diff --git a/tests/v1/worker/test_gpu_profiler.py b/tests/v1/worker/test_gpu_profiler.py
index 3e236a73b1..347b5b379a 100644
--- a/tests/v1/worker/test_gpu_profiler.py
+++ b/tests/v1/worker/test_gpu_profiler.py
@@ -112,6 +112,34 @@ def test_max_iterations(default_profiler_config):
     # Should have stopped now
     assert profiler._running is False
     assert profiler.stop_call_count == 1
+    # And fully reset, not just paused -- a later start_profile must not be a
+    # permanent no-op just because max_iterations already fired once.
+    assert profiler._active is False
+    assert profiler._active_iteration_count == 0
+    assert profiler._profiling_for_iters == 0
+
+
+def test_restart_after_max_iterations(default_profiler_config):
+    """A start_profile after an auto-stop must actually restart, not be
+    silently ignored (regression test: auto-stop used to leave _active
+    True forever, so start() always bailed out early)."""
+    default_profiler_config.max_iterations = 2
+    profiler = ConcreteWorkerProfiler(default_profiler_config)
+
+    profiler.start()
+    profiler.step()
+    profiler.step()
+    profiler.step()  # exceeds max, auto-stops
+    assert profiler._running is False
+    assert profiler.start_call_count == 1
+
+    profiler.start()
+    assert profiler._active is True
+    assert profiler._running is True
+    assert profiler.start_call_count == 2
+
+    profiler.step()
+    assert profiler._running is True
 
 
 def test_delayed_start_and_max_iters(default_profiler_config):
diff --git a/vllm/profiler/wrapper.py b/vllm/profiler/wrapper.py
index 2f56d0fc7c..877a390d6d 100644
--- a/vllm/profiler/wrapper.py
+++ b/vllm/profiler/wrapper.py
@@ -120,11 +120,13 @@ class WorkerProfiler(ABC):
             and self._running
             and self._profiling_for_iters > self._max_iters
         ):
-            # Automatically stop the profiler after max iters
-            # will be marked as not running, but leave as active so that stop
-            # can clean up properly
+            # Automatically stop the profiler after max iters. Go through the
+            # public stop() (not _call_stop() directly) so _active and the
+            # iteration counters reset too -- otherwise a later start_profile
+            # is silently ignored forever, since start() bails out early
+            # whenever _active is still True.
             logger.info_once("Max profiling iterations reached. Stopping profiler...")
-            self._call_stop()
+            self.stop()
             return
 
     def _profiler_step(self) -> bool: