A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 07:30 UTC
[Bugfix] deepseek-V3.2 self.weights_proj has no bias (#30841)
Signed-off-by: baoqian <1354987947@qq.com> Signed-off-by: baoqian426 <1354987947@qq.com>
baoqian426 committed
Dec 17, 2025 at 19:32 UTC
84896fda22d3de74398a88b5769c98eef14258f1
1 file changed
+5
-1
vllm/model_executor/models/deepseek_v2.py
+5
-1
@@ -835,7 +835,11 @@ class Indexer(nn.Module):
835
)
836
self.k_norm = LayerNorm(self.head_dim, eps=1e-6)
837
self.weights_proj = ReplicatedLinear(
838
- hidden_size, self.n_head, quant_config=None, prefix=f"{prefix}.weights_proj"
838
+ hidden_size,
839
+ self.n_head,
840
+ bias=False,
841
+ quant_config=None,
842
+ prefix=f"{prefix}.weights_proj",
843
)
844
self.softmax_scale = self.head_dim**-0.5
845