[Bugfix] Report stepwise Qwen-Image profiler metrics (#4517)

Signed-off-by: akshatvishu <akshatnayak197@gmail.com>

vllm_omni/diffusion/worker/utils.py

@@ -15,6 +15,41 @@ if TYPE_CHECKING:
     from vllm_omni.inputs.data import OmniDiffusionSamplingParams, OmniPromptType
 
 
+def clear_pipeline_stage_durations(pipeline: Any) -> None:
+    clear_records = getattr(pipeline, "clear_profiler_records", None)
+    if getattr(pipeline, "enable_diffusion_pipeline_profiler", False) and callable(clear_records):
+        clear_records()
+
+
+def consume_pipeline_stage_durations(pipeline: Any) -> dict[str, float]:
+    if not getattr(pipeline, "enable_diffusion_pipeline_profiler", False):
+        return {}
+    stage_durations = getattr(pipeline, "stage_durations", None)
+    if not isinstance(stage_durations, dict):
+        return {}
+    result = {stage: float(duration) for stage, duration in stage_durations.items()}
+    clear_pipeline_stage_durations(pipeline)
+    return result
+
+
+def merge_stage_durations(
+    state: DiffusionRequestState,
+    stage_durations: dict[str, float],
+) -> None:
+    if not stage_durations:
+        return
+    for stage, duration in stage_durations.items():
+        state.stage_durations[stage] = float(state.stage_durations.get(stage, 0.0)) + float(duration)
+
+
+def attach_stage_durations(
+    state: DiffusionRequestState,
+    output: DiffusionOutput,
+) -> None:
+    if state.stage_durations:
+        output.stage_durations = dict(state.stage_durations)
+
+
 @dataclass
 class DiffusionRequestState:
     """Per-request mutable state across all pipeline stages.
@@ -79,6 +114,9 @@ class DiffusionRequestState:
     # For example: Wan condition tensors / masks, or Bagel KV contexts.
     extra: dict[str, Any] = field(default_factory=dict)
 
+    # ── Runner-owned profiling metadata ──
+    stage_durations: dict[str, float] = field(default_factory=dict)
+
     # Peak device memory observed while this request is active in step mode.
     peak_memory_mb: float = 0.0