diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index b02400e1d..e0c9f7e25 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -309,6 +309,10 @@ class Envs:
     # Scheduler: others:
     SGLANG_EMPTY_CACHE_INTERVAL = EnvFloat(-1)  # in seconds. Set if you observe high memory accumulation over a long serving period.
     SGLANG_DISABLE_CONSECUTIVE_PREFILL_OVERLAP = EnvBool(False)
+    # Force-enable the WAR (write-after-read) barrier for the overlap scheduler
+    # even when is_cuda() is False (e.g. AMD/ROCm). On CUDA the barrier is
+    # already enabled regardless of this flag (see start_event_loop).
+    SGLANG_ENABLE_WAR_BARRIER = EnvBool(False)
     # PP: skip output send/recv when the entire batch consists of non-final chunked prefill requests,
     # since process_batch_result_prefill discards next_token_ids for those anyway.
     SGLANG_PP_SKIP_PURE_CHUNKED_OUTPUT_COMM = EnvBool(False)
diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index baac7d214..0603e5473 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -1424,7 +1424,9 @@ class Scheduler(
         # DFLASH fences its shared req_to_token writes with verify_done /
         # plan-stream deps, so the global WAR barrier only serializes plan
         # overlap. TODO: generalize this global-barrier enablement policy.
-        self._war_barrier_enabled = is_cuda() and not self.spec_algorithm.is_dflash()
+        self._war_barrier_enabled = (
+            is_cuda() or envs.SGLANG_ENABLE_WAR_BARRIER.get()
+        ) and not self.spec_algorithm.is_dflash()
         with self.device_module.StreamContext(self.schedule_stream):
             dispatch_event_loop(self)