apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 16:22:57 +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

15
fastdeploy/model_executor/layers/linear.py
View File

@@ -57,6 +57,7 @@ class UnquantizedLinearMethod(QuantMethodBase):
{
**extra_weight_attrs,
"weight_loader": extra_weight_attrs.get("weight_loader", default_weight_loader(layer.fd_config)),
"model_format": extra_weight_attrs.get("model_format", ""),
},
)
@@ -343,7 +344,9 @@ class ColumnParallelLinear(LinearBase):
weight_loader=(
self.weight_loader if hasattr(self, "weight_loader") else default_weight_loader(self.fd_config)
),
model_format=fd_config.model_config.model_format,
)
if self.nranks > 0:
if self.with_bias:
# col parallel
@@ -402,6 +405,9 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
)
def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = None):
model_format = getattr(param, "model_format", "")
if model_format == "torch":
loaded_weight = loaded_weight.transpose([1, 0])
output_dim = getattr(param, "output_dim", None)
assert output_dim is not None
shard_dim = -1 if output_dim else 0
@@ -424,7 +430,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
# Tensor parallelism splits the weight along the output_dim
if self.nranks != 1:
dim = -1 if output_dim else 0
if isinstance(loaded_weight, np.ndarray):
if isinstance(loaded_weight, (np.ndarray, paddle.Tensor)):
size = loaded_weight.shape[dim]
else:
size = loaded_weight.get_shape()[dim]
@@ -523,6 +529,9 @@ class QKVParallelLinear(ColumnParallelLinear):
assert output_dim is not None
dim = -1 if output_dim else 0
head_dim = param.shape[dim] // (self.num_heads_per_rank + 2 * self.kv_num_heads_per_rank)
model_format = getattr(param, "model_format", "")
if model_format == "torch":
loaded_weight = loaded_weight.transpose([1, 0])
if loaded_shard_id is None:
# Loaded weight is already fused on disk
shard_offsets = [
@@ -541,7 +550,8 @@ class QKVParallelLinear(ColumnParallelLinear):
assert loaded_shard_id in ["q", "k", "v"]
# Tensor parallelism splits the weight along the output_dim
if self.nranks != 1:
if isinstance(loaded_weight, np.ndarray):
dim = -1 if output_dim else 0
if isinstance(loaded_weight, (np.ndarray, paddle.Tensor)):
size = loaded_weight.shape[dim]
else:
size = loaded_weight.get_shape()[dim]
@@ -712,6 +722,7 @@ class RowParallelLinear(LinearBase):
weight_loader=(
self.weight_loader if hasattr(self, "weight_loader") else default_weight_loader(self.fd_config)
),
model_format=fd_config.model_config.model_format,
)
if self.nranks > 0:
if self.with_bias: