[Spec] Publish the final multi-layer EAGLE shared-read event (#36752)

Co-authored-by: Qiaolin-Yu <liin1211@outlook.com>

python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py

@@ -812,6 +812,7 @@ class MultiLayerEagleMultiStepDraftExtendCudaGraphRunner:
         out = runner.replay(
             self.bs, self.seq_lens_sum, self._replay_spec_info, self.seq_lens_cpu
         )
+        self._publish_shared_read_done(step)
         raw_bs = self.raw_bs
         raw_num_tokens = self.raw_num_tokens
         num_logit_rows = raw_bs if self.prune_draft_extend_logits else raw_num_tokens
@@ -829,6 +830,14 @@ class MultiLayerEagleMultiStepDraftExtendCudaGraphRunner:
             out.topk_index[:raw_bs],
         )
 
+    def _publish_shared_read_done(self, step: int) -> None:
+        if step != self.speculative_num_steps - 1:
+            return
+        runner = self.runners[step]
+        read_done = runner.device_module.Event()
+        read_done.record()
+        runner.model_runner.shared_read_done_event = read_done
+
     def clone_draft_probs(self) -> torch.Tensor:
         """Materialize the in-graph-written proposal q [raw_bs, num_steps, vocab]
         after replay; the clone must land before a later replay rewrites the
@@ -989,4 +998,6 @@ class OneGraphMultiLayerEagleMultiStepDraftExtendCudaGraphRunner(
                     out.topk_p[:raw_bs],
                     out.topk_index[:raw_bs],
                 )
-        return self._cached[step]
+        result = self._cached[step]
+        self._publish_shared_read_done(step)
+        return result

python/sglang/srt/speculative/multi_layer_eagle_worker_v2.py

@@ -962,8 +962,7 @@ class MultiLayerEagleWorkerV2(BaseSpecWorker):
 
     @property
     def last_shared_read_runner(self):
-        # Multi-layer eagle has no draft forward, only draft extend.
-        return self._draft_worker.draft_runner
+        return self._draft_worker.draft_runner_list[-1]
 
     @property
     def spec_v2_attn_backends(self) -> tuple: