[HiCache] Fix deadlock when creating new group (#15805)

Signed-off-by: Xuchun Shang <xuchun.shang@gmail.com>
2025-12-29 17:13:04 +08:00
parent de2799f3f5
commit 4ab66d956f
2 changed files with 55 additions and 2 deletions
@@ -309,9 +309,13 @@ class HiCacheController:
            # create a new communication group for synchronizing storage operations across TP workers
            self.tp_world_size = torch.distributed.get_world_size(group=tp_group)
            if self.tp_world_size > 1:
+                from sglang.srt.distributed.parallel_state import (
+                    create_custom_parallel_group,
+                )
+
                group_ranks = torch.distributed.get_process_group_ranks(tp_group)
-                self.prefetch_tp_group = torch.distributed.new_group(
-                    group_ranks, backend="gloo"
+                self.prefetch_tp_group = create_custom_parallel_group(
+                    group_ranks=group_ranks, backend="gloo"
                )

            # Select the get and set functions