Hot fix for hicache with new page aligned radixtree (#4397)

python/sglang/srt/managers/cache_controller.py

@@ -248,6 +248,8 @@ class HiCacheController:
         if device_indices is None:
             return None
         self.mem_pool_host.protect_load(host_indices)
+        # to ensure the device indices are ready before accessed by another CUDA stream
+        torch.cuda.current_stream().synchronize()
         self.load_queue.put(
             CacheOperation(host_indices, device_indices, node_id, priority)
         )

python/sglang/srt/managers/scheduler.py

@@ -434,6 +434,7 @@ class Scheduler(SchedulerOutputProcessorMixin):
                     req_to_token_pool=self.req_to_token_pool,
                     token_to_kv_pool_allocator=self.token_to_kv_pool_allocator,
                     tp_cache_group=self.tp_worker.get_tp_cpu_group(),
+                    page_size=self.page_size,
                 )
             else:
                 self.tree_cache = RadixCache(