apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +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

13
fastdeploy/model_executor/models/tp_utils.py
View File

@@ -72,6 +72,7 @@ class TensorSplitMode(Enum):
"""TensorSplitMode"""
GQA = "is_gqa"
TP_ROW_BIAS = "is_tp_row_bias"
TRANSPOSE = "transpose"
QKV = "is_old_qkv"
PairFused = "is_naive_2fuse"
@@ -212,7 +213,7 @@ def gqa_qkv_split_func(
"""
def fn(x, is_column=True):
"""fucn"""
"""func"""
def get_shape(tensor):
"""get_shape"""
@@ -430,7 +431,15 @@ def split_or_merge_func_v1(
def fn(x, **kwargs):
"""func"""
is_gqa = kwargs.pop("is_gqa", False)
if is_gqa:
is_tp_row_bias = kwargs.pop("is_tp_row_bias", False)
if is_tp_row_bias:
tensor = x[:, ...]
if isinstance(tensor, paddle.Tensor):
res = tensor / tensor_parallel_degree
else:
res = paddle.to_tensor(tensor, paddle.get_default_dtype()) / tensor_parallel_degree
return res
elif is_gqa:
func = split_or_merge_qkv_func(
is_split=is_split,
tensor_parallel_degree=tensor_parallel_degree,