[Core] Skip JIT monitor when JIT warmup is disabled (#58590)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Kimi Code <noreply@moonshot.cn>

tests/v1/worker/test_gpu_worker.py

@@ -258,3 +258,32 @@ def test_execute_model_waits_previous_pp_send_before_forward(
 
     assert log == ["wait:prev-tensor", "forward", "isend"]
     assert worker._pp_send_work == [tensor_handle]
+
+
+def test_jit_monitor_activation_follows_enable_jit_warmup(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    """The post-warmup JIT monitor must stay off when JIT warmup is disabled
+    (e.g. by enforce_eager): runtime compilation is then expected, and
+    warning/erroring on it would be noise."""
+    from vllm.utils import jit_monitor
+
+    calls = []
+    monkeypatch.setattr(jit_monitor, "activate", lambda **kwargs: calls.append(kwargs))
+
+    def worker(enable_jit_warmup):
+        return SimpleNamespace(
+            vllm_config=SimpleNamespace(
+                kernel_config=SimpleNamespace(enable_jit_warmup=enable_jit_warmup)
+            ),
+            observability_config=SimpleNamespace(
+                jit_monitor_mode="warn", jit_monitor_verbose=False
+            ),
+        )
+
+    gpu_worker.Worker._maybe_activate_jit_monitor(worker(True))
+    assert calls == [{"mode": "warn", "verbose": False}]
+
+    calls.clear()
+    gpu_worker.Worker._maybe_activate_jit_monitor(worker(False))
+    assert calls == []

vllm/v1/worker/gpu_worker.py

@@ -966,12 +966,7 @@ class Worker(WorkerBase):
 
         # All warmup is done — start monitoring for unexpected JIT
         # compilations that would cause latency spikes during inference.
-        from vllm.utils.jit_monitor import activate as activate_jit_monitor
-
-        activate_jit_monitor(
-            mode=self.observability_config.jit_monitor_mode,
-            verbose=self.observability_config.jit_monitor_verbose,
-        )
+        self._maybe_activate_jit_monitor()
 
         # Freeze the worker heap so the GC won't scan static objects
         # (model weights, KV caches, CUDA graphs) during inference.
@@ -987,6 +982,20 @@ class Worker(WorkerBase):
             encoder=self.compilation_config.encoder_compilation_time,
         )
 
+    def _maybe_activate_jit_monitor(self) -> None:
+        # When JIT warmup is disabled (e.g. enforce_eager), runtime JIT
+        # compilation is expected, so monitoring would only produce noise
+        # (or spurious errors in "error" mode).
+        if not self.vllm_config.kernel_config.enable_jit_warmup:
+            return
+
+        from vllm.utils.jit_monitor import activate as activate_jit_monitor
+
+        activate_jit_monitor(
+            mode=self.observability_config.jit_monitor_mode,
+            verbose=self.observability_config.jit_monitor_verbose,
+        )
+
     def _get_cudagraph_capture_context(self) -> AbstractContextManager[None]:
         """Let the configured profiler observe CUDA graph capture."""
         if not self.use_v2_model_runner: