[GG] Measure repeatable activation peak after kernel warmup

tests/v1/worker/test_gpu_worker.py

@@ -111,6 +111,7 @@ def test_memory_profile_replays_model_after_kernel_warmup(
 ):
     worker = object.__new__(Worker)
     calls = []
+    worker.device = "cuda:0"
     worker.model_runner = SimpleNamespace(
         profile_run=lambda: calls.append("profile"),
     )
@@ -127,10 +128,21 @@ def test_memory_profile_replays_model_after_kernel_warmup(
         "_warmup_kernels_once",
         lambda: calls.append("kernel_warmup"),
     )
+    monkeypatch.setattr(
+        gpu_worker_module.torch.accelerator,
+        "reset_peak_memory_stats",
+        lambda device: calls.append(("reset_peak", device)),
+    )
 
     worker._profile_model_with_kernel_warmup()
 
-    assert calls == ["profile", "compressor", "kernel_warmup", "profile"]
+    assert calls == [
+        "profile",
+        "compressor",
+        "kernel_warmup",
+        ("reset_peak", "cuda:0"),
+        "profile",
+    ]
 
 
 @pytest.mark.parametrize("video_backend", [None, "opencv"])

vllm/v1/worker/gpu_worker.py

@@ -486,6 +486,12 @@ def _profile_model_with_kernel_warmup(self) -> None:
         )
         self._warmup_kernels_once()
 
+        # The first pass may include one-time compiler or loader temporaries.
+        # Keep every persistent allocation initialized above, but exclude that
+        # startup-only high-water mark from the activation measurement. The
+        # second pass below establishes the repeatable serving peak.
+        torch.accelerator.reset_peak_memory_stats(self.device)
+
         # Kernel warmup can create persistent modules and communication pools.
         # A second pass is required so the measured peak contains both those
         # allocations and the model's transient activation workspace.