[BugFix] fix hang in async scheduling while open ENPU (#8354)

### What this PR does / why we need it?
1. there is no synchronization between steps. However, in async
scheduling with aclgraph, it is possible that the CPU's record event for
the current iteration completes before the previous iteration's graph
execution has finished. If cpu is fast enough, device will hang on
event_wait in interation i+1 (assume that event_record is executed
immediately on update stream of device).
2. Under ENPU, eagle proposers also need to follow event.record first,
and then event.Wait.

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?

---------

Signed-off-by: 1zzk <785396250@qq.com>

vllm_ascend/worker/model_runner_v1.py

@@ -336,6 +336,15 @@ class NPUModelRunner(GPUModelRunner):
             self.input_ids = self._make_buffer(max_buffer_num_tokens, dtype=torch.int32)
             self.positions = self._make_buffer(max_buffer_num_tokens, dtype=torch.int64)
 
+        self.use_eagle = (
+            vllm_config.speculative_config.method in ("eagle", "eagle3", "mtp")
+            if vllm_config.speculative_config
+            else False
+        )
+        # When True, run update_full_graph_params before self.model (ENPU / graph capture order).
+        # Internal / non-public toggle: read C getenv ``ENPU_ENABLE`` from enpu code (not in envs.py).
+        _enpu = get_c_env("ENPU_ENABLE")
+        self.enable_enpu = _enpu is not None and _enpu.lower() == "true"
         self._set_up_drafter()
 
         # kv role
@@ -424,9 +433,6 @@ class NPUModelRunner(GPUModelRunner):
             self.cudagraph_batch_sizes = []
         self.mamba_state_idx: dict[str, int] = {}
         self._mamba_copy_bufs: mamba_utils.MambaCopyBuffers | None = None
-        env_enpu_enable = get_c_env("ENPU_ENABLE")
-        # When True, run update_full_graph_params before self.model (ENPU / graph capture order).
-        self.enable_enpu = env_enpu_enable is not None and env_enpu_enable.lower() == "true"
 
     @property
     def use_cp(self) -> bool:
@@ -1795,6 +1801,9 @@ class NPUModelRunner(GPUModelRunner):
             and not forward_context.capturing
             and not self.use_sparse
         ):
+            if self.enable_enpu:
+                torch.npu.current_stream().synchronize()
+
             assert positions is not None
             update_full_graph_params(
                 self.attn_backend,
@@ -2592,7 +2601,13 @@ class NPUModelRunner(GPUModelRunner):
         # wrap the model with full graph wrapper if needed.
         if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
             self.update_stream: torch.npu.Stream = torch.npu.Stream()
-            self.model = ACLGraphWrapper(self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL)
+            self.model = ACLGraphWrapper(
+                self.model,
+                self.vllm_config,
+                runtime_mode=CUDAGraphMode.FULL,
+                use_eagle=self.use_eagle,
+                enable_enpu=self.enable_enpu,
+            )
 
     def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> None:
         """