apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 00:57:33 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

qwen3 0.3B fix (#3255)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source
This commit is contained in:
bukejiyu
2025-08-08 11:35:40 +08:00
committed by GitHub
parent fac2f64837
commit b76b17fc1b
2 changed files with 22 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
fastdeploy/model_executor/layers/linear.py
View File

@@ -266,10 +266,6 @@ class ReplicatedLinear(LinearBase):
) )
self.hidden_size = fd_config.model_config.hidden_size self.hidden_size = fd_config.model_config.hidden_size
self.weight_shape = [
self.input_size,
self.output_size,
]
assert self.quant_method is not None assert self.quant_method is not None
self.quant_method.create_weights( self.quant_method.create_weights(
@@ -311,24 +307,21 @@ class ColumnParallelLinear(LinearBase):
add_bias (bool): Whether to add bias in the current layer or in the pre/post layer. Defaults to False. add_bias (bool): Whether to add bias in the current layer or in the pre/post layer. Defaults to False.
skip_quant (bool): Whether to skip quantization. Defaults to False. skip_quant (bool): Whether to skip quantization. Defaults to False.
""" """
super().__init__(
fd_config=fd_config,
prefix=prefix,
input_size=input_size,
output_size=output_size,
with_bias=with_bias,
add_bias=add_bias,
skip_quant=skip_quant,
)
self.fd_config = fd_config self.fd_config = fd_config
self.nranks = fd_config.parallel_config.tensor_parallel_size self.nranks = fd_config.parallel_config.tensor_parallel_size
self.input_size = input_size self.input_size = input_size
self.output_size = divide(output_size, self.nranks) # Split the output_size using TP inference. self.output_size = divide(output_size, self.nranks) # Split the output_size using TP inference.
self.hidden_size = fd_config.model_config.hidden_size self.hidden_size = fd_config.model_config.hidden_size
self.weight_shape = [
self.input_size, super().__init__(
self.output_size, fd_config=fd_config,
] prefix=prefix,
input_size=self.input_size,
output_size=self.output_size,
with_bias=with_bias,
add_bias=add_bias,
skip_quant=skip_quant,
)
assert self.quant_method is not None assert self.quant_method is not None
self.quant_method.create_weights( self.quant_method.create_weights(
@@ -634,15 +627,6 @@ class RowParallelLinear(LinearBase):
add_bias (bool): Whether to add bias in the current layer or in the pre/post layer. Defaults to False. add_bias (bool): Whether to add bias in the current layer or in the pre/post layer. Defaults to False.
skip_quant (bool): Whether to skip quantization. Defaults to False. skip_quant (bool): Whether to skip quantization. Defaults to False.
""" """
super().__init__(
fd_config=fd_config,
prefix=prefix,
input_size=input_size,
output_size=output_size,
with_bias=with_bias,
add_bias=add_bias,
skip_quant=skip_quant,
)
self.fd_config = fd_config self.fd_config = fd_config
self.skip_quant = False self.skip_quant = False
self.nranks = fd_config.parallel_config.tensor_parallel_size self.nranks = fd_config.parallel_config.tensor_parallel_size
@@ -654,11 +638,15 @@ class RowParallelLinear(LinearBase):
self.input_size = divide(input_size, self.nranks) self.input_size = divide(input_size, self.nranks)
self.output_size = output_size self.output_size = output_size
self.weight_shape = [ super().__init__(
self.input_size, fd_config=fd_config,
self.output_size, prefix=prefix,
] input_size=self.input_size,
self._dtype = self._helper.get_default_dtype() output_size=self.output_size,
with_bias=with_bias,
add_bias=add_bias,
skip_quant=skip_quant,
)
assert self.quant_method is not None assert self.quant_method is not None
self.quant_method.create_weights( self.quant_method.create_weights(

3
fastdeploy/model_executor/models/qwen3.py
View File

@@ -286,6 +286,9 @@ class Qwen3ForCausalLM(ModelForCasualLM):
weight_loader = getattr(param, "weight_loader", default_weight_loader(self.fd_config)) weight_loader = getattr(param, "weight_loader", default_weight_loader(self.fd_config))
weight_loader(param, loaded_weight) weight_loader(param, loaded_weight)
if self.tie_word_embeddings:
self.lm_head.linear.weight.set_value(self.model.embed_tokens.embeddings.weight.transpose([1, 0]))
@paddle.no_grad() @paddle.no_grad()
def set_state_dict(self, state_dict): def set_state_dict(self, state_dict):
""" """