[Spec] Publish DFLASH verify read-done event for fine-grained WAR barrier (#29541)

python/sglang/srt/model_executor/runner/decode_cuda_graph_runner.py

@@ -995,14 +995,13 @@ class DecodeCudaGraphRunner(BaseCudaGraphRunner):
         )
         with timer_ctx, self.backend.replay_session():
             self.load_batch(forward_batch, pp_proxy_tensors)
-            # Snapshot built -- publish a read-done event for the WAR barrier.
+            # Publish a read-done event for the WAR barrier: a cuda-graph forward
-            # Only plain DECODE: the captured decode graph reads only its static
+            # finishes its shared req_to_token / SWA reads at this pre-replay
-            # snapshot, so the forward is done reading the shared pool here. Spec
+            # snapshot, so plain DECODE and DFLASH TARGET_VERIFY both qualify.
-            # verify (target_verify/dllm_extend) replays on this runner too, but
+            if forward_batch.forward_mode.is_decode() or (
-            # those are NOT the step's last shared-buffer-reading phase (eagle
+                forward_batch.forward_mode.is_target_verify()
-            # publishes from draft_extend; ngram/dflash must not publish here),
+                and self.model_runner.spec_algorithm.is_dflash()
-            # and some verify graphs may read beyond the snapshot in replay.
+            ):
-            if forward_batch.forward_mode.is_decode():
                 read_done = self.device_module.Event()
                 read_done.record()
                 self.model_runner.war_fastpath_read_done_event = read_done