step_")
@@ -213,7 +220,7 @@ class StageProfiler:
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
-        if not ((self.log_timing and self.timings) or self.log_stage_start_end):
+        if not ((self.log_timing and self.metrics) or self.log_stage_start_end):
             return False
 
         if (
@@ -239,17 +246,17 @@ class StageProfiler:
                 f"[{self.stage_name}] finished in {execution_time_s:.4f} seconds",
             )
 
-        if self.log_timing and self.timings:
+        if self.log_timing and self.metrics:
             if "denoising_step_" in self.stage_name:
                 index = int(self.stage_name[len("denoising_step_") :])
-                self.timings.record_steps(index, execution_time_s)
+                self.metrics.record_steps(index, execution_time_s)
             else:
-                self.timings.record_stage(self.stage_name, execution_time_s)
+                self.metrics.record_stage(self.stage_name, execution_time_s)
 
             # capture memory snapshot after stage if requested
             if self.capture_memory and torch.cuda.is_available():
                 snapshot = capture_memory_snapshot()
-                self.timings.record_memory_snapshot(
+                self.metrics.record_memory_snapshot(
                     f"after_{self.stage_name}", snapshot
                 )
 
@@ -269,7 +276,7 @@ class PerformanceLogger:
     def dump_benchmark_report(
         cls,
         file_path: str,
-        timings: "RequestMetrics",
+        metrics: "RequestMetrics",
         meta: Optional[Dict[str, Any]] = None,
         tag: str = "benchmark_dump",
     ):
@@ -279,25 +286,25 @@ class PerformanceLogger:
         """
         formatted_steps = [
             {"name": name, "duration_ms": duration_ms}
-            for name, duration_ms in timings.stages.items()
+            for name, duration_ms in metrics.stages.items()
         ]
 
         denoise_steps_ms = [
             {"step": idx, "duration_ms": duration_ms}
-            for idx, duration_ms in enumerate(timings.steps)
+            for idx, duration_ms in enumerate(metrics.steps)
         ]
 
         memory_checkpoints = {
             name: snapshot.to_dict()
-            for name, snapshot in timings.memory_snapshots.items()
+            for name, snapshot in metrics.memory_snapshots.items()
         }
 
         report = {
             "timestamp": datetime.now(UTC).isoformat(),
-            "request_id": timings.request_id,
+            "request_id": metrics.request_id,
             "commit_hash": get_git_commit_hash(),
             "tag": tag,
-            "total_duration_ms": timings.total_duration_ms,
+            "total_duration_ms": metrics.total_duration_ms,
             "steps": formatted_steps,
             "denoise_steps_ms": denoise_steps_ms,
             "memory_checkpoints": memory_checkpoints,
@@ -309,14 +316,14 @@ class PerformanceLogger:
             os.makedirs(os.path.dirname(abs_path), exist_ok=True)
             with open(abs_path, "w", encoding="utf-8") as f:
                 json.dump(report, f, indent=2)
-            logger.info(f"Metrics dumped to: {abs_path}")
+            logger.info(f"Metrics dumped to: {CYAN}{abs_path}{RESET}")
         except IOError as e:
             logger.error(f"Failed to dump metrics to {abs_path}: {e}")
 
     @classmethod
     def log_request_summary(
         cls,
-        timings: "RequestMetrics",
+        metrics: "RequestMetrics",
         tag: str = "total_inference_time",
     ):
         """logs the stage metrics and total duration for a completed request
@@ -326,21 +333,21 @@ class PerformanceLogger:
         """
         formatted_stages = [
             {"name": name, "execution_time_ms": duration_ms}
-            for name, duration_ms in timings.stages.items()
+            for name, duration_ms in metrics.stages.items()
         ]
 
         memory_checkpoints = {
             name: snapshot.to_dict()
-            for name, snapshot in timings.memory_snapshots.items()
+            for name, snapshot in metrics.memory_snapshots.items()
         }
 
         record = RequestPerfRecord(
-            timings.request_id,
+            metrics.request_id,
             commit_hash=get_git_commit_hash(),
             tag="pipeline_stage_metrics",
             stages=formatted_stages,
-            steps=timings.steps,
-            total_duration_ms=timings.total_duration_ms,
+            steps=metrics.steps,
+            total_duration_ms=metrics.total_duration_ms,
             memory_snapshots=memory_checkpoints,
         )
 
diff --git a/python/sglang/multimodal_gen/test/server/perf_baselines.json b/python/sglang/multimodal_gen/test/server/perf_baselines.json
index 2543fa547..970089f62 100644
--- a/python/sglang/multimodal_gen/test/server/perf_baselines.json
+++ b/python/sglang/multimodal_gen/test/server/perf_baselines.json
@@ -542,7 +542,7 @@
                 "7": 93.97,
                 "8": 94.32
             },
-            "expected_e2e_ms": 1192.92,
+            "expected_e2e_ms": 1292.92,
             "expected_avg_denoise_ms": 83.75,
             "expected_median_denoise_ms": 93.58
         },