[Core][Distributed] use cpu/gloo to initialize pynccl (#4248)
This commit is contained in:
@@ -298,12 +298,9 @@ def init_worker_distributed_environment(
|
||||
elif parallel_config.world_size > 1:
|
||||
# NOTE(woosuk): We don't initialize pynccl process group when world size
|
||||
# is 1.
|
||||
pynccl_utils.init_process_group(
|
||||
world_size=parallel_config.world_size,
|
||||
local_rank=local_rank,
|
||||
rank=rank,
|
||||
init_method=distributed_init_method,
|
||||
)
|
||||
# NOTE(kaichao): By default, pynccl will use information inside
|
||||
# `parallel_state` for initialization.
|
||||
pynccl_utils.init_process_group()
|
||||
|
||||
ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
|
||||
parallel_config.pipeline_parallel_size)
|
||||
|
||||
Reference in New Issue
Block a user