未验证 提交 ee4a20cc 编写于 作者: J james 提交者: GitHub

Bugfix: xpu now only support single node multi-card, bkcl_comm_num should always set to 1 (#48961)

上级 121b7429
...@@ -137,7 +137,7 @@ class GraphExecutionOptimizer(MetaOptimizerBase): ...@@ -137,7 +137,7 @@ class GraphExecutionOptimizer(MetaOptimizerBase):
attrs={ attrs={
"trainers": trainer_endpoints, "trainers": trainer_endpoints,
"trainer_id": trainer_id, "trainer_id": trainer_id,
"nccl_comm_num": build_strategy.nccl_comm_num, "bkcl_comm_num": build_strategy.bkcl_comm_num,
"use_hierarchical_allreduce": build_strategy.use_hierarchical_allreduce, "use_hierarchical_allreduce": build_strategy.use_hierarchical_allreduce,
"hierarchical_allreduce_inter_ranks": build_strategy.hierarchical_allreduce_inter_nranks, "hierarchical_allreduce_inter_ranks": build_strategy.hierarchical_allreduce_inter_nranks,
}, },
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册