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 8, 16:30 UTC
[Bugfix] Avoid duplicate k-proj weight emission in helper (#34142)
Signed-off-by: Artus KG <artuskg@gmail.com>
Artus Krohn-Grimberghe committed
Feb 9, 2026 at 20:17 UTC
eadb4e868bae8acd2e9b764f5827c0500ec44c34
1 file changed
+2
-2
vllm/model_executor/models/whisper.py
+2
-2
@@ -958,8 +958,8 @@ def _create_fake_bias_for_k_proj(
958
So that the bias for k_proj in qkv_proj can be initialized with zeros.
959
"""
960
for name, weight in weights:
961
+ yield name, weight
962
if name.endswith(fake_bias_key_name):
963
bias = torch.zeros(weight.size(0))
964
bias_name = name.replace("weight", "bias")
964
- yield from [(name, weight), (bias_name, bias)]
965
- yield name, weight
965
+ yield bias_name, bias