[diffusion] profile: support performance metric dumping and comparison (#13630)

python/sglang/multimodal_gen/runtime/managers/gpu_worker.py

@@ -3,6 +3,7 @@
 # SPDX-License-Identifier: Apache-2.0
 import multiprocessing as mp
 import os
+import time
 from typing import List
 
 import torch
@@ -25,6 +26,10 @@ from sglang.multimodal_gen.runtime.utils.logging_utils import (
     init_logger,
     suppress_other_loggers,
 )
+from sglang.multimodal_gen.runtime.utils.perf_logger import (
+    PerformanceLogger,
+    RequestTimings,
+)
 
 logger = init_logger(__name__)
 
@@ -86,27 +91,35 @@ class GPUWorker:
             f"Worker {self.rank}: Initialized device, model, and distributed environment."
         )
 
-    def execute_forward(self, batch: List[Req], server_args: ServerArgs) -> OutputBatch:
+    def execute_forward(self, batch: List[Req]) -> OutputBatch:
         """
         Execute a forward pass.
         """
         assert self.pipeline is not None
         # TODO: dealing with first req for now
         req = batch[0]
-        output_batch = self.pipeline.forward(req, server_args)
-        if req.perf_logger:
-            logging_info = getattr(output_batch, "logging_info", None) or getattr(
-                req, "logging_info", None
-            )
-            if logging_info:
-                try:
-                    req.perf_logger.log_stage_metrics(logging_info)
-                except Exception:
-                    logger.exception(
-                        "Failed to log stage metrics for request %s", req.request_id
-                    )
-            req.perf_logger.log_total_duration("total_inference_time")
-        return output_batch
+        output_batch = None
+        try:
+            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 output_batch.timings:
+                output_batch.timings.total_duration_ms = duration_ms
+                PerformanceLogger.log_request_summary(timings=output_batch.timings)
+        except Exception as e:
+            if output_batch is None:
+                from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import (
+                    OutputBatch,
+                )
+
+                output_batch = OutputBatch()
+            output_batch.error = f"Error executing request {req.request_id}: {e}"
+        finally:
+            return output_batch
 
     def set_lora_adapter(
         self, lora_nickname: str, lora_path: str | None = None