Fix expert distribution dumping causes OOM (#6967)

python/sglang/srt/managers/expert_distribution.py

@@ -703,6 +703,7 @@ class _StatAccumulator(_UtilizationRateAccumulatorMixin):
             dtype=torch.int32,
             device=self._server_args.device,
         )
+        self._first_dump = True
 
     def append(
         self,
@@ -727,9 +728,15 @@ class _StatAccumulator(_UtilizationRateAccumulatorMixin):
             num_logical_experts=self._expert_location_metadata.num_logical_experts,
             physical_to_logical_map=self._expert_location_metadata.physical_to_logical_map,
         )
+
+        if self._first_dump:
+            self._first_dump = False
+            torch.cuda.empty_cache()
+
         torch.distributed.all_reduce(
             logical_count_of_buffered_step, op=torch.distributed.ReduceOp.SUM
         )
+
         output = dict(
             rank=self._rank,
             logical_count=logical_count_of_buffered_step,