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

[Features] support hugging face qwen3 dense and qwen2 model (#3574)

Browse Source 
* support qwen2 and qwen3 hugging face

* fix moe

* defualt_v1 loader

* hugging_face_format deprecated

* modify hugging_face_foramt to model_format

* model_format auto

* fix environemt

* fix bug

* fix qwen3-0.6 bug

* model_format is str

* fix
This commit is contained in:
lizexu123
2025-08-26 10:54:53 +08:00
committed by GitHub
parent 66c5addce4
commit c43a4bec00
10 changed files with 182 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
fastdeploy/model_executor/layers/lm_head.py
View File

@@ -22,7 +22,7 @@ from paddle import nn
from paddle.distributed import fleet
from fastdeploy.config import FDConfig
from fastdeploy.model_executor.utils import set_weight_attrs
from fastdeploy.model_executor.utils import default_weight_loader, set_weight_attrs
from .utils import get_tensor
@@ -61,6 +61,7 @@ class ParallelLMHead(nn.Layer):
self.use_ep: bool = fd_config.parallel_config.use_ep
self.column_cut = True
self.nranks = fd_config.parallel_config.tensor_parallel_size
self.fd_config = fd_config
ColumnParallelLinear = fleet.meta_parallel.ColumnParallelLinear
RowParallelLinear = fleet.meta_parallel.RowParallelLinear
@@ -90,7 +91,14 @@ class ParallelLMHead(nn.Layer):
weight_attr=None,
has_bias=True if self.bias_key is not None else False,
gather_output=need_gather,
fuse_matmul_bias=False, # False diff更小
fuse_matmul_bias=False,
)
set_weight_attrs(
self.linear.weight,
{
"weight_loader": default_weight_loader(self.fd_config),
"model_format": self.fd_config.model_config.model_format,
},
)
if self.nranks > 1:
set_weight_attrs(self.linear.weight, {"output_dim": True})
@@ -102,8 +110,16 @@ class ParallelLMHead(nn.Layer):
weight_attr=None,
has_bias=True if self.bias_key is not None else False,
input_is_parallel=False,
fuse_matmul_bias=False, # False diff更小
fuse_matmul_bias=False,
)
set_weight_attrs(
self.linear.weight,
{
"weight_loader": default_weight_loader(self.fd_config),
"model_format": self.fd_config.model_config.model_format,
},
)
if self.nranks > 1:
set_weight_attrs(self.linear.weight, {"output_dim": False})