[Bugfix][Memory] Profile startup memory under inference mode again (#1391)

Inserting _prefill_graph_reserve between the decorator and
determine_available_memory moved @torch.inference_mode() off the profiling
entry point. The V1 runner's profile_run is not decorated itself, so it ran the
vision tower with autograd on and kept every block's activations alive: 27B
NVFP4 without a draft profiled a 17.89 GiB activation peak and refused to start
at 256K for lack of KV memory.

Co-authored-by: yangzhuxinyzx <Susannah_Carterect@gmx.ca>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

tests/v1/worker/test_gpu_worker_memory_profile.py

@@ -30,10 +30,12 @@ class _FakeRunner:
         self._device = device
         self.calls = 0
         self.graph_calls = 0
+        self.inference_modes: list[bool] = []
         self._kept: list[torch.Tensor] = []
 
     def profile_run(self) -> None:
         self.calls += 1
+        self.inference_modes.append(torch.is_inference_mode_enabled())
         if self.calls == 1:
             scratch = torch.empty(1024 * MiB, dtype=torch.uint8, device=self._device)
             self._kept.append(
@@ -95,8 +97,11 @@ def test_kv_budget_ignores_cold_compile_scratch(
     worker.vllm_config = SimpleNamespace(
         compilation_config=SimpleNamespace(
             cudagraph_mode=graph_mode,
+            cudagraph_capture_sizes=None,
             runtime=GraphPolicy(estimate_graph_memory=estimate_graphs),
-        )
+        ),
+        scheduler_config=SimpleNamespace(max_num_seqs=1),
+        speculative_config=None,
     )
     runner = _FakeRunner(device, weights_bytes)
     worker.model_runner = runner
@@ -117,6 +122,8 @@ def test_kv_budget_ignores_cold_compile_scratch(
     )
     assert abs(charged - 64 * MiB) <= 32 * MiB
     assert runner.calls == 2
+    # With autograd on, an encoder forward keeps every block's activations alive.
+    assert runner.inference_modes == [True, True]
     assert runner.graph_calls == int(graph_bytes > 0)
     assert worker.cudagraph_memory_estimate == graph_bytes
     del weights

vllm/v1/worker/gpu_worker.py

@@ -515,7 +515,6 @@ class Worker(WorkerBase):
             )
         self.model_runner.reload_weights(*args, **kwargs)
 
-    @torch.inference_mode()
     def _prefill_graph_reserve(self, activation_peak: int) -> int:
         """Graph pool bytes to keep out of the KV budget without the estimator.
 
@@ -546,6 +545,7 @@ class Worker(WorkerBase):
         )
         return max(0, activation_peak)
 
+    @torch.inference_mode()
     def determine_available_memory(self) -> int:
         """Profiles the peak memory usage of the model to determine how much
         memory can be used for KV cache without OOMs.