squall/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix ReplicatedLinear weight loading (#6793)

Browse Source
This commit is contained in:
QQSong 2024-07-25 19:24:58 -07:00 committed by GitHub
parent 2eb9f4ff26
commit 062a1d0fab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/linear.py
View File

@ -199,12 +199,16 @@ class ReplicatedLinear(LinearBase):
self.input_size, self.input_size,
self.output_size, self.output_size,
self.params_dtype, self.params_dtype,
weight_loader=self.weight_loader,
prefix=prefix) prefix=prefix)
if bias: if bias:
self.bias = Parameter( self.bias = Parameter(
torch.empty(self.output_size, dtype=self.params_dtype)) torch.empty(self.output_size, dtype=self.params_dtype))
set_weight_attrs(self.bias, {"output_dim": 0}) set_weight_attrs(self.bias, {
"output_dim": 0,
"weight_loader": self.weight_loader,
})
else: else:
self.register_parameter("bias", None) self.register_parameter("bias", None)