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 2, 08:30 UTC
[Bugfix] fix wrong partial_rotary_factor calculation for bailing_moe model. (#43770)
Signed-off-by: zzt <zengzetang.zzt@antgroup.com> Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
zzt committed
Jun 1, 2026 at 17:42 UTC
8796838910a0c12149084c151a221ed384dea2dd
1 file changed
+6
-1
vllm/model_executor/models/bailing_moe.py
+6
-1
@@ -130,7 +130,12 @@ class BailingAttention(nn.Module):
130
prefix=f"{prefix}.dense",
131
)
132
133
- rotary_dim = getattr(config, "rotary_dim", self.head_dim)
133
+ rotary_dim = getattr(config, "rotary_dim", None)
134
+ if rotary_dim is None:
135
+ partial_rotary_factor = getattr(config, "partial_rotary_factor", 1.0)
136
+ rotary_dim = int(self.head_dim * partial_rotary_factor)
137
+ if rotary_dim is None:
138
+ rotary_dim = self.head_dim
139
config.rope_parameters["partial_rotary_factor"] = rotary_dim / self.head_dim
140
141
self.rotary_emb = get_rope(