[MLX] Fix step-bounded profiling for bench tools on Apple Silicon (#29217)

python/sglang/benchmark/offline_throughput.py

@@ -284,7 +284,15 @@ def throughput_test_once(
         dir = os.getenv("SGLANG_TORCH_PROFILER_DIR")
         if not profile_steps:
             known_files = set(os.listdir(dir))
+        # With --profile-steps the scheduler auto-stops mid-run after N steps, so
+        # a second stop here raises "not in progress"; a run shorter than N steps
+        # never hit the target and still needs this explicit stop. Either way we
+        # must stop before monitor_trace_file, which loops forever waiting for a
+        # trace that would otherwise never be finalized.
+        try:
             backend.stop_profile()
+        except RuntimeError:
+            pass
         monitor_trace_file(known_files, dir)
 
     if backend_name == "runtime":

python/sglang/srt/hardware_backend/mlx/scheduler_mixin.py

@@ -86,6 +86,16 @@ class SchedulerMlxOverlapMixin:
     """Mixin that adds MLX overlap scheduling to :class:`Scheduler`."""
 
     def _finalize_mlx_pending_job(self: Scheduler, pending: MlxPendingJob):
+        # Account for this completed forward step. The standard scheduler does
+        # this inside run_batch(), but the MLX overlap loop bypasses run_batch,
+        # so without this forward_ct never advances on MLX. That stalls the
+        # watchdog liveness counter and, more importantly, breaks step-bounded
+        # profiling: _profile_batch_predicate auto-starts/stops based on
+        # forward_ct, so `--profile-steps` (and the server /start_profile
+        # num_steps path) only takes effect once the counter moves here.
+        self.forward_ct += 1
+        self.profiler_manager._profile_batch_predicate(pending.schedule_batch)
+
         result = self.tp_worker.finalize_mlx_result(
             pending.prefills,
             pending.extends,