diff --git a/tests/fmha_v3_stage_c_example7.py b/tests/fmha_v3_stage_c_example7.py index 374e4124..d8b98568 100644 --- a/tests/fmha_v3_stage_c_example7.py +++ b/tests/fmha_v3_stage_c_example7.py @@ -435,11 +435,11 @@ class FmhaV3StageCMulti: cute.arch.fence_view_async_tmem_load() # Async-proxy fence so the TMA store sees the SMEM writes. cute.arch.fence_proxy("async.shared", space="cta") - cute.arch.barrier_arrive( + cute.arch.mbarrier_arrive( barrier_id=self.epilog_sync_bar_id, number_of_threads=32 * len(self.epilogue_warp_id), ) - cute.arch.barrier_wait( + cute.arch.mbarrier_wait( barrier_id=self.epilog_sync_bar_id, number_of_threads=32 * len(self.epilogue_warp_id), )