apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Supports DP+TP+EP hybrid parallel deployment strategy (#3489)

Browse Source 
* Support DP+TP+EP hybrid parallel deployment strategy

* Support DP+TP+EP hybrid parallel deployment strategy

* fix conflict

* add moe_tp_ep function split_allgather_out

* del tp_group in moe_cutlass_backend

* for ci

* fix parallel_config for ci

* del log
This commit is contained in:
lzy
2025-08-26 15:04:01 +08:00
committed by GitHub
parent 52eda7fdb3
commit d339df2e90
15 changed files with 304 additions and 224 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
fastdeploy/distributed/communication.py
View File

@@ -47,15 +47,20 @@ try:
@paddle.jit.marker.unified
def tensor_model_parallel_all_reduce(
input_: paddle.Tensor,
group_: paddle.distributed.communication.group.Group = None,
) -> paddle.Tensor:
"""All-reduce the input tensor across model parallel group."""
global _TP_AR
if _TP_AR is not None and _TP_AR.should_custom_ar(input_):
# TODO: supports different_group custom allreduce
_TP_AR.custom_all_reduce(input_)
elif paddle.in_dynamic_mode():
hcg = fleet.get_hybrid_communicate_group()
mp_group = hcg.get_model_parallel_group()
dist.all_reduce(input_, group=mp_group)
if group_ is not None:
dist.all_reduce(input_, group=group_)
else:
hcg = fleet.get_hybrid_communicate_group()
mp_group = hcg.get_model_parallel_group()
dist.all_reduce(input_, group=mp_group)
else:
dist.all_reduce(input_)