diff --git a/vllm/v1/worker/gpu/async_utils.py b/vllm/v1/worker/gpu/async_utils.py
index b3d6f5e4d..e4659104f 100644
--- a/vllm/v1/worker/gpu/async_utils.py
+++ b/vllm/v1/worker/gpu/async_utils.py
@@ -24,7 +24,8 @@ class AsyncOutput(AsyncModelRunnerOutput):
         self.model_runner_output = model_runner_output
         self.sampler_output = sampler_output
         self.num_sampled_tokens = num_sampled_tokens
-        self.copy_event = torch.cuda.Event()
+        # Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
+        self.copy_event = torch.cuda.Event(blocking=True)
 
         with stream(copy_stream, main_stream):
             copy_stream.wait_stream(main_stream)
@@ -81,7 +82,8 @@ class AsyncPoolingOutput(AsyncModelRunnerOutput):
         self.model_runner_output = model_runner_output
         self.pooler_output = pooler_output
         self.is_valid = is_valid
-        self.copy_event = torch.cuda.Event()
+        # Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
+        self.copy_event = torch.cuda.Event(blocking=True)
 
         with stream(copy_stream, main_stream):
             copy_stream.wait_stream(main_stream)
diff --git a/vllm/v1/worker/gpu/spec_decode/utils.py b/vllm/v1/worker/gpu/spec_decode/utils.py
index 4ab45b2ae..8e73ec8ab 100644
--- a/vllm/v1/worker/gpu/spec_decode/utils.py
+++ b/vllm/v1/worker/gpu/spec_decode/utils.py
@@ -12,7 +12,8 @@ class DraftTokensHandler:
     def __init__(self, device: torch.device | None = None):
         self.device = device
         self.copy_stream = torch.cuda.Stream(device)
-        self.copy_event = torch.cuda.Event()
+        # Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
+        self.copy_event = torch.cuda.Event(blocking=True)
 
         self.req_ids: list[str] = []
         self.draft_tokens_np: np.ndarray | None = None
diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py
index 74938a823..8756c97a1 100644
--- a/vllm/v1/worker/gpu_model_runner.py
+++ b/vllm/v1/worker/gpu_model_runner.py
@@ -254,7 +254,8 @@ class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
         self._invalid_req_indices = invalid_req_indices
 
         # Event on the copy stream so we can synchronize the non-blocking copy.
-        self.async_copy_ready_event = torch.Event()
+        # Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
+        self.async_copy_ready_event = torch.cuda.Event(blocking=True)
 
         # Keep a reference to the device tensor to avoid it being
         # deallocated until we finish copying it to the host.
@@ -375,7 +376,8 @@ class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
         self._model_runner_output = model_runner_output
 
         # Event on the copy stream so we can synchronize the non-blocking copy.
-        self.async_copy_ready_event = torch.Event()
+        # Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
+        self.async_copy_ready_event = torch.cuda.Event(blocking=True)
 
         # Keep a reference to the device tensors to avoid them being
         # deallocated until we finish copying it to the host.
@@ -697,7 +699,9 @@ class GPUModelRunner(
         self.prepare_inputs_event: torch.Event | None = None
         if self.use_async_scheduling:
             self.async_output_copy_stream = torch.cuda.Stream()
-            self.prepare_inputs_event = torch.Event()
+            # Blocking (sleep) event to avoid busy-polling the CUDA driver lock;
+            # under TP contention that spin can balloon and make the rank a straggler.
+            self.prepare_inputs_event = torch.cuda.Event(blocking=True)
 
         # self.cudagraph_batch_sizes sorts in ascending order.
         if (